Welcome to the reading list, a weekly roundup of news and links related to buildings, infrastructure, and industrial technology. This week we look at Icon’s fundraising, employment in China’s solar industry, grid-scale battery reliability, Tesla’s flying roadster, and more. Roughly 2/3rds of the reading list is paywalled, so for full access become a paid subscriber.

Housekeeping items this week:

• I was on MTS Live talking about robots, prefabricated construction, and other things.

• I’m quoted in this Dispatch article about data center construction.

Housing and Cities

From the annals of “suspicious discontinuities,” New York City has a tax break for new projects that set aside some proportion of the units as affordable housing called “485-x.” If a project has more than 99 units, the required share of units that are “affordable” goes up, and so do the wages that need to be paid for construction workers, resulting in a very large number of projects that are exactly 99 units. [X] [Housing Wire]

DR Horton, the largest homebuilder in the US, signed an agreement to access the services of Canadian construction automation startup Promise Robotics for more than 1250 homes. Promise seems to offer some sort of AI-driven, robotic factory-based construction service. “Promise Robotics applies AI across the homebuilding production process by interpreting existing home designs, generating production plans and robotic instructions, orchestrating manufacturing, and connecting that production with site installation.” [Promise Robotics] One possible future I see for construction automation is “AI + robotics changes the logic of prefab construction and makes it more attractive,” so it’ll be interesting to see if things like this play out.

Also in construction automation news, construction 3D printing startup Icon raises a $200 million series D, at a $1.4 billion valuation. [Dealroom] This is surprising to me: I had sort of assumed things weren’t going well for the company. Though this does seem to be a “down round” — in 2022 its valuation was supposedly “approaching $ 2 billion” — I’m nevertheless adjusting my probabilities of success for Icon upward somewhat.

Some interesting data from Harvard’s Joint Center for Housing Studies on immigrants in the US construction industry. One stat I found notable: between 2013 and 2024, the number of foreign-born workers who were construction laborers or carpenters increased massively, even as the number of native-born workers in those occupations declined. But for electricians and plumbers (licensed professions), most of the increase came from native-born workers. [Harvard JCHS]

Manufacturing

GlobalFoundries, the semiconductor manufacturer that was spun off from AMD in 2009, will operate two new fabs planned for New York, funded by $21 billion from Japan. The US government is also taking a 4% equity stake in the company. [GlobalFoundries]

ASML says it can’t currently meet customer demand for lithography machines for 2027 and 2028, and is looking for ways to increase capacity. [ASML]

Ruthless competition in China’s solar industry: employment at China’s five biggest solar manufacturers fell 21% last year, and 31% the year before. [Reuters]

The US isn’t the only country with a small number of companies that make the grain-oriented electrical steel (GOES) needed for power transformers. Only two EU companies, Thyssenkrupp and Stalprodukt, make it, and as in the US they can’t make the highest quality steel. “The global leaders—the companies that make the world’s best GOES—by and large are in [South] Korea and Japan,” Alimento says. “In the U.S. and in the EU, domestic manufacturers are not capable of making that quality of GOES, because it is a heavily protected trade secret.” [IEEE Spectrum]

Thanks to ongoing memory shortages as a result of AI data center construction, smartphone shipments are down 20% year over year. [Counterpoint]

Read more

The scaling hypothesis regards the blessings of scale as the secret of AGI: intelligence is ‘just’ simple neural units & learning algorithms applied to diverse experiences at a (currently) unreachable scale. As increasing computational resources permit running such algorithms at the necessary scale, the neural networks will get ever more intelligent.

Likewise, pseudonymous AI researcher Nostalgebraist wrote in 2024 that “the entire current landscape dates back to the realization that ML models improve steadily and predictably as you train them on more data. And that everything since then consists of training bigger models on larger datasets, enabled by years of rapid improvement in hardware.” Epoch AI found that since the early 2010s, the size of training datasets for AI language models has doubled about every six months. The biggest AI models today are trained on tens of trillions of tokens.

As more and more effort gets devoted to applying modern AI to robotics, an important question is “where will the training data come from?” Language models could be trained on huge existing corpuses of text. GPT-3, for instance, was trained on about 300 billion tokens’ worth of data scraped from the internet. In 2024 Epoch AI estimated the total existing stock of quality, accessible text data is around 300 trillion tokens. But there’s no equivalent stock of good, public robot training data. Figuring out where the data for robotics will come from and how much of it we’ll need has become a key problem for the field.

What is robot training data?

Training an AI model, whether it’s an LLM or a model used for controlling robots, fundamentally requires lots and lots of examples of something being done correctly (or enough knowledge of what a correct action looks like that you can score an attempt). You feed a model some set of inputs, you compare the model’s output to the actual, correct output, and you use the difference between the two to modify the parameters of the model to reduce this difference via some flavor of gradient descent. (We went into this process in some depth earlier this month.)

With text, a training example can be just an existing chunk of text; you take the chunk, feed the model part of it, and have the model predict what comes next in the sequence. One chunk of text can thus generate many training examples by feeding the model different portions of the text.

To train a robot, an example is typically some sort of recording of a task being accomplished or attempted. There are a variety of different methods used to create these recordings and turn them into training data.

Historically, one of the most common methods for creating these recordings is called “teleoperation.” With teleoperation, a person manipulates some controls that are directly connected to a robot, and uses them to move a robot through some particular task (folding a shirt, picking up and moving objects, and so on). The image below, for instance, shows a person teleoperating a pair of robot arms, using them to pour water from a pitcher into a glass.

As the person performs the task, data of the robot’s state is recorded, along with images from cameras that are recording the task (often mounted to the robot itself), which are combined to create a training example. An example might consist of feeding a model some data on the robot’s state and what the cameras can see at some particular point in the sequence. The model will try to predict what the next robot actions are (in the form of something like joint angles), and then compare that to what the actual actions were. This sort of training strategy, using examples generated by an expert (a human) and then training a model to mimic them, is known as behavioral cloning, and it’s part of the broader category of “imitation learning.”

The major benefit of this method is that the data it generates very closely matches the robot’s actual movement capabilities, because it’s produced from the actual robot you want to train, and it remains widely used today. You sometimes see videos or pictures from huge data collection factories, filled with people doing nothing but teleoperating robots to complete various tasks, and you sometimes see teleoperation accidents when the person forgets that they’re controlling the robot and suddenly moves in a way that the robot can’t handle.

But teleoperation has drawbacks. The biggest one is probably expense: to get the data, you need to have an actual robot, along with a person to sit there and manipulate it, which makes it hard to scale this method to produce very large amounts of data. Another drawback is that it’s somewhat clunky for a person to control a robot this way; they often can’t get any sort of tactile feedback from the object being manipulated to guide their motions, so this won’t necessarily produce really good data on dexterous manipulation. The clunkiness can also make it hard to collect good teleoperation data, and as I understand it, only a fraction of total teleoperation data collected ends up being useful for training.

It’s also possible to generate training data using an existing robot by having it move around on its own, recording its actions, and using some sort of system to categorize these behavioral sequences as successful or not. This sort of strategy, where the robot moves autonomously and you score its actions in some way, is known as “reinforcement learning.” Google was doing something like this around 2016 with an “arm farm”: a series of robot arms moving autonomously, with success scored by whether an arm picked up or grasped an object or not.

This sort of arm farm doesn’t seem to be widely used today, but using real-world deployed robots will probably eventually be the largest source of robot training data. You often hear robotics folks talk about the need for a “data flywheel,” where you have robots deployed in the real world, you use the data from them to train even better models, resulting in more robot deployments, resulting in better models, and so on.

Robotic AI company Physical Intelligence, for instance, is experimenting with this sort of “deploy a robot to do real work, and use it to gather data to improve the model” proto-flywheel at a chocolate factory near its offices. But while this might work eventually, right now there’s a very small number of AI-controlled robots doing anything resembling real work. Real-world data also tends to be extremely repetitive — robots doing the same task over and over again in very similar conditions — which makes it much less useful for training. So for now most data needs to come from other sources.

Another very common strategy for generating robotic training data is simulation: instead of running a physical robot in the real world, run a digital one in a simulated world instead. This is also commonly done with reinforcement learning, in which you have the robot try to navigate some simulated space, after which you score it on how well it does. Another flavor of this is the teacher/student paradigm where you have a “teacher” model that has full access to all the data within the simulated environment, and a “student” model that can only see what the real robot would see (images from its cameras and its own internal state), and you try to get the student to mimic the teacher.

The benefits of simulation are that it’s much cheaper and faster to simulate a robot than it is to deploy one in the real world, and that you can get very large amounts of data by simulating the behavior of thousands of robots. Figure AI taught its humanoid robot to walk via reinforcement learning on thousands of simulated robots. OpenAI taught a robotic hand to solve a Rubik’s Cube via the equivalent of 13,000 years of simulated experience. By 2019, Waymo had driven 10 million miles in the real world, but 10 billion miles in simulation.

But simulation also has drawbacks. The biggest one is that the quality of the data it produces is a function of the accuracy of the simulation. For some things, such as locomotion (having a robot move through some physical space), simulation is fairly accurate, because the behavior is mostly well-characterized enough that we can represent it digitally. So you can train a quadruped robot to navigate some terrain, or a bipedal robot to dance, or a drone to fly along some path in simulation, and the simulated environment will be a reasonably good approximation of the real world. But for things such as object manipulation (particularly for soft or deformable objects), it’s much harder to build a simulation that accurately captures real-world behavior, and using it as a source of training data is much more difficult. This is generally referred to as the “sim-to-real gap.”

Because of simulation inaccuracies, one popular simulation technique is called “domain randomization.” Since you don’t know what various parameters like friction will be in the real world, you decide to vary the parameters randomly when running the simulation, in the hopes of creating a training envelope that encompasses the parameters that will be encountered in the real world. When OpenAI trained its robotic hand to solve a Rubik’s Cube, for instance, it continuously expanded the range of parameters being tested, forcing the model to learn to solve the cube in more and more difficult environments, in the hopes of giving it capabilities that would generalize to the real world. The robotics team there used this strategy specifically because accurate physical modeling was too difficult:

The biggest challenge we faced was to create environments in simulation diverse enough to capture the physics of the real world. Factors like friction, elasticity and dynamics are incredibly difficult to measure and model for objects as complex as Rubik’s Cubes or robotic hands and we found that domain randomization alone is not enough.

To overcome this, we developed a new method called Automatic Domain Randomization (ADR), which endlessly generates progressively more difficult environments in simulation. This frees us from having an accurate model of the real world, and enables the transfer of neural networks learned in simulation to be applied to the real world.

But despite this technique, and really quite a lot of simulation (13,000 simulated years), performance was only OK — for maximally difficult starting arrangements, success was only 20% — which is maybe why this technique doesn’t seem to be widely used for manipulation training currently. Environmental parameter randomization does seem to be fairly standard for locomotion training though, used for things like randomly varying the coefficient of friction of the surface being traversed.

Another way to use simulation is to create synthetic training data: taking real training data as an input and using it to create lots of similar training examples. Nvidia, for instance, has a page showing how you can use its Isaac GR00T robot simulation platform to generate synthetic robot training data using teleoperation data as an input. Synthetic data generation also seems to be widely used for training autonomous vehicles.

We can also create robot training data by recording humans doing tasks and creating some sort of mapping from the recorded task data to a robot. This might be as simple as fitting a person with a head-mounted camera (creating what’s known as “egocentric video”), and using it to record the person doing various tasks with their hands. The video is then processed so the person’s motions are interpreted as robot motions, and used to create training examples. (This is one reason for the popularity of the humanoid form factor in robotics: humans doing tasks will be a reasonably good source of training data.) Figure AI, for instance, recently announced Index, its program for gathering training data by sending people head-mounted cameras to record themselves doing various tasks (alternatively, Figure will also send a crew equipped with these cameras to your home or business to do chores for you). Companies like Shift are also doing this, sending people with hat-mounted cameras to folks’ houses to record themselves doing chores.

You can also add more sensors to collect more elaborate human task data. The now-defunct startup Eidon AI stuck sensors on folks’ arms to collect data on joint positions for training robot arms. Neocambrian AI recently announced “Mitts,” which it describes as “state of the art tactile gloves that capture high-precision touch data for robots.”

Another flavor of this strategy is to use what’s called a Universal Manipulation Interface, or UMI. With UMI, a person holds a pair of specially designed manipulators, that typically have a simple pincer grip at the end (though some have multiple fingers). The person does various tasks with the manipulators, which have cameras mounted to them that record exactly what the manipulators are doing. This video is then used to create training examples for robots that have similar manipulators. Robot model startups Generalist AI and Sunday Robotics, for instance, seem to be primarily using this strategy for gathering training data.

The benefit of using human-collected task data is that it’s much cheaper to collect than teleoperation data: all you need is a camera, or some camera-equipped manipulators, rather than an entire robot. And it’s probably much easier to get a wide variety of different task recordings by just having a person walk around and do different things, rather than constantly repositioning a robot to do them. And because the data is gathered in the real world, by people manipulating real objects, it’s more accurate than simulation. The drawback is that because this data comes from a person rather than a robot, it won’t map onto a robot’s capabilities perfectly well. Even UMI, which minimizes the human-robot differences by having a person do tasks using a sort of robotic manipulator, won’t necessarily map perfectly, though the discrepancy is small and many of the most impressive robotics demos we’ve seen (such as Generalist AI’s) have been trained using UMI data.

It’s possible to use multiple training strategies in a single robot model. Agility Robotics’ Digit, for instance, walks using a small, 1-million-parameter neural network trained in simulation using reinforcement learning, while object manipulation seems to have been trained using behavioral cloning. Dyna Robotics similarly seems to have used a combination of reinforcement learning in simulation and human video for training data.

It’s also often useful to train a robot model using data other than task completion data. Vision-Language-Action models, as we noted earlier, are built from Vision-Language models trained on text and images. And because of the ubiquity of general video data on things like YouTube or TikTok, there’s a great deal of interest in finding ways to use generic web videos as a source of training data. Nvidia’s Cosmos, for instance, is a foundation model trained on video data that can be used as a backbone for a robot policy model.

Robot dataset sizes

How much of this action-relevant training data have we gathered so far, and how much do we need?

As with AI models more generally, robotic training datasets seem to be getting bigger. An early dataset that was considered large at the time was the Open X-Embodiment (OXE) dataset, created by Google DeepMind in 2023 by stitching together dozens of existing datasets. OXE consisted of over 1 million robot action trajectories spanning more than 160,000 different tasks, totaling around 4,000 hours’ worth of task data. Physical Intelligence’s Pi 0 model, released in October of 2024, was trained on 10,000 hours of robot task data. Generalist AI’s GEN 1 model, announced earlier this year, was trained on around 500,000 hours of UMI data. Figure AI’s Index announcement stated that it was gathering 30 minutes of video data every second, or 43,000 hours a day. Dyna Robotics and Nature Will both claim that their latest models are trained on over 1 million hours of human egocentric video data. I saw one estimate that the robot data collection industry is collectively producing around 500,000 hours of training data a month.

It’s not amazingly easy to compare these datasets to the text datasets used to train LLMs. For one, different robotic training sets have different levels of quality — teleoperation or UMI data is probably higher quality than pure human egocentric video data for training purposes, and a highly variable dataset with lots of different sorts of tasks is much more useful than an extremely repetitive one. And video data for robotics training is informationally different from text data: for one, video data tends to be highly correlated — on short timescales, one frame is pretty similar to another frame. Thus even though video recordings contain a lot of data, they’re probably less informationally dense for training purposes than text data. As robotics researcher Chris Paxton notes, “machine learning works best the closer you are to an independent and identically distributed (iid) dataset; neither web data nor robot data is iid, but robot data is far less iid than web data.”

But we can still try. One very rough way to compare dataset sizes, used by Physical Intelligence researcher Kevin Black in this presentation and robotics professor Ken Goldberg in this presentation, is to estimate how long a dataset of text tokens would take for a person to read, and compare that to the hours of robot training data we have. Under this metric, GPT-2 was trained on the equivalent of around 475,000 hours of data, GPT-3 was trained on 26 million hours, and GPT-4 was trained on 685 million hours. Using this (somewhat dubious) metric, even several-year-old AI models were trained on something like 1,000 times the data that current robot models are trained on. Goldberg calls this the “100,000-year data gap.”

On his Substack “It Can Think!” Paxton makes an estimate from the other direction, converting hours of robot data into tokens, and comes up with a fairly similar estimate. He guesses that for video running at 10 frames per second, each frame will contain the equivalent of one useful training token. Every hour of video thus contains about 10 × 60 × 60 = 36,000 training tokens. He then divides this by 10, on the assumption that you can’t expect every second of video data to actually be useful, getting you about 3,600 training tokens per hour of video. This would make Dyna Robotics’s 1 million hours of training equal to around 3.6 billion tokens, on the order of 5,000 times smaller than the datasets used to train recent AI models.

If you assume that we need about 1,000 times as much real-world robot training data as we actually have — on the order of 1 billion hours of data — how much would that cost to generate? Eidon AI reports robot data prices of $60 to $100 per hour. Pantheon similarly reports UMI data costs from $60 to $150 per hour, but that its engineers developed a pipeline capable of generating data for $10 per hour. Shift claims you can earn up to $20 per hour recording egocentric video data and selling it to them. So a naive estimate suggests that getting enough data to get robot models to a similar level of capability as recent AI models might cost on the order of $10 billion to $100 billion.

How much data we actually need, however, seems to be something of an open question. If we don’t care about generalization, and only want to train robot models to perform various specific tasks, the data requirements might be substantially less. Alternatively, it might turn out that frontier general AI models like Fable or Astra are extremely capable robot controllers after being fed a comparatively small amount of robotic training data; some folks are already predicting that special-purpose robot AI models might get replaced with more generic AI models. Likewise, we might find robotic AI model architectures that are much more sample efficient — learning much more from each training example — than current models.

On the other hand, it could be the case that moving around and manipulating objects in the real world is a fundamentally more difficult task than manipulating text on a computer, and that models require a lot more training to become competent at this real-world behavior. Small errors in generated text, for instance — misspellings, or slightly wrong word choice — don’t necessarily compromise the text as a whole, but when you’re manipulating objects a single tiny positioning error can be the difference between success and complete failure. Likewise, Ken Goldberg argues that robotic movement is a much higher-dimensional problem than generating text (though I’m not sure I buy the version of the argument he articulates in that presentation, but I have seen this sentiment echoed by other robotics researchers).

Conclusion

LLMs showed the benefits of scale when training AI models — bigger models, more data, trained using much more computation. We’re now in the process of seeing how well this scaling works for robotics, and robotics researchers are working hard to gather enormous amounts of robot task training data in the hopes of seeing the same sort of trends with robotic capabilities that we have with AI more generally.

Welcome to the reading list, a weekly roundup of news and links related to buildings, infrastructure, and industrial technology. This week we look at propane in air conditioners, Anduril making submarine parts, nuclear reactors on the moon, the robotic hand landscape (handscape?), and more. Roughly 2/3rds of the reading list is paywalled, so for full access become a paid subscriber.

Housing

Someone created a city building game where you try to navigate local opposition to get housing built in San Francisco. [X]

I somehow missed this paper from last year analyzing the impact of International Building Code changes from 2012 to 2021 on construction costs. “Between 2012 and 2021, the International Code Council (ICC) adopted 130 revisions to the International Building Code (IBC) affecting multifamily development. Using a detailed prototype of a three-story apartment building and holding material prices and labor productivity constant at 2024 levels, we find that 16 of these revisions raised national average construction costs by 9.9 percent relative to a 2009 IBC baseline, with fire protection requirements accounting for the largest share.” [SSRN]

Should we be using propane as a refrigerant in our heat pumps and air conditioners? [The Rebuild]

Another look at what’s in the Senate’s new permitting reform bill. [Abundance and Growth]

Manufacturing

Atomic Machines, a manufacturing startup using a vaguely described manufacturing process that “builds working micro machines from code alone” (whatever that means) emerged from stealth and got a lot of attention on Twitter. [X] A NYT article on the company has a little more information, but not much. “Using the same techniques that drive a chatbot, Atomic Machines is building A.I. technology that can discover new ways of designing extremely small devices. By analyzing thousands of experiments involving standard manufacturing materials and finding patterns in how these materials interact, it learns to piece them together in ways that produce viable devices.” [NYT]

The CEO of OSH Cut (a similar sort of business to SentCutSend) on what he thinks the future of manufacturing looks like. A few similarities here to my concept of “AWS for Everything.” [X]

The Navy awarded defense startup Anduril a $2.9 billion contract to manufacture components for Virginia-class nuclear submarines. “Arsenal-2 will build complex components and large scale assemblies of the Virginia-class submarine and deliver them to Electric Boat and Newport News Shipbuilding for final assembly. Adding another source of this work will allow more submarine construction to happen in parallel across the industrial base and increase the overall capacity available to the Virginia-class program.” [Anduril]

Domestic car sales in China have apparently not risen at all since roughly 2015? [X]

Related, Chinese car sales are really taking off in Britain. “The Jaecoo 7, made by China’s Chery Automobile, was Britain’s bestselling vehicle in September, an important sales month in the U.K. In the first half of 2026, almost every sixth car sold in Britain bore a Chinese brand, according to Schmidt Automotive Research, a higher proportion than in any other European country except Norway.” [WSJ]

Read more

In 1960, a team at the University of Illinois released the first version of their PLATO “computerized instruction system,” a network of what would eventually be thousands of computer terminals connected to large mainframe computers. Users of PLATO could use the system to learn about various topics, with courses on subjects such as chemistry, math, music, and foreign languages.

The first iteration of PLATO used a simple TV screen for a terminal display, but this was at best a stopgap measure. TV screens used cathode ray tubes (CRTs), which used a magnetically deflected beam of electrons to paint an image on a glass screen. This beam redrew the entire screen 30 times a second, which meant that a computer had to feed the CRT screen a constant stream of high-bandwidth video data. This was fine for broadcast TV, since the same signal was being sent to every TV that received it, but on the PLATO system every terminal needed to be sent different data, depending on what the user was doing. Not only would this require high-bandwidth data connections between the mainframes and the PLATO terminals, it would force each terminal to have a large amount of memory storage to hold the video data, which at the time was extremely expensive. CRT-based screens worked well enough for a small system with a few terminals, but it was felt that the technology wouldn’t scale to the thousands or millions of terminals the PLATO designers envisioned. What was needed was a new type of screen, one that had a built-in “memory.” On such a screen, only the things that changed on the screen would need to be redrawn, which would dramatically reduce the bandwidth and data storage such a screen required compared to a CRT.

Donald Bitzer, one of the architects of the PLATO system, considered possible methods that might be used to make such a display. One potential option was gas plasma — running an electric current through a low-pressure gas, which would cause that gas to glow. Scientists had been making low-pressure gases glow using electric current since the 1600s, and the phenomenon had been used in neon signs since the early 20th century. But no one had yet found a way to turn the phenomenon into a computer display, which would require an array of thousands or even millions of independently controlled tiny gas plasma lamps, one for each pixel on the screen. Bitzer decided it was worth trying.

In 1962, Bitzer’s graduate student Robert Willson built a one-cell model of a plasma display screen, the equivalent of a single pixel. It consisted of a “sandwich” of three thin sheets of glass, with wires on either side of a tiny pocket of argon gas contained in a hole in the middle glass sheet. When electric current ran from the wires through the gas, the argon would glow brightly. Willson’s model suggested that a plasma screen might work, but also that there were many problems that needed to be overcome: the brightness of the cell was inconsistent, components burned out quickly, leaks were rampant, and over time the electricity-excited gas would deposit itself on the electrodes, reducing the flow of current. Willson’s design also required each individual cell to have its own resistor to prevent an electric arc forming between the electrodes, which was projected to add significant manufacturing difficulties on a full-sized display: a 500x500 pixel screen would require 250,000 individually attached resistors.

A breakthrough occurred in 1964, when Bitzer’s team had the idea to move the electrodes to the outside of the glass sandwich, which eliminated the need for a separate resistor. When building a test display of the new model, the team made a serendipitous discovery: a small amount of air accidentally introduced into the cell gave the display the sought-after memory effect: the gas would continue to glow even when the voltage at the electrodes was reduced. By 1967, the team had scaled up their single-pixel display to a 16 by 16 array of pixels (weber 2006), but this was still little more than a fragile prototype: leaks and contamination remained common, and the extremely thin sheets of glass (originally used for microscope slides) broke easily.

In 1967 Owens-Illinois, a glass manufacturer, licensed this technology, known as a plasma display, and began working to turn it into a real, practical product. Owens-Illinois eliminated the middle sheet of glass and the individual gas-filled pockets along with it, and simply filled the entire space between the two outer glass sheets with gas. Even though the gas within the panel was continuous, it would only glow at the intersection of charged electrodes, allowing individual pixels to be controlled. In 1968 Owens-Illinois debuted a 128 x 128 pixel plasma display, far sturdier and more reliable than the University of Illinois prototypes, and in 1971 the company shipped the world’s first commercial plasma display panel, a 512 x 512 screen dubbed the “Digivue.” The PLATO system, which would shortly be scaled up to thousands of terminals worldwide, was its first customer.

But the newly released plasma screen wasn’t alone in the market for electronic displays. The plasma screen had originally been conceived as an alternative to CRT screens and the bulky and expensive external memory they required. But memory technology wasn’t standing still either: in the late 1960s, as Illinois researchers were building plasma screen prototypes, researchers at IBM developed DRAM, a type of computer memory that stored data using transistors and capacitors. Computer memory was soon being pulled along by Moore’s Law, getting smaller, faster, and cheaper with every generation, and before long the memory requirement for CRT screens was a minor concern.

Plasma screens did have other advantages, beyond the way their physical properties enabled a kind of “memory.” Unlike a CRT, which required a bulky glass tube, a plasma display could be made extremely thin. And plasma screens could potentially be much more rugged than CRTs, which had fiddly and easily disrupted electron guns.

But plasma screens also had drawbacks: they were very power hungry compared to CRTs, weren’t as bright, and were far more expensive. And while Illinois researchers tested a color plasma screen in the late 1960s, it would be many years before a color plasma screen robust enough to be commercialized was developed.

What’s more, plasma wasn’t the only new display technology that appeared in the early 1970s. During the 1960s, scientists at Westinghouse and RCA had experimented with displays based on liquid crystals: substances that were liquid but displayed organized, repetitive structure like a crystal. Liquid crystal molecules were found to change their orientation when an electric field was applied to them, and if polarized light (light oscillating within a single plane) passed through a liquid crystal, its plane of polarization would rotate. By the late 1960s, researchers had used these phenomena to construct a display. The first liquid crystal display (LCD) screens appeared in calculators and other small electronic devices in the 1970s, and by the 1980s were being used in portable TVs.

Plasma didn’t slow down either. Over the next several decades, plasma display technology steadily improved. Engineers developed better electrodes, reduced power consumption, and invented a method for creating grayscale by rapidly flickering the pixels off and on (the more a pixel was “off,” the darker the pixel). Grayscale, in turn, opened the door for full-color plasma displays, and these began to appear in the late 1970s.

As their capabilities improved, plasma displays began to carve out a small niche in the electronic display market. IBM introduced a line of electronics for banks with plasma displays in 1973, and in the mid-1970s plasma displays began to be used for various military applications thanks to their rugged, low-maintenance designs. In the 1980s, laptops and portable computers with monochrome plasma screens began to appear. By the 1990s, however, LCD technology had advanced to the point where high-quality color screens could be made for laptops, entirely displacing plasma from that market.

As it was forced from the laptop display market, plasma was finding success in a new market: large-screen, high-definition televisions. Up through the 1990s CRT had been the standard technology used to make TVs, but the size of the tube required made it difficult to scale CRTs up to very large screens of 40 inches or more: when Sony introduced a 45-inch CRT TV in 1988, it weighed over 400 pounds. Thinner plasma screens, by contrast, could more easily scale to very large screens. Japanese electronics manufacturer Fujitsu demonstrated a 31-inch plasma TV in 1990, and then a 42-inch plasma in 1995. US display startup Plasmaco debuted a 60-inch plasma screen in 1999, and Samsung displayed an enormous 102-inch plasma screen in 2005.

However, LCD was entering the large-screen TV market at the same time, and by the early 2000s LCD manufacturers were making screens of similar size as plasma screens. Plasma had some advantages over LCD TVs — they could offer richer colors and higher refresh rates — but also some disadvantages. Plasma screens consumed more power, put out more heat, and were less bright than LCD screens. And for a given screen size, an LCD screen could be made much higher resolution than a plasma screen.

But the most important issue was likely manufacturing: it proved harder to achieve high yields and low manufacturing costs with plasma screens than with LCD screens, possibly because manufacturers invested far more in LCD manufacturing than they did plasma.

By the mid-2000s, plasma’s screen quality advantages had either been whittled away or simply gone unnoticed by most consumers: in 2006 reviewers found them roughly equal in quality, and LCD was significantly outselling plasma screens. Plasma and LCD battled over the TV market through the early 2010s, but by 2013 plasma had lost nearly all its market share. By the end of the 2010s, plasma screen manufacturing had essentially ceased, and today the technology is all but dead.

The classic pattern of new technology adoption is something like an S-curve: it starts out serving niche uses for users who need its particular combination of attributes, but as the technology gets more capable and less expensive, it gets adopted more broadly, until it reaches some particular peak. When a technology gets replaced, in turn, it tends to follow this pattern in reverse, getting removed from more and more uses but continuing to be used by users who value its particular combination of attributes. Sometimes this combination allows an “outdated” technology to continue to be used far into the future; vacuum tubes were replaced by the transistor for most uses decades ago, but they’re still manufactured today for use in certain audio equipment. But sometimes this isn’t enough to save a technology; while some folks prefer the look of plasma screens, this preference evidently wasn’t enough to maintain production of it.

Welcome to the reading list, a weekly roundup of news and links related to buildings, infrastructure and industrial technology. This week we look at a Senate permitting reform bill, the potential of Tesla’s Cybercab, the Navy’s $200 billion worth of shipyard renovations, a possible plan to return a SR-71 to service, and more. Roughly 2/3rds of the reading list is paywalled, so for full access become a paid subscriber.

Housing and Cities

The senate has released the text for a permitting reform bill, the Bipartisan American Affordability and Jobs Act. My IFP colleague Aidan Mackenzie has a brief overview of what’s in it, as does Thomas Hochman at FAI. [X] [Green Tape]

A tweet confirming my biases that people mostly don’t care about building aesthetics, and that improving aesthetics/design is not a particularly good avenue for addressing NIMBYism. [X]

A New York judge rules that NYC’s new tax on second homes wasn’t implemented correctly, and needs to be redone. “The judge, Wayne M. Ozzi of State Supreme Court on Staten Island, sided with a group of homeowners who had sued the city. The homeowners had argued that the Mamdani administration’s Department of Finance did not try hard enough to determine who should owe the tax before moving toward collecting it.” [NYT]

Manufacturing

Friend of the newsletter Austin Vernon on the potential of Tesla’s Cybercab, and the ways its driving down manufacturing costs. “Painting cars has become an enormous pain. The paint shops are expensive, trigger all sorts of environmental thresholds, and they complicate production. Usually doors have to be put on the car, painted, taken off for assembly, then put back on at the end. Painting really has to go. Tesla tried a stainless steel exoskeleton with Cybertruck, but that ended up being challenging. The Cybercab has injection-molded plastic body panels instead of typical painted sheet metal. Cars have had plastic body panels before, but the panels didn’t come out with a good finish or color, requiring painting. Tesla’s “Reactive Injection Molding” parts do come with a smooth finish and with color incorporated so the paint shop can finally disappear, greatly simplifying production.” [Austin Vernon]

Mesabi Metallics is planning a $15 billion steel mill in Iowa, the “largest single investment ever made in a US steel facility.” It looks like instead of a blast furnace, the mill will use direct reduction of iron ore, which will then be fed into electric arc furnaces. [ENR]

Raytheon gets a contract to produce AMRAAM missiles worth up to $20.7 billion. [Breaking Defense] And in addition to building the F-47, Boeing has also won the competition to build the next generation fighter for the US Navy. [Breaking Defense]

A Federal Reserve paper estimates that semiconductor shortgages in 2021-2022 reduced US car production by more than 2 million vehicles. [Federal Reserve]

The Washington Post editorial board argues that we shouldn’t ban Chinese-made cars in the US. I didn’t realize that a substantial number of cars already had large numbers of Chinese parts, and that some actually had final assembly in China. “Two cars for the American market from U.S. brands, the Lincoln Nautilus and Buick Envision, have final assembly in China using nearly all-Chinese parts. Thirty-nine other automobiles from model year 2026 that are on the roads have Chinese parts, including several that are assembled in the U.S., according to the National Highway Traffic Safety Administration.” [Washington Post]

SpaceX’s Terafab is scheduled to be constructed and begin operations very quickly. [X]

Claims that China is preventing firms from setting up lithium-iron-phosphate battery plants in India. “LFP is not available anywhere in the world outside of China, and right now they are guarding it like a weapon.” [X]

Read more

We’ve also seen some impressive-looking robotics demos over the last year, though these must be evaluated with a large grain of salt. Figure showed off its robots sorting packages for hours at a time and autonomously unloading a dishwasher. Physical Intelligence showed off its robot model making coffee and folding boxes in a chocolate factory. Generalist AI showed off its GEN-1.5 model learning new tasks from just a few examples.

Some progress in robots is a function of hardware advances: better actuators, and so on. But the lion’s share of it is due to advances in robot AI, using specially developed AI models to control a robot. As more general AI continues to rapidly advance in capabilities, one obvious question is whether we’re going to see the same sort of acceleration in robotic AI. Right now robotic AI capabilities still seem very limited, but given how fast more general AI improved, I can imagine this changing very quickly. Because of this, it’s worth understanding how the AI models being developed to control robots work.

There are a few different sorts of robotic AIs, which are sometimes referred to as “policies” (where a “policy” is something that maps a particular set of inputs — robot state, sensor data, instructions it’s been given — to a set of robot actions). For this essay, we’ll look at one commonly used architecture, employed by companies such as Figure, Unitree, Physical Intelligence, and Nvidia: the vision-language-action model (VLA). More specifically, we’ll look at one popular VLA, Physical Intelligence’s open-weight π0.5 VLA, which was released in 2025 and has become widely used (though it’s not Physical Intelligence’s most advanced model).

While a large language model takes text as an input and gives text as an output, a VLA takes text, images, and robot information as an input and gives a series of robot actions as an output. It does this using many of the same components that made LLMs so successful, specifically attention and the transformer architecture. It’s not amazingly obvious if VLA’s will remain the primary robotic AI paradigm, but as of now they’re a very common method for controlling robots.

Linear algebra basics

To learn how AI works (be it a VLA or anything else), it’s useful to know a very small amount of linear algebra. Specifically, we want to understand a few different operations for manipulating arrays of numbers, since this is most of what AI models do.

For instance, say we have the following two-dimensional arrays (or matrices) of numbers:

One thing we can do to these arrays is add them together. This works exactly like how you’d expect: each number in one array is added to the corresponding number in the other array, giving you a new array with all the resulting additions.

Another thing we can do is multiply or divide an array by a single value. This also works more or less like you’d expect: each value in the array gets multiplied or divided by the respective value.

But what if we want to multiply two matrices together? To do this, we need an operation called the dot product. The dot product takes two one-dimensional arrays of numbers (also known as vectors), multiplies each value in one vector by the corresponding value in the other vector, then sums the result.

You can think of the dot product as measuring two things: how large two vectors are, and how similar they are to each other. (If you normalize the vectors, scaling the values so everything is between 0 and 1, then the dot product is entirely a measure of how similar two vectors are to each other.)

When we multiply two matrices together, we’re simply doing a bunch of dot products: each value in the resulting matrix is the dot product of the row of one array and the column of another. Because the dot product requires two lists of equal length, multiplication of matrices must be done in a certain way: the number of columns of one matrix must be equal to the number of rows of the other matrix. And the size of the output matrix will be a function of the sizes of the input matrices.

I find it easiest to understand matrix multiplication by putting one matrix on the left and the other matrix on the upper right. The output matrix fits in the space between them, each value the dot product of the first matrix’s rows and the second’s columns:

Neural network basics

Most modern AI models, as we know, are built using neural networks. The classic image of a neural network is something like this:

You have some series of input neurons that represent your input data, which are connected to various intermediate layers of neurons, which then get connected to output neurons. Depending on what data is fed into them and how the connections between different neurons have been set, neurons will “activate” to various degrees, taking some input from the neurons in the previous layer and sending some output along to the neurons in the next layer. This structure is sometimes called a “multilayer perceptron,” or MLP.

However, I think it’s much easier to understand neural networks by looking at how they’re actually implemented, which is done by multiplying arrays of numbers (in machine learning, these arrays are sometimes called “tensors,” where a tensor is just an array of numbers that can be any dimension. Google, for instance, has a machine-learning library called “TensorFlow”). The neural network above, for instance, would actually be implemented via something like this:

The three input neurons become a one-dimensional array, or vector, with three numbers in it, which is our input data. This vector then gets passed to the hidden layer, where it goes through several steps. First, it gets multiplied by a two-dimensional matrix (W1 in the figure). This matrix has four rows (corresponding to the four neurons in the original diagram) and three columns (one for each value in the input vector): it takes a vector three numbers long (or 3-vector) as an input and spits out one four numbers long as an output. The resulting vector of four numbers is then added to another 4-vector (b1 in the figure), the values of which are known as “biases.” The weights and biases of the various layers are known as the “parameters” of the network and are what get modified when the neural network is being trained.

Once the biases have been added, the resulting vector (known as a “pre-activation,” z in the figure) then goes through what’s called a nonlinearity, which is just a function that does different transformations to an input depending on what that input is. A common nonlinearity, which the example above uses, is “ReLU,” for rectified linear unit. All ReLU does is replace any negative numbers in the vector it’s given with zeroes. Other nonlinearities are the sigmoid (an S-shaped function that squeezes values to be between 0 and 1) and GELU (a somewhat ReLU-like function that’s smooth instead of sharply kinked).

In a basic MLP like this one, each hidden layer in a neural network will consist of this matrix multiplication, then bias, then nonlinearity. The network above has just one hidden layer, so after the nonlinearity the resulting vector is then passed to the output neurons. These transform the data again using another matrix multiplication (W2) and bias addition (b2), yielding a vector of two numbers as the output (y).

This structure is simple but powerful. To see why, imagine a neural network that takes a single value as an input and gives a single value as an output but has many neurons in its hidden layer (the number of neurons is known as the “width” of the layer). Before the nonlinearity, each individual neuron calculates wx + b, which is the formula for a straight line; the weights determine the slope of the line, while the biases determine its vertical location. The nonlinearity, in turn, lets the layer draw different line segments, with different slopes and intercepts, at different points on the graph. Together, this allows a sufficiently wide neural network to, in principle, approximate any continuous function by drawing a large number of different line segments1: the more hidden-layer neurons we add, the more line segments we can draw, and the closer we can approximate the function. In higher dimensions — larger numbers of inputs and outputs — the geometry gets more complicated, but the basic idea remains the same.

Training a neural network

Other than their ability to approximate very complex functions, the other key property of neural networks is that they can be trained, by comparing the actual output of a network to the output you’d like it to have, and nudging the weights and biases in response to make the output more accurate.

Training a network starts with a forward pass, giving it an example input and then calculating an output. This calculated output is then compared to the actual, correct output using some loss function, which measures the difference between the two. For a simple case where the output is a single number, the loss function might be mean squared error, which you can calculate by squaring the difference between the actual output and the expected output: the bigger the difference, the larger the loss. If you’re trying to predict the next symbol of a sequence, however, a common loss function is cross entropy, which is based on how likely the model thought the actual next symbol was; the less likely the model thought the correct symbol was, the higher the cross entropy. As a model gets trained and does a better and better job at approximating the correct output, the loss will (hopefully) trend downward, and reports on AI training runs will often show training loss curves.

Once the loss has been determined, the next step is to go back through the network, layer by layer, and, using the chain rule, calculate how much the loss would change if you made small adjustments to each weight and bias. This is known as backpropagation and produces a long list of numbers, each number telling you how much the loss would change when you adjust the value of one parameter in the model. This list, called the gradient, can be thought of as the direction you’d need to “step” in a high-dimensional space to maximally increase the loss for the current values of the model’s parameters.

The gradient in hand, the next step is to take a small step in the opposite direction (since we want to decrease the loss, not increase it). This is done by multiplying each value in the gradient by a small positive number (known as the learning rate), and then subtracting the resulting number from each parameter. With the parameters changed, you can then run another input through the model, calculate the new loss, and modify the parameters again. This process of repeatedly calculating the gradient, then using it to modify the parameters of the model to (hopefully) reduce the loss is known as gradient descent.

In practice, most models are trained using modified, more complex versions of gradient descent that do things like use the average of recent gradients to update the parameters and use different learning rates at different points in the training process. It’s also uncommon in practice to update the model after every individual training example. Models are instead trained in batches (sometimes called “minibatches”): dozens, hundreds, or even more training examples are run through the model, their losses are averaged, and the gradient is calculated based on the average loss. Settings that determine how a model is trained, such as learning rate or batch size, are known as “hyperparameters,” and figuring out the best hyperparameters to train a model with is an important area of machine learning.

Inputs to a neural network

For an AI to do useful work, we’d like to be able to feed it things like images and text data. How do we convert images, text, and other data into a form that a model can understand and operate on?

For text, there are two main steps. First, a string of text is broken up into chunks, or “tokens.” Then, each token is converted into a long vector of numbers, called an “embedding,” which can then be processed by the network. Let’s take each step in turn.

Tokenizing is a way to break a string of text into a series of chunks, which is useful because it allows the model to learn something about the meaning of individual chunks of text. One option would be to break apart text using a predefined list of words, but a fixed list of whole words would prevent us from processing words or character combinations that haven’t been seen before. So instead it’s more common to use token libraries, which contain both individual characters and clusters of characters (including whole words as well as portions of words) that have been found to occur frequently in actual bodies of text. The token vocabulary for GPT-3, for instance, has around 50,000 items in it. You can browse the GPT token vocabulary here and see that it contains individual characters, chunks of words, and entire words (often with punctuation or spaces included).

Once a text is broken up into individual tokens, each token then gets converted into an embedding, a long vector of numbers. The embedding vector for GPT-3, for instance, is 12,288 numbers long. You can think of this embedding vector as representing coordinates in some very high-dimensional space. The conversion is done using an embedding matrix, which is effectively a huge matrix where each row is the embedding for one particular token. The image below shows an example of converting the string “the cat sat on the mat” to a series of embedding vectors using a (highly simplified) token vocabulary and embedding table.

One key fact about the embedding matrix is that the values in it — the numbers in each embedding vector — are model parameters that are learned, just like the weights and biases in a multilayer perceptron. As a model is trained, the embedding vectors will get updated via whatever flavor of gradient descent is being used, and tokens that the model learns are similar or even substitutable might end up closer to each other in “embedding space.”

With images, we can use a similar process, but it must be modified somewhat. In something like a vision transformer (an AI model used for things like image classification), an image is first broken into a series of smaller chunks (say, a square of 16 × 16 pixels) called patches, the image equivalents of tokens. However, unlike with tokens, it’s not feasible to have a unique embedding vector for every possible image patch, because the number of possible patches is too large. Even if an image was limited to just black and white pixels, the number of unique 16 × 16 patches is 2^256, which is not all that much less than the estimated number of atoms in the universe. This is far too many items to list in a table.

Instead, patches are converted to embedding vectors using a learned embedding function. A patch is first converted to a very long vector containing the numerical color values of each pixel. This vector is then converted to an embedding vector using some learned embedding function (such as multiplying by a learned matrix), the parameters of which are updated via training.

For both text and images, this process produces a matrix, where each row is an embedding vector for one particular token or patch. Because the embedding vector on its own doesn’t give any information on where a token occurs in a particular bit of text, the list of embedding vectors is often modified by adding a position-encoding vector to each embedding vector.

Attention and the transformer

Sticking with text for a moment, we have our tokenizer and embedding matrix, which convert words into long vectors of numbers. We add position information to those vectors, and then feed them into a neural network, modifying them in various ways and eventually producing some output.

One weakness of this arrangement is that while the network can learn about the general similarities of various chunks of text by way of the embedding table, embedding vectors don’t include any information about what a token means in a particular string of text. The vectors for the tokens “red” and “crimson” will reflect what the model has learned about how these are used generally, and since they describe similar colors they might end up with similar embedding vectors near each other in embedding space. But this won’t give any information about what they mean in a sentence like “the harvard crimson staff were seeing red on sunday when the president shut down their office.” If we want to accurately predict the next token in a sequence of them, it would be useful if we could modify the embedding vectors to include information about the surrounding context.

One way of doing this that has proved to be very successful is called “attention.” Attention is basically a way to modify an embedding vector with information about its surrounding context. This is done by turning the embedding vector for a token into three other vectors: a query vector, a key vector, and a value vector. Each one is created by multiplying the embedding vector by some matrix. As with the embedding table, these matrices are learned: the values in them will be modified via the training process. In GPT-3, the query, key, and value vectors are each 128 numbers long.

The query vector can be thought of as representing something like “what other tokens are important for interpreting me?” The key vector can be thought of as representing something like “what sort of thing am I useful for?” And the value vector can be thought of as representing “what sort of information will I provide?” The resulting vectors are combined together in a query matrix Q, a key matrix K, and a value matrix V, each row the vector for some particular token.

To calculate attention, the query matrix Q is multiplied by the transposed key matrix K (where “transposed” means “rows turned into columns and vice versa”). This gives a new matrix, QKᵀ, whose values indicate how important each token is for understanding every other token. There are then a few mechanical steps done to this matrix. First, we divide each value in the QKᵀ matrix by the square root of the length of the key vector, to prevent the values from getting too large. When predicting the next token in a sequence, we don’t want earlier tokens to be able to “see” later ones. We correct this by what’s called “causal masking” — essentially setting the values for later tokens on earlier ones to negative infinity. We also scale the values in each row so that they’re between 0 and 1 and sum to 1, using a function called “softmax.” (If you had the vector [10 6 4], softmax would convert it to [0.5 0.3 0.2]). You can think of this as creating a matrix that tells us how much information each token should take from other tokens.

The scaled, masked QKᵀ matrix is then multiplied by the value matrix V. The output of this multiplication is a list of embedding vectors, where each vector has been modified to have information about the token’s surrounding context.

In practice, it’s common to run this attention mechanism multiple times simultaneously (this is called “multiple attention heads”). GPT-3, for instance, has 96 attention heads. Each attention head learns different matrices for calculating Q, K, and V, and might learn different things to look for in the surrounding context: you can imagine one attention head “looking” for related nouns, one looking for previous mentions of a word, and so on. (These are just conceptual examples; in practice the things an attention head learns to “attend to” don’t necessarily map cleanly to concepts like this.) The output of each attention head will get stitched together into one matrix, which then gets multiplied by another learned matrix, which mixes the information from different heads together. The resulting matrix can then get processed by the rest of the neural network.

Probably the most common way to use attention is via what’s called the transformer architecture. A GPT-style transformer consists of multiple “transformer blocks,” each one of which contains an attention step followed by a multilayer perceptron. GPT-3, for instance, contains 96 transformer blocks. In the first transformer block, the model is fed the embedding vectors for the tokens, which are then processed by the 96 attention heads, producing an output matrix whose rows are modified and processed embedding vectors. Each row of the matrix then goes through an MLP with a single hidden layer of neurons: while the attention mechanism mixes information between tokens, the MLP processes each embedding vector it receives separately, modifying them further. The output then gets fed into another transformer block, which does the same thing: running the attention mechanism over the output from the previous block (using totally new learned matrices for calculating Q, K, and V), then sending it to an MLP. This is done again and again, through all the transformer blocks, until at the end the vector for the last token is converted into a list of scores, one for each possible token, representing the model’s guess at how likely each token is next in the sequence.

Vision, language, action

A transformer architecture can also be used to process images. The Vision Transformer (ViT) base model described in a 2020 paper by a team from Google uses 12 transformer blocks, each one with 12 attention heads. One difference with the vision transformer is that there’s no causal masking: every token can “see” every other token when processed by the attention mechanism. Another difference is the output: while a GPT-style language transformer takes a string of tokens as an input and predicts the next token as an output, the initial vision transformer takes a string of image “tokens” as an input and produces a classification of the image as an output.

With a transformer capable of processing vision or text, it’s not much of a leap to design one that can process BOTH image and text: say by feeding it an image and a text query (“What is this a picture of?”) and having the model generate an answer. This is known as a vision-language model, or VLM. LLaVA (short for Large Language and Vision Assistant) is one such model. In LLaVA-1.5, an image is fed into a vision transformer, which breaks it into patches, converts those patches into embeddings, and then sends the embeddings through a series of vision transformer blocks, producing a series of vectors (one for each patch). Those vectors then get converted via what’s known as a “projector,” an MLP that changes their size and modifies their representation into something that an LLM expects. These converted vectors then get fed, along with embeddings from the input text tokens, into an LLM, which consists of another series of transformer blocks. Attention in this transformer can look at both the vectors from the image and text, modifying them based on what’s in the image and what’s in the text. The data flows through the transformer blocks, eventually getting spit out as a predicted string of text (“It’s a picture of a dog”). This basic architecture — a vision transformer, a projector that converts its output into LLM-sized vectors, and then an LLM that processes the text and image data — is common, though various models implement it in different ways.

You can similarly use a transformer architecture to drive a robot, by feeding it information on robot state along with images and text, and having it output a series of robot actions. This is called a vision-language-action model, or VLA. The architecture for the π0.5 VLA is shown in the image below:

Image data from the robot’s cameras, text instructions, and data representing the robot’s state are all converted into token embeddings via various encoders. These then get fed into a language backbone, which consists of a series of 18 transformer blocks. (Interestingly, these have multiple attention heads, but ONLY for the query matrix: the key and value matrices are the same for each attention head. This is called “multi-query attention.”)

Running in parallel with this vision-language model is another transformer, the action transformer, which also consists of 18 transformer blocks (the same as the vision-language model). This action transformer gets fed an input of token embeddings representing essentially 50 random actions. These make their way through the transformer, and at each attention step for each action vector the action transformer looks not only at the other action vectors, but at the vectors from the images, text, and robot state at the corresponding block in the vision-language model. It then modifies those action embeddings, moving them away from random actions based on the information that it obtains from the text and image vectors. As the action vectors make their way through the action transformer, they continue to be modified using information from the processed image, text, and state vectors. The output of the action transformer then gets used to correct the input “noise” actions, and the whole process runs again. After 10 iterations, the output of the action transformer is, hopefully, a useful sequence of 50 actions. (The model processes 50 actions at a time because outputting a single action, and then running the entire model again for the next action, would make the robot move very slowly. This is called “action chunking.”) These actions take the form of joint-angle or gripper-position targets, which then get fed into a traditional robot controller that converts them into motor torques to actually move the robot.

This system was assembled from various pre-existing components and then trained together. The transformer that converts image patches into embedding vectors is called SigLIP, a 400-million-parameter model first created in 2023 and trained on pairs of images and text. In 2024, Google combined SigLIP with an existing, 2-billion-parameter LLM called Gemma (which had been trained on 3 trillion tokens of text) to create the PaliGemma VLM, which was then further trained on over a billion training examples. These examples included tasks like captioning images and answering questions about them. To create the π0.5 VLA, Physical Intelligence took PaliGemma, combined it with a system for converting robot actions into tokens (which they call FAST) and a newly created action transformer (a 300-million-parameter model), and then trained the combined system on various robotic manipulation tasks (as well as some additional image description tasks).

Conclusion

Before looking into this I knew, vaguely, that robot AI models were similar to the models used for AI chatbots that I was familiar with. But I was struck by how much overlap between the two is, with VLAs literally using LLMs trained on text as a major building block, and the basic transformer architecture from LLMs being applied to generate robot actions. This doesn’t necessarily mean that we’ll see the same sorts of rapid improvements with robot capabilities that we’ve seen with AI more generally, but it does make me think it might be possible.

Robots and the AI that controls them is something that I expect to become increasingly important, and it’s a subject I plan on spending more and more time on, with more and deeper looks at the technology and how it’s progressing.

Welcome to the reading list, a weekly roundup of news and links related to buildings, infrastructure, and industrial technology. This week we look at ASML’s lack of European sales, a California distributed power plant, Starlink as a planetary barometer, and more. Roughly 2/3rds of the reading list is paywalled, so for full access become a paid subscriber.

No essay this week, but a more involved piece about how the AI that powers modern robots works will be out next week.

Housing and Cities

Following the passage of the 21st Century ROAD to Housing Act, manufactured home producer Cavco has unveiled a HUD-code manufactured home without a permanent steel chassis. [GlobeNewswire]

Related, both Texas and Florida passed laws requiring cities to allow the placement of manufactured homes. As of September 1st Texas’ SB 785 requires cities to allow “by right” placement of manufactured homes (meaning that there’s no discretionary approval or review process) in at least one residential district. [Smart Cities Dive] And as of July 1st Florida requires “qualifying off-site constructed” homes to be allowed by right in any zoning district where single family homes are allowed. [HousingWire]

Austin’s recent decline in housing prices apparently isn’t only due to dramatic increase in supply: a fall in demand in 2023 and 2024 also contributed. [Free Range]

Kalshi has a prediction market for whether a permitting reform law will get passed in 2026. Odds are up to almost 80%. I didn’t look close enough at this, this market already resolved in July with “yes”! [Kalshi]

Why young Europeans can’t afford to buy homes. “The OECD data covers the average across 20 EU countries. For example, 60% of people born around 1975 owned a home by age 32. This fell to 58% for the 1980 birth cohort and to 52% for those born in 1985.” [Euronews]

A Census map showing the change in housing for US states between 2020 and 2026. Not much terribly surprising here, though I didn’t expect Washington State to be among the states with a large housing increase. [Census Bureau]

Trump wants his Washington DC arch to be equipped with drones and snipers. [NYT]

Manufacturing

ASML, the European manufacturer of semiconductor lithography machines, says that it doesn’t actually sell any machines in Europe currently because no new fabs are being built there. [NL Times]

The US passed Japan to be the second largest installer of industrial robots in 2025, with 38,400 newly-installed robots against Japan’s 36,200. China is still far ahead of both with 354,000, almost 10x the US. [IFR]

From the annals of the ongoing US power transformer shortage: Hitachi Energy plans to build a $500 million plant in Mississippi that would double its capacity to produce power transformers in the 10 to 160 megavolt-ampere range. [Utility Dive]

China’s progress in chipmaking machine market share over the last year. Looks like maybe a net trend upward in marketshare-weighted terms, though a pretty small one. [CSET]

A teardown of two DJI drones, aimed at shedding some light on how it achieves its impressively low manufacturing costs (and what the US might have to do to compete). “By our estimates, DJI not only sustained exceptionally high hardware margins but also grew them from 72% to 79% between 2019 and 2023. For comparison, Apple, a $4.6 trillion dollar company famous for its supply chain, achieved a 53% gross margin on the iPhone 15 Pro Max.” [Arena Physica]

The death of the small car in the US. Between 1981 and 2025 the average weight of new cars rose from 3200 pounds to over 4400 pounds, and the fraction of new cars that were small passenger cars fell from 80% in 1975 to 34% in 2025. “As of 2026, the traditional small car market has shrunk to a handful of remaining options like the Nissan Sentra and the Kia K4. Every other manufacturer has either killed off small-sized models or upgraded them to midsized cars and crossover utility vehicles (CUVs).” [World of Systems]

Trump is working on a “massive” deal to import potash (an ingredient in fertilizer manufacturing) from Belarus, likely in the hopes of reducing potash imports from Canada. [Kyiv Independent]

Saudi Arabia launches a line of EVs, Ceer. [AGBI]

Energy

Read more

Welcome to the reading list, a weekly roundup of news and links related to buildings, infrastructure, and industrial technology. This week we look at AirBnB’s housing accelerator, rising steel prices, a resurgence of US coal power, low-head dams, and more. Roughly 2/3rds of the reading list is paywalled, so for full access become a paid subscriber.

Housing and Cities

Halina Bennet on what she’s learned from a year of covering housing for Matt Yglesias’ substack Slow Boring. “...the most consequential housing wins of the past year came from states preempting local control, not from localized enthusiasm for new building. The order of operations matters; the people who’ve led building efforts without state encouragement have often found trouble at the ballot box.“ [Slow Boring]

This week AirBnb announced a “housing accelerator” program, which is investing $250 million in “last-dollar financing” for promising multifamily housing projects. The program will also “support housing advocates tackling issues like outdated zoning, permitting delays, and complex processes.” [Airbnb]

We’ve previously noted the rising opposition to property taxes around the country, with states like Florida considering ending them entirely for certain types of property. A new paper from law professor David Schleicher takes a broad look at recent opposition to property taxes, what caused it, and what impacts it’s had. “In the last three years, a number of states have substantially reformed their property tax systems, providing huge tax benefits to owner-occupied homes and shifting the burden of paying for services like schools and police to commercial property owners (including rental apartment buildings), to other local taxes, and to state funding, which itself largely comes from sources like sales taxes and income taxes. A few, including Florida, Ohio, North Dakota, and Texas, have considered going further, either completely abandoning property taxation for owner-occupied housing or even abandoning property taxation entirely...Perhaps most notably, these property tax reforms will lead to higher housing costs, and thus constitute a substantial transfer of wealth towards people who already own homes and have already seen their net worth increase substantially in the post-COVID housing market.“ [SSRN]

Median and average home sizes aren’t actually all that much larger today than they were in 1999 (there’s been about a 10% increase), but averages don’t tell you what the overall distribution of house sizes looks like over time. If you look only at smaller houses, less than 1400 square feet, the fraction of them has fallen precipitously. [X]

A new paper from Ed Glaeser on the impacts of various transportation technologies on urban density. [NBER]

The Fed raises interest rates 0.25%, the first raise since 2023. [Fox]

Manufacturing

Steel prices are up 25% year over year, and lead times are now at 55 weeks, thanks to (you guessed it) data center demand. [Construction Dive]

Trump says that he’d be ok with Chinese companies building cars in the US. [Reuters]

Supersonic aircraft startup Hermeus gets a license to manufacture Pratt and Whitney’s F100 jet engine, which powers aircraft such as the F15 and F16. The propulsion system Hermeus is developing, which is called a “turbine based combined cycle,” combines a traditional jet engine with a ramjet (which uses the aircraft’s forward motion to compress air into the combustion chamber, rather than a compressor). [X] [Hermeus]

The US house of representatives passes a bill requiring AM radios in cars. AM radio is apparently difficult for EVs to accommodate, because the large electric currents in EVs interfere with AM radio reception. [WSJ]

Read more

If you’re reading this, there’s a very good chance you’ve seen this famous image of three iterations of SpaceX’s Raptor rocket engine. The Raptor engine was developed for SpaceX’s Starship spacecraft (the Falcon 9 and Falcon Heavy use the Merlin engine); it was first test-fired in 2016, first flew on Starhopper in 2019, and first flew on a Starship prototype in 2020 and on the full Starship stack in 2023. Since then, it’s continued to improve, going from the tangle of pipes and wires you can see on the Raptor 1 to the smooth, streamlined design of the Raptor 3, which first flew in May of this year.

The evolution is so dramatic that many folks initially believed that it wasn’t real; Tory Bruno, the then-CEO of space launch company United Launch Alliance, tweeted that there was “no need to exaggerate this by showing a partially assembled engine,” which was followed by SpaceX president Gwynne Shotwell tweeting a picture of the Raptor 3 firing successfully:

This streamlining has come alongside meaningful gains in performance, with the Raptor 3 providing about 35% more thrust than the Raptor 1.

I wanted to better understand how this evolution actually happened. What, specifically, did SpaceX change that allowed it to go from the tangle of wires and pipes to the svelte, streamlined engine on the right?

There turned out to be less detail available here than I hoped. SpaceX doesn’t publish any official Raptor schematics, and no one has done a teardown of a Raptor engine. But thanks to the occasional comment from Elon Musk, and the speculations of an army of SpaceX fans, we can get some idea of what the major changes have been.

How the Raptor engine works

The Raptor engine is a “full-flow staged combustion” (FFSC) engine. What exactly does that mean?

A rocket engine works by throwing mass (“propellant”) out of a rocket nozzle — due to Newton’s Third Law (“for every action there is an equal and opposite reaction”), this pushes the rocket in the other direction. The more mass you can throw out, and the faster you throw it, the more thrust your rocket engine will produce.

The simplest way to do this is to simply fill a tank full of pressurized gas, and vent some of the gas out. The venting gas propels the rocket, in the same way that letting the air out of a balloon pushes the balloon. This sort of rocket, called a “cold gas thruster,” is often used for making minor adjustments to a spacecraft’s position or orientation. Cold gas thrusters are found on, among other equipment, NASA’s Manned Maneuvering Unit, and the control thrusters for SpaceX’s Falcon 9 rocket.

These engines are simple and reliable, but limited; there’s only so much thrust you can practically get from a cold gas thruster. So how can we get more thrust? One obvious way is instead of a single propellant, we can use two: take a fuel (like kerosene) and an oxidizer (like oxygen) and then burn them together in a combustion chamber, creating hot gas that escapes out the back of your rocket nozzle. Now instead of just using the energy stored in the pressurized propellants, we’re also using the energy stored in their chemical bonds, and converting that energy into thrust. An engine that uses the pressure of the propellant tanks to force fuel and oxidizer into the combustion chamber is called a pressure-fed engine. The engine used for the ascent stage of the Apollo Lunar Module was this type of engine: it used nitrogen tetroxide (N2O4) as an oxidizer and a fuel called Aerozine 50, both pressurized.

But now you have a problem. To prevent the burning gases in the combustion chamber from flowing back into the fuel and oxidizer lines (starving the combustion chamber of incoming propellant), the pressure in the combustion chamber has to be lower than the pressure in the fuel and oxidizer tanks and lines. But for a compact, powerful rocket engine, we want to have very high pressures in the combustion chamber. We could deal with this by increasing the pressure in our propellant storage, but this quickly gets impractical: the higher the pressures, the thicker and heavier our storage tanks and propellant lines need to be in order to maintain them without rupture. A better solution is to store the propellants at low pressure, and then feed them through a pump that increases their pressure before they reach the combustion chamber. This is the standard way to build a powerful rocket engine, and virtually every engine used for putting a rocket into orbit uses a pump of some kind to pressurize the fuel and oxidizer.

Now we have a new problem: because of the volume of fluid they must handle and the amount of pressure they must add, these pumps require an enormous amount of power to operate. The turbopump on the F-1 engine used on the Saturn V, for instance, required around 41 megawatts of power, slightly less than the power the S8G nuclear reactor delivers to the propeller shaft of an Ohio-class submarine. One way to provide this power is with a battery — Rocket Lab’s Electron rocket has an engine with a battery-powered electric pump — but the more common strategy, employed by virtually every large booster rocket, is to use the rocket’s own propellant as a power source, burning a small amount of fuel and oxidizer to drive a turbine that in turn drives the pump.

There are different ways that a system like this can be configured. The simplest is to route a small amount of fuel and oxidizer to a turbine, and then vent the resulting exhaust. This is known as a gas generator cycle, and it’s what’s used on the F-1 engine, as well as SpaceX’s Merlin engine.

Another option is to burn some of the fuel and oxidizer to drive the pump, but then route the burned propellant through the main combustion chamber along with the unburned fuel/oxidizer. This is known as “staged combustion,” and the smaller combustion chamber used to drive the pumps is called the “preburner.” This is a more complex arrangement than a gas generator cycle, but it’s also more efficient, since the hot exhaust that drives the turbopump isn’t just wastefully vented.

For most staged combustion engines, only a portion of the propellant gets routed through the preburner: the rest gets routed around and goes directly to the combustion chamber. But SpaceX’s Raptor uses a particular type of staged combustion known as “full flow.” In a full-flow engine there are two preburners, one that drives the fuel pump and one that drives the oxygen pump. All the propellant gets routed through the preburners: in one of them, a small amount of oxidizer is burned in the presence of a large amount of fuel (leaving most of the fuel unburned), while in the other a small amount of fuel is burned in the presence of a large amount of oxidizer (leaving most of the oxidizer unburned). The fuel-rich and oxidizer-rich exhaust streams then both enter the main combustion chamber, where they’re burned together.

A full-flow staged combustion engine is very complex, and prior to the Raptor only two had been built, neither of which successfully flew on a rocket. The Russians developed an FFSC engine in the 1960s, the RD-270, but it never flew, and the US built part of an FFSC engine called the Integrated Powerhead Demonstrator in the 1990s and 2000s, but it was never developed into a full engine. (When SpaceX began working on the Raptor in 2012, it obtained some of the equipment used on the Integrated Powerhead Demonstrator.) But an FFSC engine has several advantages, one of which is (theoretically) reliability: because so much mass flows through the turbines driving the pumps, the turbines can run cooler and at lower pressure, making them (in theory) more reliable. If you’re a company betting heavily on reusable rockets, a more reliable turbine is obviously attractive.

How the Raptor evolved

SpaceX is, understandably, tight-lipped about many of the specifics of its advanced rocket technology, and there’s less official information on the exact details of how the Raptor operates than you might hope. Nobody has opened up a Raptor engine to show what’s inside, and much of the information that exists comes from Elon Musk’s tweets or his comments in Everyday Astronaut interviews.

There are, however, a lot of SpaceX enthusiasts out there, many of whom do things like obsessively photograph every Raptor engine leaving the factory, and these folks have put a lot of effort into speculating about how the Raptor engine works. The most useful output of this speculation, for me, is the various fan-made schematics that show how the Raptor engine is thought to operate. These schematics aren’t official — they’re made by SpaceX enthusiasts piecing together information from various sources — so they must be taken with a large grain of salt. But they’re still a useful starting point for getting an idea of how people think various versions of the Raptor engine have worked.

To start, let’s look at some schematics of version 1 of the Raptor. The image below was created by Elisei Maslov, a Russian propulsion engineer, in 2019.

Another useful Raptor version 1 schematic is the one below, created by NASASpaceFlight forum member “hisdirt” in December 2019. This is particularly useful because it calls out the various engine components on an actual 3D model of the engine (modeled in Revit, of all things).

You can see in these schematics the basic components of a full-flow staged combustion engine: the oxygen pump, turbine, and preburner are on top of the engine, while the fuel pump, turbine, and preburner are the assembly on the side. You can also see how the fuel flows through the outside of the nozzle before going back into the preburner — this cools the nozzle so it doesn’t melt from the heat of the rocket exhaust, and is known as regenerative cooling.

The next schematic, made by NASASpaceFlight users “Livingjw” and “HVM” in February of 2022, is from Wikipedia, and it shows version 2 of the Raptor. This is less detailed than Maslov’s schematic — it basically just shows the flow of oxygen and methane — but it shows the same basic arrangement: oxygen pump assembly on the top, fuel pump assembly on the side.

And this schematic, posted by Twitter user “TheSpaceEngineer” in January 2025, shows version 3 of the Raptor.

Assuming that these schematics aren’t grossly misleading, we can see that the major architecture hasn’t changed between version 1 and version 3 of the engine. It’s still a full-flow staged combustion engine, and still has the basic arrangement of the oxygen pump, turbine, and preburner assembly on the top, with the fuel pump, turbine, and preburner mounted to the side, with fuel being used to regeneratively cool the nozzle. And you can see that a lot of the smaller lines are the same on versions 1 and 3. Both show nitrogen lines used for purging (clearing propellant out of the system), both show fuel and oxidizer lines going to the igniters, and both have fuel and oxidizer lines carrying gaseous propellant back to the tanks to keep them pressurized as they empty (this is known as autogenous pressurization).

So what changed? If we look closely at the version 1 and version 3 schematics (remembering again that these are unofficial and speculative), we can see a few differences. Version 3 shows the smaller lines bundled together in a “common umbilical,” and if we look at photos of the Raptor 3 connected to a rocket we can see this:

The version 1 schematic also shows helium being used to spin up the fuel and oxygen turbines initially, while in version 3 these lines have been eliminated, and nitrogen is used instead. (It’s not amazingly obvious to me if this evolution is correct — some Reddit commenters on the version 1 schematic stated that helium wasn’t used — but there’s some evidence that it is.) Per these schematics, the helium used to control some valves present on version 1 is also absent on version 3, and the version 3 schematic shows no helium lines whatsoever.

The version 3 schematic also notes that a heat exchanger has been eliminated, something echoed by various folks on the NASASpaceFlight forums — this appears to be a gaseous oxygen heat exchanger near the preburner. There’s also a fuel line going to the preburner that’s present on version 1 but is absent on the version 3 schematic.

The more reliable source of changes, of course, is statements by Elon Musk and SpaceX. A 2022 NASASpaceFlight article about SpaceX’s update on the Starship progress noted that “everything from turbomachinery to chamber nozzle to electronics” was redesigned on Raptor 2, the turbopump had shrunk, and the preburner controllers “had been moved to boxes rather than being all over the engine.” An Everyday Astronaut article about the same update also notes that “many valves were combined into valve plates.” In a recent response to a September 5 tweet by Yishan Wong, the former CEO of Reddit, Musk ran down a short list of changes in the Raptor over time:

One big category here is sensors, and the various wires and cables they require. From what I understand, Raptor version 1 was in large part a development engine, and thus required a lot of extra sensors monitoring things like temperature and pressure in various parts of the engine to determine how it was behaving. As the engine got dialed in, a lot of these sensors could be eliminated (though some of them may have been moved internally).

Another thing Musk calls out is the main chamber spark igniters. Igniters, per their name, ignite the fuel and oxidizer in the main combustion chamber. These were present in version 1 of the Raptor, but by version 2 they had been eliminated. Musk doesn’t say what they were replaced with, but this change may have something to do with the fact that the hot fuel and oxidizer exhaust streams from the preburners probably need very little encouragement to combust. (A NASASpaceFlight comment from 2019 wonders if main chamber igniters are even necessary, given the high temperature of the preburner exhaust streams.)

Musk also notes that a lot of bolted flange connections were replaced by welded connections, reducing mass and joints that can leak at the expense of serviceability. This appears to be still taking place, as early Raptor 3 photos show a large bolted flange on the main body of the engine that in later photos has been eliminated:

Another substantial change on Raptor 3 is that much of the piping hasn’t been eliminated, but moved internally, by way of 3D printing many of the components (Musk has previously noted that SpaceX “has the most advanced 3D metal printing technology in the world,” and in 2024 the company licensed the 3D metal printing technology of Velo3D).

Close-up photos of the Raptor 3, in fact, appear to show 3D printing lines:

The purpose of this internalization is the removal of more parts, in particular the heat shield and fire suppression systems. The huge mass of wire and pipes on version 1 and version 2 of the Raptor required a large, heavy heat shield to protect it from the heat of the rocket exhaust. Internalizing the various propellant lines and adding internal cooling systems allows this heat shield and fire protection system to be eliminated. In fact, the biggest mass change from version 1 to version 3 of the Raptor comes from mass removal of “vehicle side” engine hardware, which is almost certainly largely the heat shield.

Conclusion

SpaceX’s Raptor engine has undergone impressive evolution, going through several major versions, a dramatic increase in performance, and a dramatic decrease in mass and external components in a short amount of time. The fundamental architecture of the engine hasn’t changed — it’s still a full-flow staged combustion engine, using the same primary components — but the various ancillary elements and components supporting its operations have been modified significantly.

It’s worth noting that while this streamlining has made the Raptor engine simpler in the sense that various individual parts have been eliminated, it was still an (as Musk notes) enormously complex task to get it to work, and there’s a huge amount of internal complexity added to the Raptor 3 that these pictures don’t show. And the problems haven’t been completely ironed out: the first launch attempt of Starship’s flight test 13 was aborted automatically by the vehicle’s flight software at T-0 (right before liftoff) this past July when several Raptor 3 engines failed to start. So the picture of several versions of the Raptor is a snapshot of a piece of technology that’s continuing to evolve.

Welcome to the reading list, a weekly roundup of news and links related to buildings, infrastructure, and industrial technology. This week we look at city permitting speeds, more efficient ground source heat pump drilling, an Amazon Air crash in Miami, cooling paint, and more. Roughly 2/3rds of the reading list is paywalled, so for full access become a paid subscriber.

Housing and Cities

Economists Evan Soltas and Leonardo D’Amico have a new paper looking at permitting and construction speeds for housing in various cities in the US. The fastest cities are done with construction before the slowest cities have even issued a permit! [X]

Another setback for California Forever, the California city being planned in Solano County. A proposal to let the planned shipyard avoid completing a CEQA-required years-long environmental review (by using an existing environmental impact report) has been punted “to an unspecified date,” effectively killing the proposal for the rest of the year. [NYT]

A recent study suggests that home values don’t decline when a wind or solar farm is built nearby. “Researchers at Ball State University analyzed Indiana home sales from 2004 through 2024, comparing prices before and after large wind and solar projects came online. They also compared homes close to the projects with those farther away. The researchers found no statistically significant negative effect on residential sale prices near either commercial wind turbines or utility-scale solar farms.” [Electrek]

An interesting building technology, supposedly used in China for residential construction: hollow prefabricated wall panels lined with steel mesh and (it looks like) plastic sheets. Once in place, the panels are filled with concrete. [YouTube]

A cool map which shows how LA got built year by year, starting in 1880. [Parcelscope]

Manufacturing

Fab2, the semiconductor equipment startup co-founded by Sam Zeloof (who famously made transistors in his parents garage) raised $500 million in funding at a $3.7 billion valuation. [X]

ASML begins construction of a new production facility in the Netherlands. “The investment will eventually add ​space for up to 20,000 workers, with a first phase targeting completion by 2029 that will add offices, logistics and a so-called cleanroom space — a controlled environment for specialist equipment manufacturing. ASML said the design of a new “Flow ​Factory” will enable it to build and assemble its giant chip-printing tools more quickly and ​efficiently.” [Reuters]

Intel has processed more than 1 million wafers using ASML’s most advanced High NA EUV machines, more than the rest of the semiconductor industry combined. [Tom’s Hardware]

Energy

A huge, nine-year long investigation to find evidence of cold fusion turned up nothing. “We searched a wide swath of the loading, stimulation, and material parameter space associated with LENR. Although we didn’t test every possible corner of the parameter space, we consider our null result definitive in the spaces we did search. We found no evidence for any variety of cold fusion phenomenon. Furthermore, we found prosaic (non-nuclear) explanations for most prior reports of “excess heat” in electrochemical cells.” [Callisto Report]

We’ve previously noted how advances in drilling technology, such as the PDC drill bit, have enabled the rise of enhanced geothermal energy (EGS). EGS requires a lot of drilling footage, and so the economics only work when drilling gets sufficiently cheap. Interestingly, another type of geothermal energy, ground source heat pumps, might also be made cheaper by a different type of declining drilling costs. Ground source heat pumps work by drilling holes into the earth, tapping the soil as a heat source/sink. A startup, Dig Energy, thinks it can make these systems cheaper by reducing the cost of drilling. “In the standard process, the borehole is drilled first, typically with a metal bit or compressor, and shored up along the way with a casing to prevent collapse before a heat exchange pipe is inserted. Dig’s system, by contrast, uses a high-pressure nozzle that sprays water at 10,000 pounds per square inch to bore through soil and rock. The approach was explored by the oil and gas industry decades ago, but never gained traction because it didn’t increase productivity in those sectors, Madden said.” [Canary Media]

Read more

I wanted to get a better sense of what the arc of construction automation looks like over time: what things had been tried when? What worked, and what fell by the wayside? To do this, I used AI to trawl through thousands of old issues of digitized construction trade journals going back to the 19th century, looking for any example of a construction task being automated.

My main takeaway is that successful construction automation basically falls into two categories. One is construction-related tasks that can be done off-site in a factory: this includes things like steel fabrication, precast concrete, window and door manufacturing, and so on. The other category is on-site tasks that are, or can be made to be, extremely factory-like: they can be done by a machine moving along a simple path, doing some repetitive motion over and over again. Historically, almost anything outside these two categories hasn’t succeeded.

Method

I started by downloading every back issue of Architectural Forum (950 issues) and House & Home (370 issues) on USModernist, and every back issue of Engineering News-Record (4,300+ issues) available on archive.org. These are trade journals that cover various aspects of the construction industry, including construction practices and methods, and cover the period from the mid-19th century up to the early 1980s. I then had Claude (Opus 5 Max) search through these issues and flag any instance of some construction task being automated. In addition to this, I also had Claude do a more general web search for examples of construction automation, which included things like looking through the archives of the International Association for Automation and Robotics in Construction (IAARC).

The purpose of this wasn’t for data analysis: I wasn’t counting the number of cases by year or by category, or anything like that. This was purely an existence survey so I could see what sorts of things had been tried and were promising enough that they made it into a trade journal.

As usual with a task like this, the results will be a function of what definitions you choose to operationalize the terms in your prompt. By “automation,” I mean something like “a machine does some substantial portion of a task without requiring human direction.” So this is distinct from “mechanization” or “industrialization,” where a machine is used to reduce the labor required to do some task, but that machine might still require continuous direction or direct operation from a human operator. So something like a power tool or a crane wouldn’t count as automation, because in both cases each individual movement/action of the machine is orchestrated by the human as the work is being done. A bricklaying machine that automatically placed individual bricks without a human operator directing it brick by brick, on the other hand, would count as automation.

Like most definitions, this one gets fuzzy around the edges. In practice most automated systems require some degree of human oversight while they’re running, and whether that oversight counts as “directing individual movements” can be something of a judgment call. So I’d consider a road-paving machine that automatically lays down asphalt “automation,” even with a human driver monitoring the machine continuously. But something like a ride-on concrete finishing machine seems much more borderline, since that seems to require a lot more human direction on a second-by-second basis.

Factory methods

If some construction task can be done outside an actual construction site, in a factory or factory-like environment, there’s a very good chance that, like any factory task, it can at least be partly automated.

An early example of this is concrete batch plants. Through most of the 19th century, concrete was mixed manually by workers with shovels, but by the late 19th century concrete mixing started to be done by machine, and during the first half of the 20th century these became more and more automated, with things like automatic cement weighing, automatic water dispensing, and so on. On the Hoover Dam, for instance, concrete batching was “completely automated.”

Steel fabrication is similar. Steel sections are, of course, a factory-produced product and always have been, and it’s not much of a leap to turn raw steel elements into building components at the factory and to partly automate those tasks. Machines to do tasks like automatically drilling holes in steel beams, automatically welding pipe seams, and automatically painting steel sections started to appear in the late 19th century and remain widely used today.

More generally, essentially any building material — steel, brick, concrete block, plywood, drywall, shingles — will be produced in large volumes in a factory and thus various portions of the production process will be automated.

Concrete, as we’ve noted, started out as a site-produced material but eventually had its production moved off-site. We see the same trajectory with other building components. Windows and doors, for instance, used to be fabricated on-site, but those are now produced in off-site factories with many of the assembly operations automated. Floor and roof trusses were similarly originally site-produced components but are now mostly made in off-site factories with varying degrees of automation. A typical truss factory will automate the cutting of the various wood sections and the pressing of the truss plates, and some truss factories have even higher levels of automation.

Precast concrete is another factory-produced building component: large concrete elements like beams, walls, and columns are manufactured in an off-site plant, then trucked to the jobsite and assembled there. Precast manufacturing doesn’t necessarily have a great deal of automation, particularly in the US, but it can be highly automated.

This is, of course, part of the reason there’s been so much historical interest in prefabricated construction: once you move the actual construction process into a factory, you can potentially automate large fractions of it, achieving (it’s hoped) labor and cost savings. While sometimes prefabricated construction is merely conventional, labor-intensive construction done inside a warehouse, other times the process is highly automated. Levitt and Sons built a highly automated prefab plant as part of Operation Breakthrough in the 1970s, and today companies such as Autovol in the US and Gropyus in Germany run partly automated plants.

On-site continuous processes

The other category of successful automation is on-site tasks that can be made factory-like: a simple, repetitive motion, done over and over again as a machine moves along a simple path.

A lot of the examples in this category involve concrete or asphalt. Concrete is a good candidate for this sort of automation because it’s a thick, viscous fluid; it can be squeezed into the shape of some particular container and then maintain that shape when the container is removed. So you have things like automatic concrete slipforming machines, which extrude a layer of concrete as they advance forward. Oriented horizontally, these can be used to make things like curbs, trenches, and sidewalks. Oriented vertically, these can be used to make vertical concrete structures (though I believe vertical operation is somewhat less automated).

We also see this sort of “automatic extrusion machine” for metal. Gutter-making machines are basically portable roll-forming machines: they take in sheet metal from a coil and shape it into gutters. Similar portable machines exist for metal roof deck and metal siding.

You also see this sort of automation used successfully in the construction of horizontal, continuous infrastructure. Roads, for instance, are paved with a series of specialized machines called a paving train. Working together, the machines work in sequence to rip up and roughen the existing surface of the road and then lay down a new layer of asphalt.

Parts of railway construction can similarly be done by a series of automated machines. Maintenance tasks — things like replacing worn-out railroad ties and tamping ballast — in particular seem to be highly automated. Since the rail is already in place, this machinery can be rail-mounted, performing its various specialized tasks as it moves along the track.

Tunneling is similar. Historically tunnels were done by hand, but around the turn of the 20th century tunneling began to be done by self-propelled boring machines.

Everything else

From what I can tell, up through the early 2000s (and even into the 2010s), almost everything that didn’t fall into the two categories (tasks that can be done in an off-site factory, and tasks that can be done in a factory-like way on-site) didn’t succeed.

Automatic bricklaying machines are the classic example of unsuccessful automation here. As I noted in a previous essay, these sorts of machines have been tried over and over again since the early 20th century, and historically none of them succeeded.

Automatic setting of bricks or concrete blocks seems like it would be the perfect construction task to automate: it’s both physically strenuous and extremely repetitive. But in practice, the task isn’t quite simple and repetitive enough to build a machine to easily do it. Spreading a layer of mortar, and then placing a brick on top of it, while ensuring that both the individual brick and overall wall remain straight and level turns out to be hard to do with a uniform, repetitive mechanical motion: in practice the bricks often end up slightly askew, requiring small, individual adjustments to the bricks to make sure that the wall is being built level. Making nice, smooth mortar joints similarly isn’t easy for a machine to do. SAM, the Semi-Automated Mason manufactured by Construction Robotics, required masons to follow behind the robot to finish the joints and occasionally level bricks that were set askew. Fastbrick Robotics’ Hadrian block-laying robot avoided these problems entirely by using a custom adhesive that doesn’t require spreading a layer of mortar.

Brick walls also often aren’t repetitive enough in the sense that they rarely consist of long, uninterrupted runs of uniformly placed brick. Walls are often short, they have window and door openings, they have joints, they have corners and jogs, and various other deviations from uniformity. A simple machine doing a purely repetitive motion, or even a robot with a greater degree of flexibility, can’t easily handle these sorts of deviations, and historically bricklaying machines needed to be manually readjusted by a person whenever they encountered one of these situations. The SAM bricklaying robot really only worked well on very long, uniform stretches of wall; because it couldn’t turn corners or even set bricks at the end of walls, it needed to be set up fresh at every new wall, which took time and effort.

This sort of difficulty is very common with construction automation. Often there’s some repetitive task that can be automated, but it’s not quite repetitive enough to make it worth setting up some machine to do it. If a person has to manually reset a machine every few minutes and move it to a new area, that’s often not any easier than simply having a person do the task. Builders tried automatic welding machines on jobsites as early as the 1930s, for instance, but you don’t see many such machines used today: buildings rarely have long, uninterrupted continuous stretches of welding that would make such a machine worth it.

Similarly, tying rebar — wrapping a thin wire around two pieces of rebar to tie them together — is simple enough to be automated, but historically it hasn’t been quite repetitive enough to field a dedicated machine for it. (In my experience, it’s often not repetitive enough to even give workers power tools that can do this.) Even today the calculus appears pretty unfavorable: Advanced Construction Robotics started offering the TyBOT in 2020, but you only ever see it demonstrated on concrete bridge decks, which have very large expanses of rebar to tie. (It also has a rebar-placing robot, IronBOT, but not only does this also require large expanses of rebar to make sense — its website states that “[s]etup requires planning and coordination due to the IronBOT’s size and capabilities” — but the actual automation is fairly thin, requiring a human to constantly feed it individual bars.)

The lack of sufficient repetition, combined with difficulty in making a construction site “factory-like,” is what doomed Japan’s push to develop construction robots at the end of the 20th century. Starting in the late 1970s, Japan undertook a major effort to develop all manner of construction robots — column-welding robots, fireproofing-installation robots, wall-painting robots, and so on. By the 1990s, over 100 different types of construction robots had been developed. As I noted in a previous essay, while these systems often worked, they generally weren’t particularly productive:

Though the robots could often perform their tasks extremely efficiently, they required a great deal of setup and teardown time, and often encountered difficulties in buildings that hadn’t been specifically designed to accommodate them (such as narrow spaces that the robots couldn’t reach). They also had trouble dealing with the somewhat unpredictable nature of a conventional construction site — their need to work uninterrupted, combined with safety limits on how closely workers could approach, tended to constrain the rest of the construction process. And even when tasks could be completed faster, it often meant that the bottleneck just shifted to some other part of the process…

To try and overcome these issues, Japan’s largest contractors tried to make jobsites as factory-like as possible. They created huge “skyscraper factories,” huge climbing platforms that would build a skyscraper floor by floor with the aid of all manner of construction robots. But this failed to solve the problem:

The long factory setup time made them impractical for smaller buildings, and even on buildings that were large enough, extensive upfront design and coordination was required to design the building for robotic assembly. In some ways, the move from single-task construction robots to fully automated construction sites just shifted where the difficulties occurred — instead of a long setup (and takedown) time for each individual task, now you had one enormous and costly setup over the entire first portion of the project.

Future

Historically, construction automation has been successful on certain types of tasks: tasks that can be moved off-site into a factory, or tasks that are simple and repetitive enough that they can be done in a factory-like manner on-site. Automation outside of these areas generally hasn’t succeeded. Often this is because the task isn’t quite simple and repetitive enough: while you can build a machine to do it, the time and effort it takes to get the machine set up and doing the work, and orienting the jobsite to accommodate it, obviates any advantage you get from the automation.

I think, however, that we’re on the cusp of this changing.

This historical pattern we’ve observed is, in large part, a technological limitation. For most of history machines have had, at best, very narrow information-processing abilities: they were limited in what they could sense about their environment, and the extent to which they could modify their actions based on that information. Often they simply did the exact same motion over and over again regardless of the circumstances. When they could modify their behavior based on environmental feedback, it was in very limited ways. Without robust information-processing abilities, automation is limited to being done in a factory (where you can break apart your task into a series of small tasks, each one done by some particular machine) or to processes that are naturally factory-like.

The history of welding automation provides an example. The earliest welding machines were fixed automation: they performed some specific welding motion — such as welding the seams on pipes — over and over again. Starting in the 1960s, welding began to be done with robots: now instead of needing to build a new machine, you could simply reprogram the robot, feeding it a different sequence of actions to take. But that sequence was still dependent on a human programmer to determine.

Over the next several decades robotic welding gradually accumulated information-processing abilities: by the 1980s a welding robot had a limited ability to determine its position with respect to the weld and modify its actions accordingly. But the scope of this ability was very narrow: a welding robot could do things like measuring the electrical conductivity through the weld to determine its position, but not anything more complex:

They couldn’t compensate for things like a clamp that impeded the path of a robot, or a part that was too dirty to weld properly, or a part that was oriented vertically instead of horizontally, or any of the other thousand things that a manual welder could automatically adjust for.

Thus even automated robotic welding required a fairly repetitive set of tasks to be worth the time and effort it took to set up.

But this technological limitation is changing, and we’re getting machines that have far more robust information-processing capabilities. Modern automation is increasingly capable of being given a set of instructions and figuring out what actions it needs to take, step by step, to achieve them. This is, in large part, due to modern AI technology: if you feed some AI model a huge amount of training data, you can get something that responds effectively in different situations — a self-driving car that navigates a route and a traffic pattern that it’s never observed before, an LLM that answers a question that it’s never seen before, and so on.

This changing constraint is reflected among the emerging set of construction automation startups. There’s a new crop of construction automation startups that are leveraging modern AI technology and are getting traction and raising large amounts of funding. Roughly half of the VC funding for construction robot startups has gone to companies founded post-2022, and most other large funding rounds for construction robotics have been post-2020. (And this doesn’t include the enormous amounts of funding that have gone into more general-purpose, non-construction-specific robots, like Figure.) It’s increasingly becoming possible to automate construction tasks that were previously resistant to it, and I only see this trend continuing.

It’s not clear to me exactly what form this will take: will we see an army of Optimus robots on construction sites capable of a broad array of tasks, or will it be a bunch of special-purpose robots like Dusty and Canvas and Rufus? I suspect that regardless of how it develops, the future of construction automation does not look like the past.

Welcome to the reading list, a weekly roundup of news and links related to buildings, infrastructure, and industrial technology. This week we look at SpaceX’s turbine blade factory, Fervo’s deal with Google, Delhi’s cheap metro, why recycling is overrated, and more. Roughly 2/3rds of the reading list is paywalled, so for full access become a paid subscriber.

Housing and Cities

Claims that increased immigration raids on residential construction jobsites are straining US homebuilders. “In some markets, homebuilders say the impact has been significant; in others, it has been an existential threat to their ability to continue operating as businesses, complete homes, and repay construction loans.” [HousingWire]

YIMBY in Practice on the different strategies states can use to pre-empt local control over housing construction in the hopes of getting more housing built, and how likely they are to succeed. “The hierarchy is clear. Whatever policy gives the state the most power over housing tends to be the best. Whatever policy gives the local governments the most power tends to be the worst. The trouble is that what’s good policy is not always good politics. State preemption bills are often the most difficult for pro-housing legislators to pass. And in the heat of negotiations, they often end up somewhere in the middle of the spectrum. Preserving local discretion may not be a good idea, but it’s often what is needed to get a bill to pass.” [YIMBY Action]

The Free Press on four decades of attempts to build a gravel road between a remote Alaskan town and an runway 18 miles away, and the environmental activists who keep opposing it. [The FP]

Surprised I hadn’t heard of these before: “Costco doors” are doors that homebuilders put between the garage and the pantry to make unloading groceries easier. [X]

Manufacturing

Not satisfied with the 100% tariffs, an alliance of several major automakers, including GM, Ford, Toyota, Honda, and Hyundai, want congress to ban import of Chinese cars completely. [Reuters] And China is exporting so many cars that it’s straining global supplies of car carrier ships. [Freightwaves]

SpaceX is apparently building a turbine blade factory to help get gas turbine power plants for AI data centers online more quickly. Turbine blades are famously complex, but rockets have a lot of turbomachinery — turbo pumps are used to pump fuel into the motor — so this isn’t a completely wild jump. “According to the outlet’s account of the Morgan Stanley report, the foundry would supply single-crystal nickel superalloy parts to both the turbopumps powering SpaceX’s Raptor engines and to natural gas power turbines.” [QZ]

Energy

The US continues to install a lot of battery storage: between April and June of this year, US grid energy storage capacity increased by about 10%. [Canary Media]

Also on the subject of batteries, consumers are starting to use home battery storage for energy arbitrage: buying power off the grid when it’s cheap, and selling it back when it’s expensive. [Bloomberg]

Enhanced geothermal startup Fervo signs a deal to supply 396 megawatts of power to Google for a potential Utah datacenter. [Fervo] (I haven’t been tracking Fervo since the IPO, and was surprised to see its stock down about 50% since it went public.) [Yahoo]

While US shale oil output was growing steadily for many years, for most major US shale basins production has now flattened out. [X]

Nuclear reactor startup Oklo had a hybrid power project (combining nuclear, fuel cells, ans gas generation) dropped from a PJM interconnection planning process; PJM claims that that the “company never showed the project could ride through a sudden drop in grid voltage.” Oklo has filed a complaint with FERC to get the project reinstated. [Utility Dive]

Read more

This is my third and final essay looking at the 21st Century ROAD to Housing Act, and how it’s likely to affect housing supply. As I’ve noted in my previous two essays, this is a sprawling piece of legislation, with 59 sections spread across 12 different “titles.” So far we’ve looked at Titles I, II, and III, which together contain 24 individual sections. In this post, I’ll look at the remaining nine titles of the package. There are a lot of sections (35!), but many of them don’t apply to housing at all, and as with earlier titles the major potential impacts are concentrated in a small number of sections (chiefly sections 501, 502, 1001, and possibly 504).

Title IV — Accessing the American Dream

This title contains five sections, and while all of them are related to housing in some way, none of them are likely to have a major impact on housing supply.

Section 401 — Creating Incentives for Small-dollar Loan Originators and Section 402 — Small-dollar Mortgage Points and Fees. These two provisions require the Consumer Financial Protection Bureau (CFPB) to study how mortgage loan compensation practices (401) and points and fee thresholds (402) affect the availability of small-dollar mortgages. As we noted when we looked at section 105, small-dollar mortgage availability is a real problem, but all this requires is for the CFPB to study the problem. Conceivably this section could have some impact down the line if these studies inspire further action, but on its own this does nothing except create some reports.

Section 403 — Appraisal Industry Improvement Act. This section loosens various restrictions on appraisers (such as what credentials FHA appraisers need to have), changes some education requirements, and provides some grants for training more appraisers. These changes might help expand the effective size of the appraiser workforce, making it easier or cheaper to get an appraisal (and potentially reducing disparities in home valuation), but it’s not something likely to have any real impact on housing supply.

Section 404 — Helping More Families Save Act. This section modifies the Family Self-Sufficiency (FSS) program, a sort of savings program administered by HUD. When a family is living in HUD-subsidized housing, they pay 30% of their income in rent, with HUD covering the balance. So if rent for an apartment is $1,500 a month, but 30% of a family’s income is $800 a month, HUD will cover the remaining $700.

Normally, if the family’s income increases, they will still pay 30% of their income, so any rise in income also increases rent. With the FSS program, if a family’s income increases, the extra money paid in rent gets put in an account, which then gets returned to the family after several years if they meet certain requirements (such as being employed). This change creates a pilot for a modified version of this program that removes some of these requirements, and makes the program “opt-out” rather than “opt-in.” So this is in essence an experiment to test how a modified version of this savings program might work, but not something that will have any sort of housing supply impact.

Section 405 — Choice in Affordable Housing Act. This section modifies inspection requirements for federally subsidized housing, basically eliminating the inspection required before a family moves into a unit if it has been inspected in the previous 12 months as part of certain other federal housing programs. It also allows new landlords to request advance inspections. This eliminates some of the hoops that need to be jumped through when a family moves into a subsidized unit, but it’s not something that will have any housing supply impact.

Title V — Program Reform

This title contains five sections, all of which modify existing housing programs in some way. Out of Titles IV through XII, this is the only one that has some provisions that might meaningfully increase housing supply.

Section 501 — HOME Investment Partnerships Reauthorization and Reform Act. This section reauthorizes the HOME Investment Partnerships program and makes some tweaks to how it works. HOME is sort of like Community Development Block Grants (CDBG); it’s a big pot of money ($1.25 billion in 2025) that HUD gives to various jurisdictions to use on various projects. The difference is that while CDBG has historically not been used to build new housing (until changes in this bill), HOME has specifically been funding for affordable housing. As I understand it, HOME funding often gets used to bridge the funding gap on LIHTC projects.

The reauthorization is interesting: technically HOME’s authorization expired in 1994, but it has had funds appropriated every year since then anyway, so the actual impact of the reauthorization is not amazingly clear to me.

The biggest tweak in how it works now is the addition of some NEPA exclusions. Infill construction and projects of up to 15 units (up from four units previously) will now be exempt from NEPA review. Other tweaks include allowing HOME money to be used for certain non-housing things (such as sewers or sidewalks near HOME projects), changing the income requirements for subsidized tenants, and so on.

My takeaway is basically the same as the NEPA modifications in section 206 of Title II — a nice bit of streamlining that probably does not move housing supply much:

It doesn’t change the amount of funding for these projects, but it could make some of that funding stretch slightly farther if folks don’t need to take the time and effort to do an environmental review. And it also might just make folks more willing to do these sorts of projects. But the bigger impact is probably just significantly reducing the permitting difficulties for HUD-funded housing projects, making it easier to build them.

Section 502 — Rural Housing Service Reform Act. This section makes a bunch of tweaks to the USDA’s Rural Housing Service program. As we noted previously, the USDA has historically had significant involvement in rural housing construction and since 1950 has funded the “construction, purchase, or repair of over 5.5 million rural housing units.”

The list of changes section 502 makes to this program is long and complex (this section goes on for 10 pages of text in the act), but as far as I can tell the most consequential change is to how some USDA-subsidized rural rental housing is funded. These units were subsidized in two ways. First, the USDA would provide a subsidized loan for the actual construction of the rental property (this is called a Section 514 or Section 515 loan, depending on the type of housing). Second, the USDA would cover the difference between the subsidized rent (typically 30% of a tenant’s income) and the full rent of the unit.

Historically, these two sources of funding were coupled: when the mortgage was paid off, the USDA would stop subsidizing rents. Section 502 of the ROAD to Housing Act makes it possible to decouple these funding streams: once the mortgage ends, the rent subsidy can remain in place. (There has been a pilot program for this since 2024, but it was initially limited to 1000 units.)

There’s a fairly large pool of units that this change will apply to: the USDA has around 400,000 apartment units; 80% of them rely on rent subsidies; and on the order of 137,000 units will have mature mortgages by 2034. So this change will keep a lot of subsidized housing units around.

For most of these units, “keeping them around” means “keeping them subsidized,” not “keeping them in the housing stock.” In the absence of this legislation, for most of these units the owners would probably either refinance the mortgage before it expires (thus keeping the rent subsidies) or convert the units to unsubsidized housing. But some fraction of them might have their mortgage expire, lose their subsidized renters, and then stop being maintained, so this change could potentially keep a chunk of these USDA housing units from falling out of the housing stock.

Section 503 — Incentivizing Local Solutions to Homelessness. This section modifies a federal homelessness program, Emergency Solutions Grants (ESG), which is designed “to assist people with quickly regaining stability in permanent housing after experiencing a housing crisis and/or homelessness.” ESG gives jurisdictions money for various homeless programs/efforts, but has constraints on how much money can be spent on certain things. This change allows someone to request a waiver for some of those constraints, specifically “emergency shelter activities and street outreach.” Nothing here will affect housing supply at all, unless you have a very expansive definition of housing that includes homeless shelter beds.

Section 504 — Reforming Disaster Recovery Act. This section authorizes the CDBG Disaster Recovery program for three years and makes some tweaks to it. The disaster recovery program, as the name suggests, is a chunk of CDBG funding allocated for disaster recovery. Until now this program has operated somewhat unofficially: it’s just been money allocated via the CDBG mechanism that’s specifically earmarked for disaster relief. This change makes tweaks to how the program works, which will do things like allowing funds to get out the door more quickly after disasters and allow for better long-term planning about disaster recovery. The main impact on housing is probably letting damaged or destroyed housing get rebuilt somewhat more quickly than it might otherwise be following a disaster, but it’s hard to estimate what the magnitude of that might be.

Section 505 — New Moving to Work Cohort. This section authorizes a new Moving to Work cohort. Moving to Work is a program created in 1996 that gives public housing authorities more flexibility for things like the way they use federal funding and the way they calculate subsidized rents. It’s designed to “[u]se Federal dollars more efficiently, help residents find employment and become self-sufficient, and increase housing choices for low-income families.” As of 2020, there were 39 agencies in the program. This section authorizes another 25. There’s probably not much impact on housing supply here; this basically just gives more flexibility for the use of existing federal money.

Title VI — Veterans and Housing

This is a short title aimed at various veteran housing issues, none of which are likely to have much impact on housing supply.

Section 601 — Military Service Question. This section adds a disclosure to the standard residential mortgage application form (the Uniform Residential Loan Application) that veterans might be eligible for a VA home loan. This is conceivably a large amount of loans — one VA lender estimated that 58,000 possible VA home loans went “untapped” in 2024 — but it doesn’t really have any impact on housing supply. (Since a VA home loan is effectively a demand subsidy, if anything this would tend to push house prices up.)

Section 602 — Housing Unhoused Disabled Veterans Act. This section modifies HUD’s Veterans Affairs Supportive Housing program, which is a program that provides housing vouchers for veterans. It removes certain disability benefits from the income calculation to determine if someone qualifies for the program.

Section 603 — Veterans Affairs Loan Informed Disclosure (VALID) Act. This section changes mortgage disclosure rules to make it easier to compare VA loans to other types of loans.

Title VII — Oversight and Accountability

This is another short title, containing four sections that mostly modify various reporting requirements. Nothing here is likely to affect housing supply.

Section 701 — Requiring Annual Testimony and Oversight from Housing Regulators. This section requires HUD to testify before Congress annually.

Section 702 — FHA Reporting Requirements on Safety and Soundness. This section requires HUD to provide monthly reports on the FHA insurance fund’s capital ratio.

Section 703 — United States Interagency Council on Homelessness Oversight. This section requires the United States Interagency Council on Homelessness (USICH), whose goal is to “coordinate the federal response to homelessness,” to provide an annual progress report.

Section 704 — Appraisal Modernization Act. The only part of this title that’s not purely about reporting requirements, this section requires lenders for federally backed mortgages to have procedures if a consumer wants a second appraisal or a reconsideration of a previous appraisal. It also requires the GAO to report on the feasibility of a public appraisal database.

Title VIII — Accountability, Coordination, Studies, and Reporting

This is another short title that’s focused on making it easier for various government agencies to coordinate. Nothing in this title will likely affect housing supply at all.

Section 801 — HUD-USDA-VA Interagency Coordination Act. This section directs HUD, the VA, and the USDA to look for ways to coordinate their various housing programs.

Section 802 — Streamlining Rural Housing Act. This section directs HUD and the USDA to coordinate their environmental review processes. Another nice, if very minor, bit of environmental streamlining, but nothing that will move housing supply.

Section 803 — Improving Self-Sufficiency of Families in HUD-Subsidized Housing. This section directs HUD to study work requirements for public housing authorities participating in Moving to Work.

Section 804 — GAO Studies. This section requires the GAO to study various things, such as barriers to affordable housing access, housing units located near EPA “National Priorities List,” and “challenges relating to heirs property.”

Section 805 — Improving Public Housing Agency Accountability. This section basically creates various reporting requirements for certain public housing agencies.

Title IX — Strengthening Community Banks’ Role in Housing

This is a longer title — nine sections — that’s not about housing at all. Every section in this title is about various banking regulations, most of which are about helping out small, regional banks. Since this title doesn’t look at housing at all, I’m skipping it.

Title X — Home-Ownership for Main Street America

This title just has a single section, 1001 — Homes Are for People, Not Corporations, which restricts the purchase of single-family homes by large institutional investors (companies that directly or indirectly own 350 homes or more). This is a policy that’s become very popular on the right and the left, and there’s some evidence that these sorts of institutional purchases might push up home prices.

However, in practice this section probably won’t actually restrict these sorts of purchases all that much. This section includes a number of exceptions that allow an institutional investor to purchase single-family homes, some of which are likely to be relatively easy to comply with. The easiest is probably a carveout that allows these purchases if an investor has a “program to boost homeownership,” the requirements of which are modest:

• The investor needs to report on rental payments to consumer reporting agencies.

• The investor needs to give renters the first right of refusal and a 30-day window before they sell the home.

• The investor may provide “meaningful financial support,” including price concessions, for renters who want to purchase the house from the investor.

This last requirement could potentially be more burdensome, but the word “may” suggests that it will be optional or only required in certain circumstances. Assuming that compliance costs are not that high, this section probably won’t actually influence investor demand, and thus housing supply, in any meaningful way.

(This section previously would have sharply discouraged build-to-rent housing from large investors, but in the final version build-to-rent was added as an exception.)

Title XI — Central Bank Digital Currency

This title has just one provision, which temporarily “prohibits the Federal Reserve from establishing a digital currency.” No housing impact.

Title XII — Miscellaneous

This title is just administrative clarifications: one section here clarifies that no funding is authorized, and another clarifies that if one section of the act is found invalid, it doesn’t invalidate any other sections.

Conclusion

These nine titles contain most of the sections of the legislation (35 of the 59 total), and most of the actual legislative text. But most of the changes it makes seem likely to be either minor, unrelated to housing supply, or both. The sections with the biggest potential impact on housing are probably 501 (HOME Investment Partnerships reauthorization and NEPA exclusions), 502 (USDA rural housing funding), 1001 (ban on institutional investors buying single-family homes), and possibly 504 (CDBG disaster relief changes which might allow funds to get out the door quicker).

Once again, for most of these the question comes down to incentives and how strongly the various sections of the act modify them: for 501, I suspect the NEPA exclusions for HOME aren’t a strong enough incentive to change much about how many of these projects get built, for 1001 I think that the “ban” on institutional investors owning single-family homes has enough exceptions that the incentive for institutional investors to purchase them isn’t weakened all that much. With 502, the incentive seems stronger to me — keeping rent subsidies might plausibly make it worth keeping housing units around that might otherwise not be maintained — but it’s not easy to tell.

Welcome to the reading list, a weekly roundup of news and links related to buildings, infrastructure, and industrial technology. This week we look at sinking London homes, US magnesium manufacturing, the World Humanoid Robot games, catastrophic flooding in Nepal, and more. Roughly 2/3rds of the reading list is paywalled, so for full access become a paid subscriber.

Housing and Cities

New home sales are down 6.3% year over year. [Calculated Risk] And the median new home price is down to $393,800, the lowest since July 2021. [HousingWire]

As of a few weeks ago there’s a new tallest timber building in the world, for certain definitions of “timber building.” “Designed as the Sydney headquarters for technology company Atlassian, the 180-metre-high office tower topped out earlier this week. With a hybrid structure of timber, steel and concrete it overtakes the 86.6 -metre-high Ascent tower in Milwaukee, USA, by almost 100 metres, to become the world’s tallest hybrid timber tower.” [Dezeen]

Salim Furth at Market Urbanism looks at how various jurisdictions will be affected by the Build Now Act, the ROAD to Housing provision that tethers a portion of CDBG funding to housing growth rates. I noted in my look at ROAD to housing that the way this was calculated was weird, and Furth confirms that as written Build Now has some consequences that aren’t ideal. “The biggest takeaway is that most places are ineligible. The second takeaway should be that the law ought to be amended to better match what proponents say it’s meant to do: provide more federal funding to support places that are growing, even if those are in cheaper parts of the country or experienced a nasty snowstorm 2 years ago.” [Market Urbanism]

London homes are cracking as the ground sinks beneath them, thanks to subsidence caused by recent heat waves. “Insurance claims tied to subsidence — a phenomenon associated with bouts of hot, dry weather that shrink the soil and destabilize the foundations on which buildings stand — hit a record last quarter, according to data provided by the Association of British Insurers.” [Bloomberg]

A good essay on how the financialization of real estate — specifically, the rise of distant owners without any particular interest in making an area beautiful or livable — harms the built environment. “All the qualities that give a place charm or loveliness are ones that are best stewarded by people who live there. Someone who owns the plot from afar, without even visiting, can never understand the subtle details that give it life. And the middleman property developer or manager just will never care.” [New Atlantis] See also my essay on why skyscrapers became glass boxes.

Manufacturing

Following a 50% US tariff on various Canadian goods, Canada responded with tariffs up to 50% on various US aluminum and steel products. [Manufacturing Dive]

John Deere, which manufacturers earthmoving equipment in addition to tractors and farm equipment, has full order books for its construction division for the rest of the year thanks to data center demand. [Manufacturing Dive]

EV manufacturer Polestar, which will now be banned in the US due to recently-created restrictions on Chinese-owned car manufacturers, says that it was blindsided by the decision. “Polestar also said the Commerce Department had signaled it would be allowed to stay in the U.S., just like its corporate cousin, Volvo Cars. Polestar and Volvo share the same Chinese majority owner, Zhejiang Geely Holding Group, and two of their models are built on the same production line in South Carolina, with nearly identical hardware and the same software inside.” [WSJ]

Read more

The SPR has gotten a great deal of attention lately, as the Trump administration has steadily delivered oil (around 120 million barrels as of this writing) from the reserve in response to Iran closing the Strait of Hormuz. Much of this attention has centered on technical factors related to how much oil can safely be released from the reserve: various folks have claimed, for instance, that the minimum safe operating level of the SPR is 300 million barrels of oil (a level that it went below the week of August 7), and below that damage to the storage facility can occur.

It’s not clear if the specific claim about the 300 million barrel level is true, but due to the way the SPR was constructed, it does indeed have limits on how much oil can be removed from it and how many times it can be emptied. Given the ongoing importance of the SPR, it’s worth understanding how it was built and how it works now.

Origins of the Strategic Petroleum Reserve

Almost as soon as oil became strategically important to the US, people began floating the idea of stockpiling it for emergencies. The first strategic stockpiles of petroleum in the US were created for the military: a Naval Petroleum Reserve (NPR) was created in 1912 by setting aside land containing oil deposits in California, and by the 1920s there were four NPR sites across the country.1 During WWII, Secretary of the Interior Harold Ickes advocated stockpiling petroleum “for military and civilian needs,” and was appointed president of the Petroleum Reserves Corporation, a government corporation that tried (but failed) to gain control over Middle East oil fields for emergency reserve purposes. In 1952, President Truman’s Materials Policy Commission proposed maintaining an emergency oil reserve, and President Eisenhower suggested creating one following the disruption of oil imports during the Suez Crisis.

But it wasn’t until the 1973 OAPEC oil embargo that the idea for a petroleum reserve for both civilian and government use stuck. By the early 1970s, the US’s demand for petroleum was increasingly outpacing domestic production, making the country more and more dependent upon imports; between 1963 and 1970 the US’s surplus oil capacity (how much domestic production could be boosted in the event of a crisis) had fallen by 75%, and between 1967 and 1973 US oil imports more than doubled. When OAPEC declared an oil embargo in October 1973, the impact on the US was immediate; gas prices jumped by over 40%, and widespread gasoline shortages occurred.

Even before the embargo, in response to tightening oil supplies and frequent threats from Middle Eastern countries to withhold oil, the Nixon administration commissioned a study on the impact of a major oil supply disruption from the National Petroleum Council. The council’s report, released a few months before the embargo, suggested that a large petroleum reserve could act as a buffer for import shortfalls. It considered several possibilities for storing the oil, including simply leaving known reserves in the ground and aboveground steel tanks, but concluded that the most cost-effective method would be underground storage in caverns hollowed out from large salt deposits known as “salt domes.” A subsequent report by the newly created Federal Energy Administration, formed following the oil embargo (and which later became the Department of Energy), likewise suggested that underground salt dome storage was the most economical way of storing large quantities of petroleum.

This suggestion became a reality in 1975, when Gerald Ford signed the Energy Policy and Conservation Act. Among the bill’s provisions was a mandate to establish a petroleum reserve of up to 1 billion barrels to “diminish the vulnerability of the United States to the effects of a severe energy supply interruption.” Construction on the SPR began in 1977.

Constructing the SPR

When planning the SPR, several options for storing the petroleum were considered, including using large numbers of idle oil tankers. But ultimately it was decided that the best option was what the National Petroleum Council and the Federal Energy Administration had both recommended: storing the oil in large, underground salt caverns.

Salt domes are enormous underground salt structures, potentially miles deep and miles across, formed when geologic forces push underground deposits of rock salt up through the rock above. They were first discovered in the mid-19th century, and began to be surveyed extensively in the early 20th century after oil was discovered near a salt dome at Spindletop in Texas: as the salt pushes up into the rock above, it can create pockets around the dome where oil and gas can accumulate. Because rock salt is largely impermeable to oil or gas, hollowed-out caverns within salt domes were a potentially attractive place to store large amounts of crude oil.

These caverns can be created by a process known as “solution mining”: drilling a well down into the salt dome and pumping water into it. The water dissolves the salt, and the resulting salty brine gets pumped back out, leaving behind a void where the salt had been. Mining rock salt by using water to dissolve it had been practiced in Europe for hundreds of years, but modern solution mining — drilling a well into an underground salt deposit and pumping water into it — first began around the 1860s–1870s in New York.2 By the 1970s, solution mining had become a widely used method for mining salt. Companies like Morton Salt used solution mining to obtain salt for food products, and Dow Chemical used solution mining as a source of chemical feedstocks.

Initially, the underground caverns created by solution mining were simply abandoned. But by the mid-20th century, they had begun to be used for underground storage. In the 1940s Canada first considered solution-mined caverns for underground storage, and beginning in the 1950s it had become a popular method of storing liquefied petroleum gas (LPG) in North America and Europe. By 1976, private firms in the US had created over 900 solution-mined caverns, with a total storage capacity of 300 million barrels. For storing unprecedented amounts of petroleum, solution-mined salt caverns were the obvious choice, and that’s what was chosen to create the Strategic Petroleum Reserve.

The Texas-Louisiana Gulf Coast was chosen as the location for the reserve: not only did the area have numerous salt domes that could be utilized for storage, but it was close to existing oil and gas infrastructure, making it easier to fill the reserve and draw it down when required.

Construction would be split into three phases. In Phase I, several existing salt caverns would be converted into crude oil storage at five sites along the Gulf Coast:

• Four caverns at Bryan Mound, in Texas.

• Five caverns at West Hackberry, in Louisiana.

• Six caverns at Bayou Choctaw, in Louisiana.

• Three caverns at Sulphur Mines, in Louisiana.

In addition to these solution-mined caverns, the SPR would also use a conventionally mined salt mine at Weeks Island in Louisiana. Collectively, the Phase I sites would give the SPR 248 million barrels of capacity. And because the caverns already existed, they could begin to be filled with oil quickly.

In Phase II, 29 new solution-mined caverns would be created at these sites, adding another 290 million barrels of storage capacity. In Phase III, a new site would be added at Big Hill in Texas, which would bring the total storage capacity to 750 million barrels.

Due to political pressure to get oil stored as soon as possible, the construction schedule was compressed, and the first oil entered the SPR in 1977. But the project was soon mired in setbacks. The Federal Energy Administration employees tasked with overseeing the project had little experience in managing large construction projects, and cost control proved difficult. Material and equipment specifications were created before plans were fully complete, creating problems when the selections later proved to be unsuitable. Equipment such as pumps was often chosen because of availability, not because it was best suited for the job, resulting in breakdowns and costly workarounds. Contractors taking advantage of management inexperience engaged in “extensive fraud”; used valves were sold as new, and drill bits were purchased, stolen, and then re-sold back to the government. One contractor was indicted for stealing 18 truckloads of equipment worth $427,000; another was accused of making off with $8 million worth of petroleum bound for Bayou Choctaw and pumping in “hazardous and toxic wastes” instead (though the government denied that this occurred). By the late 1970s, over two dozen SPR fraud cases were being investigated simultaneously.

And while the SPR project was broadly popular, it nevertheless had its share of opponents who tried to slow down the project. The governor of Louisiana, Edwin Edwards, felt the SPR was wasteful government spending and led a “virtual one-man crusade” against the project, using various environmental protection regulations to slow it down until he was sufficiently convinced that it would benefit Louisiana. Among the concessions Edwards was able to extract were slowing down site conversions to minimize the loss of local jobs, and getting the SPR office located in Louisiana instead of the originally planned Houston.

There was also broader environmental opposition. Creating the SPR caverns during Phases II and III would require creating and then disposing of large amounts of brine; while some of this brine could be pumped back underground, the lion’s share of it would be disposed of by pumping it into the Gulf of Mexico. But local fishermen, worried about the potential impact on fish and shrimp populations, objected, ultimately forcing brine disposal pipelines to extend much farther into the Gulf of Mexico than originally planned. Securing the necessary environmental permits required an enormous amount of time and effort in the first few years of the program, and by the end of 1978 the SPR Office had completed 25 NEPA-required environmental impact statements.

There were also numerous technical challenges involved with building the SPR. Phase I of construction involved converting existing solution-mined caverns, but these caverns had not been created for the purposes of petroleum storage: some of them turned out to be unsuitable; others were workable but proved to be incapable of holding as much petroleum as originally projected due to pressure limitations. Collectively, the Phase I caverns ultimately had 87 million barrels less storage capacity than first anticipated.

Technical challenges continued during Phase II, when solution mining of the new caverns began. Historically, most solution-mined caverns had been on the order of 1 million barrels in size; a few had reached the 4 million mark. But most of the Phase II caverns would be 10 million barrels or more, far larger than previous solution-mined caverns. SPR officials learned as much as they could about solution-mining caverns for petroleum storage from studying the West German petroleum reserve, which also used solution-mined caverns and was already in operation, but the much larger scale of the SPR (700 million barrels vs. Germany’s 30 million) meant that they were operating in uncharted territory. When mining began, for instance, it was found that brine could be disposed of underground much less quickly than anticipated, and that leaching took far longer than anticipated, both of which slowed down cavern construction.

Collectively, the technical, political, and managerial challenges encountered when building the SPR drove up costs and pushed back delivery dates. In 1976, the estimated cost of the SPR through Phase II was $766 million; just two years later that had risen to $1.47 billion (~$7.5 billion in 2026 dollars). The original plan called for Phase III to be complete by 1983, with 750 million barrels of crude in the ground. But the actual fill level in 1983 was roughly half that, with many caverns yet to be completed. Between 1980 and 1993, 49 new solution-mined caverns for the SPR were created, bringing its total storage capacity up to 750 million barrels, but the actual fill level would remain at just over 500 million barrels into the 2000s.

Since its completion, the SPR has continued to evolve. The small Sulphur Mines site began decommissioning in 1992, and in 1996 the Weeks Island facility began to be decommissioned when it was realized that groundwater was slowly leaking into it. And over the last 20 years three of the remaining Phase I caverns have been decommissioned, bringing the total number of storage caverns down to 60.

Operating the SPR

At a high level, the way the SPR works is relatively simple. Each cavern in the SPR is filled up with oil, which sits on top of a heavier layer of brine beneath it. To remove oil from a cavern, water is pumped down into it, which forces oil out the top. To fill the cavern back up with oil, oil is pumped in as brine is pumped out. The caverns are thus always completely full of liquid; the only thing that changes is where the oil-brine boundary is.

But this high-level description belies a lot of complexity. We think of physical infrastructure as something relatively static and unchanging, but in the SPR various forces are constantly reshaping the caverns.

One of these forces is cavern creep. The SPR caverns are thousands of feet below ground, and the weight of the ground constantly squeezes the sides of the caverns, compressing and shrinking them over time. As the caverns are squeezed, the liquid inside the caverns is pressurized, and this pressure, in turn, tends to resist further compression of the caverns. But if that pressure is relieved, which occurs during well repair operations known as “workovers,” cavern creep can accelerate. The SPR loses up to 2.4 million barrels (~0.33%) of storage capacity each year due to the effects of cavern creep, depending in part on the number and length of workovers required.

In the other direction, removing oil from the caverns requires water to be pumped into them. And when fresh water is pumped into the caverns, it dissolves some of the salt from the walls of the caverns. Every 100 barrels of water pumped in dissolves around 15 barrels of salt; removing all the oil from an SPR cavern thus increases its size by about 15%.

Because of this, SPR caverns have a finite lifespan. After enough drawdowns, the walls of the cavern become too thin to use it for storage; they can no longer withstand the stresses and may begin to leak oil into surrounding caverns or outside the salt dome itself.

Various factors affect how the walls of a cavern wear and how many drawdowns it has available. Phase II and III caverns designed specifically to store petroleum are tall and thin, like skyscraper-sized cigars, and are spaced far enough apart to allow for significant cavern growth over time. But many of the Phase I caverns, which were originally mined for other purposes, have irregular shapes or thin cavern walls that make it infeasible to use them for multiple drawdowns. These caverns are thus known as “single-cycle drawdown” caverns. There are currently 11 Phase I caverns in use at the SPR and six caverns that have one drawdown or fewer remaining.

Cavern lifespan is also affected by how a cavern is used. When a cavern is fully emptied, wear on the cavern walls is relatively uniform, but if the cavern is subject to multiple partial drawdowns — removing a bit of oil, then filling it back up, then removing a bit more, etc. — this concentrates wear at the bottom of the cavern, resulting in “undesirable shapes” that can more rapidly reduce cavern lifespan. Conversely, it’s also possible to increase cavern lifespan by strategically emptying caverns in such a way as to wear the cavern into a more desirable shape.

Cavern wear is not simply “water dissolving the salt and the walls get thinner”; the geology of the SPR caverns is complex, and they can wear in complex ways. Changing stresses in the walls can cause “salt falls” (when a chunk of salt falls into the cavern) or open up cracks and fissures, causing leaks. Because of this, Sandia National Labs closely monitors SPR cavern conditions using tools like sonar. Each year it puts out a report estimating how many full drawdowns each cavern still has available. Because it can be hard to predict cavern behavior after repeated drawdowns, the maximum number of drawdowns it simulates is five (thus a cavern with five available drawdowns today might actually have several more than that available).

Limits of the SPR

The limited lifespan of SPR caverns, and the complex wearing forces that they’re subjected to, have led to a lot of speculation and assertion about how much oil can be safely removed from the SPR. Folks often claim, for instance, that the SPR can’t go below 300 million barrels of oil without damaging the caverns. I decided to look closely at what these limits actually are.

One limit on the SPR is that a minimum amount of “roof oil” must be kept in the caverns to prevent water at the top of the cavern from dissolving and damaging the cavern ceiling. As of 2016 this was about 12 million barrels, or around 1.7% of total SPR capacity.

Another limit you will often see mentioned is 70 million barrels: for instance, the DOE told CNBC earlier this year that 70 million barrels was the minimum amount needed to safely operate the SPR. It’s not amazingly clear what’s behind this limit, but this may be simply using the DOE’s definition of a “full drawdown,” which is when 90% of the oil in the SPR has been removed: 70 million barrels is about 10% of the SPR’s total capacity. It’s not clear if there’s an actual, physical limit at 70 million — that may be the level at which the oil-brine interface would be above the oil intake well, for instance — or if that’s simply a defined operational limit.

Above these lower limits, you see various claims about operating levels below which damage to the SPR starts to occur. Earlier this year a JPMorgan energy analyst claimed that 150–160 million barrels “must remain in place to preserve cavern stability and maintain operational flexibility.” Rapidan Energy, founded by former George W. Bush energy advisor Bob McNally, claims that this limit is at 170 million. A professor of petroleum engineering at Texas A&M, Siddharth Misra, claims that the operational limit is between 250 and 300 million barrels, below which “the risk of structural damage increases while the system loses its ability to pump oil at high speeds.” Former Biden energy advisor Amos Hochstein claims that he “know[s] plenty of people who think we can’t get near 300 [million] because physically you will damage the caverns where the oil is stored.”

Some of these claims can be verified. For instance, it’s true that draining the SPR beyond a particular point reduces the speed at which oil can be pumped out of it. Oil can only be pumped so quickly out of an SPR site, so as caverns empty the drawdown rate drops.

Additionally, because of the aging SPR infrastructure as well as the downtime needed for repairing that infrastructure, not all the oil in the SPR is available for drawdown at any given point. For instance, a recent GAO report noted that oil at several SPR sites might be unavailable during warmer months due to vapor pressure. Oil stored in the caverns gradually absorbs methane, and gradually warms up, both of which increase the vapor pressure within the oil. To be safely distributed the vapor pressure in the oil must be below a certain level, which can be achieved at a “degasification” plant. But several years ago the SPR’s degasification plant was dismantled and hasn’t yet been replaced, limiting its ability to distribute crude during warmer times of the year. More generally, the SPR has a huge maintenance backlog and has been described as being held together with “band aids.” A planned large-scale SPR renovation and maintenance program, Life Extension Phase 2, began in 2021 but was scaled back due to high costs.

There have also been various reports noting that the SPR may not have the ability to respond to a severe supply disruption below a level of 250 to 500 million barrels, and that outside of true emergencies it would be prudent to keep it above this level.3

However, no official report by the Department of Energy, the GAO, Sandia National Labs, or any other government agency mentions specific fill levels below which damage to the SPR occurs (other than the small amount of roof oil required). While Sandia puts out an annual report on the number of available drawdowns on a cavern-by-cavern basis, it doesn’t note any SPR-wide fill level limit below which damage occurs. I similarly wasn’t able to find any independent analysis or report pointing to some minimum SPR fill level and explaining specifically what drives it. So it’s not particularly clear to me what the source of claims about very high minimum fill levels is: if it’s from people talking to Sandia/DOE folks in some unofficial capacity, or if they’re inferring limits based on cavern stress data that Sandia provides but doesn’t make explicit, or something else. It’s notable to me that claims about minimum fill levels aren’t particularly consistent, and that different analysts/experts give widely different limits. It’s also notable that these claims often conflate different limits. (The Department of Energy, for its part, has denied the 300 million barrel limit.)

Conclusion

The Strategic Petroleum Reserve is the largest petroleum storage facility in the world and stores vast amounts of crude oil — over 700 million barrels — in underground salt caverns. These caverns behave in complex ways: they grow, shrink, and distort over time due to the various fluid and geologic forces that they’re subjected to, and the caverns have a limited lifespan due to water gradually dissolving the solid salt walls.

Because of these complex forces, and the gradual damage that gradually accumulates across the walls of the caverns, there’s widespread speculation and assertion that the SPR can’t be drained below some particular level without inflicting serious damage on the underground caverns.

It’s hard to know how seriously to take these claims. Many of them are made by experts, such as professors of petroleum engineering and former White House energy advisors. And there are various limits associated with removing oil from the SPR, such as the oil needed to keep water from damaging the roof of the caverns or the drop in withdrawal rates that occurs as the SPR empties. But it’s hard to find any sort of actual published estimate of these limits in a report or official document. This doesn’t mean that these claims are wrong, but it does make me wonder where the claims are coming from.

Welcome to the reading list, a weekly roundup of news and links related to buildings, infrastructure, and industrial technology. This week we look at Los Angeles’ ADUs, electrical transformer manufacturing, gas turbine power plants, how water in the Colorado River is used, and more. Roughly 2/3rds of the reading list is paywalled, so for full access become a paid subscriber.

No essay this week, but I have a very long piece in American Affairs about why US transit construction costs are so high. You can read it here.

Housing and Cities

Los Angeles built over 10,000 ADUs in 2025, 37% of all housing units built in the county. [Smart Cities Dive]

There were 511 state bills introduced in 2025 across 40 states that pre-empted county land use authority in some way, 27% of which passed. A lot of land use restrictions are downstream of the local nature of land use authority — costs of development tend to be concentrated, and benefits tend to be diffuse, so local authorities are often incentivized to restrict development. State-level pre-emption should in theory help with this. [Smart Cities Dive]

Mexico plans on making it much harder to illicitly develop along the coast, and to tear down any illegal buildings that exist. “For years, wealthy developers have treated environmental fines as simply another line item in their budgets—a minor inconvenience easily absorbed by their profit margins. Tamborrel acknowledged this problem directly: “There are companies that allocate part of their budget to pay these fines and then continue operating; offenders prefer to pay the fine rather than request a permit.” But those days appear to be ending. Profepa has already closed four real estate developments in Puerto Vallarta and promises similar action along all of Mexico’s coastal states.” [Yucatan Magazine]

Claims (with video) that Amazon’s drone delivery service dropped a package in someone’s pool. [X]

The city of Mesa, Arizona is planning to sue several firetruck manufacturers, claiming that they’re colluding to drive up prices. [Mesa Tribune]

Manufacturing

Raytheon gets a $23 billion Navy contract for Tomahawk missiles. [Manufacturing Dive]

The New York Times has a photo essay on the laborious process of electrical transformer manufacturing. It’s not amazingly clear to me why this process isn’t more automated, and whether that would help to get them built faster or if the bottleneck lies elsewhere. [NYT]

1872 is a startup founded by several former SpaceX employees that wants to build highly automated factories for making steel parts. “The company is initially focused on automating the manufacturing of steel skids—rectangular steel frames that can provide a moveable foundation for modular buildings—with the goal of supplying customers who are developing AI data centers or small modular nuclear reactors.” [Ars Technica]

Energy

There are currently 270 GW of natural gas power plants in various stages of development in the US. This is roughly equal to 20% of all current US generation capacity. [X]

FERC approves a Southwest Power Pool (a regional transmission operator) plan to reduce grid congestion by “topology optimization” — responding to congestion by rerouting how power flows through the grid, rather than redispatching power plants. MISO, another regional transmission operator, already has such a program, and estimates that it has saved them almost $100 million so far this year. [Utility Dive]

The DOE cancels three planned National Interest Electric Transmission Corridors, designations designed to make it easier to build transmission projects. The DOE stated that transmission policy “must serve the American people, not special interests or a climate-alarmist agenda.” [Utility Dive]

Read more

Welcome to the reading list, a weekly roundup of news and links related to buildings, infrastructure, and industrial technology. This week we look at home price declines, the Strategic Petroleum Reserve, The Boring Company’s $4 billion fundraise, and more. Roughly 2/3rds of the reading list is paywalled, so for full access become a paid subscriber.

Housing and Cities

California is still not building enough housing. “After years of nudging, political trolling and litigating, most cities and counties now have state-approved plans in place. But as the production numbers show, it’s one thing to plan and another to build. Almost nowhere in the state is actually seeing the new construction necessary “to meet the housing needs of all Californians,” as housing regulators have described these targets.” [CalMatters]

Following the severe buckling of a column at an office to residential conversion project in New York, the Department of Buildings inspected 180 jobsites which had a connection to the team working on that project. No hazardous structural issues “posing an imminent danger to public safety” were found. [Construction Dive]

Lennar and DR Horton, the two biggest homebuilders in the US, have been reducing their prices for the last several years. “Lennar has reduced its average selling price from $511k to $377k, a nearly 25% decline (inclusive of incentives). DR Horton has cut from $415k to $366k, a 12% decline. Four years ago, this would have been hard to imagine.” [X] This matches a broader trend of declining home prices overall since 2022. [FRED] Interestingly, the Case-Shiller index hasn’t fallen. [FRED]

A cool map showing how much people in various cities in the US actually walk, compared to what their neighborhood walk score is. [National Walking Atlas]

Manufacturing

As part of his plan to tile the sky with orbital data centers, Elon Musk wants the Terafab project to eventually be able to produce 1000 gigawatts of computing power each year. Per Musk, currently the world produces about 20 gigawatts of computing power each year. [Manufacturing Dive]

The White House released a report accusing China of a “Great Transshipment Scam” — exporting nearly-finished goods to other countries and performing a very small amount of work there to mask their country of origin and get around tariffs. [SCMP]

The Wall Street Journal on how China is increasingly a manufacturer of not just consumer goods but of high-value industrial equipment. “The transformation is threatening the economic moats of advanced-manufacturing economies such as the European Union, Japan and South Korea. Producers of chemicals, machines, batteries and other industrial goods in those economies once depended on Chinese factories as customers, but now China is a formidable competitor abroad and even in their home markets. For the first time in decades, Germany imports more advanced capital goods from China than it exports there.” [WSJ]

Apparently 100,000 of the 383,000 members of the United Auto Workers union are academics, not autoworkers? “The University of California’s academic local alone represents some 48,000 workers—more than the UAW’s total membership at the global automaker Stellantis. Harvard’s graduate students account for another 4,000.” [City Journal]

South Korean shipbuilder Hanwha, which bought a Philadelphia shipyard in 2024, wants to acquire another US shipbuilder, Austal USA. Letting a foreign shipbuilder acquire these operations, and potentially getting them running more efficiently, seems like a pretty good deal for the US. [gCaptain]

Also on the subject of shipbuilding, the White House releases a memo calling for the Pentagon to source ships from foreign suppliers, and to construct or acquire a new Navy shipyard. [Workboat]

Read more

This is the second installment of my look under the hood of the recently passed 21st Century ROAD to Housing Act, looking at what each provision actually does and trying to suss out its actual impact on housing supply. You can read my previous essay on Title I of the Act here. This time I’ll look at Title II, “Building More in America,” and Title III, “Manufactured Housing For America,” which have 13 and 4 individual provisions respectively. Some of the most consequential chunks of this legislation are in these two titles.

The sections in this title mirror the act overall: only a small number of these changes are likely to be of major consequence, and how much they increase housing stock will depend a lot on whether the constraints they modify were actually binding, which is often hard to tell.

Title II — Building More In America

This is the largest title in the act, with 13 sections, each one of which targets housing production or preservation in some way. We’ll look at each section one by one.

Section 201 — Increasing Housing in Opportunity Zones. This section allows HUD to give additional weight when awarding housing grants to projects that will be built in Opportunity Zones: areas that are economically distressed and that have tax incentives available for investing in them.

The likely impact of this section on housing supply is almost definitely zero. It just shifts which housing projects HUD decides to allocate money to. It doesn’t increase the size of the pot of money, or the number of projects that can be funded.

Section 202 — Whole-Home Repairs Act. This section establishes a HUD pilot program for providing grants to homeowners and landlords for home repairs. This is potentially a useful program — some people get funds to make housing repairs that maybe otherwise wouldn’t be able to afford them — modeled on earlier, state-level programs like Pennsylvania’s. But it seems more about increasing housing quality than housing stock: improving energy efficiency, weatherization, safety, general habitability, etc.

That said, there might be a “it’s easier to keep an existing customer than finding a new customer” logic to this. If you want to turn dollars into housing supply, in some cases a dollar probably goes farther repairing an existing home than it does building a new home. Typically homes fall out of the housing stock at a pretty slow rate, but if there’s damage (say, a leak in the roof) that probably greatly accelerates the rate at which it happens. Funding those sorts of repairs is probably pretty useful from a cost/benefit point of view. Depending on how the repair money is distributed, this could potentially be a meaningful amount of homes, though probably not a huge number. I suspect mostly this will be about addressing quality-of-life more than keeping homes from falling out of the housing stock.

Section 203 — Community Investment and Prosperity Act. This section raises the cap of how much banks are allowed to invest in public welfare projects, most of which are Low Income Housing Tax Credit (LIHTC) housing projects. The LIHTC program is a federal housing subsidy program. Each year the federal government awards several billion dollars worth of tax credits to states, who in turn award them to housing developers for completing housing projects rented to low-income families. Roughly 70,000 LIHTC housing units are completed each year.

These projects are typically funded by banks, who take essentially all the project’s equity in exchange for funding it, and get the tax credit attached to it in return. (The tax credit reduces the amount of federal income tax you owe, and thus is most useful for organizations that have a lot of tax that could be reduced.) Banks are limited in how many of these sorts of projects they can invest in at once; this section raises those limits.

I had a hard time understanding the exact mechanics of these sorts of funding requirements, but as I understand it this change is more about preventing these limits from becoming binding in the future than them being constricting right now. So this section is possibly a big counterfactual deal (in the sense that we would get fewer LIHTC units in the future without it), but not one that will boost housing supply over current levels.

Section 204 — Addition of Affordable Housing Construction as an Eligible Activity. This section allows affordable housing projects (including LIHTC housing) to be funded by Community Development Block Grants (CDBG), a pot of money HUD hands out every year for states and other jurisdictions to do various community improvements. CDBG funding is fairly large, $3.3 billion each year, and this lets up to 20% of a jurisdiction’s CDBG funding be used for housing, so it’s a potentially large amount of money for housing projects.

The mechanics of these programs and the actual binding contraints are sufficiently complex that it’s hard for me to gauge whether it will be used all that much in practice, so my error bars here are quite wide.

Section 205 — Better Use of Intergovernmental and Local Development (BUILD) Housing Act. This section basically allows some NEPA environmental reviews for HUD housing projects that previously would have been done by HUD to instead be done by some local supervising entity within the project’s jurisdiction (with HUD still needing to approve the overall environmental review). This could speed up environmental review for these projects — this sort of delegation has sped up environmental review on transportation projects by 15-30% — if the jurisdiction in question has the capability of doing this sort of review. Overall I expect a positive but marginal improvement here that probably doesn’t shift housing supply all that much.

Section 206 — Unlocking Housing Supply Through Streamlined and Modernized Reviews Act. This is another “streamlining environmental review” provision, similar to 105 and 205, and is probably the most important section in the bill so far. This section exempts a bunch of HUD activities from NEPA review (mostly administrative ones), and provides categorical exclusions for a bunch of HUD-financed construction activities: new construction or rehabilitation of multifamily projects of up to 15 units, office to residential conversions, and any infill housing construction project (where “infill” means a project in some municipality on less than five acres of land that has been built on previously).

It doesn’t change the amount of funding for these projects, but it could make some of that funding stretch slightly farther if folks don’t need to take the time and effort to do an environmental review. And it also might just make folks more willing to do these sorts of projects. But the bigger impact is probably just significantly reducing the permitting difficulties for HUD-funded housing projects, making it easier to build them.

Section 207 — Grants for Planning and Implementation Associated with Affordable Housing. This section creates a HUD grant program for jurisdictions to study and implement pro-housing strategies, such as updating zoning codes or modifying housing-related regulations. This probably addresses a real constraint — there are thousands of permitting jurisdictions in the US, and many of them probably can’t easily spare the resources to analyze and modify their zoning regulations — but it would surprise me if it were particularly binding. I suspect jurisdictions that wanted to modify their regulations to be more pro-housing would find a way to do so, though this could conceivably be enough to encourage the marginal jurisdiction to pull the trigger.

These sorts of grant programs have been created before: in 1992 as part of the Housing and Community Development Act, and more recently in the form of PRO Housing Grants (which was first implemented by HUD as something like a planning grant but evolved after congressional pressure to be more like the Innovation Fund, below), but it’s not amazingly clear to me that these have had much of an impact. Overall the likely outcome of this seems fairly minor to me.

Section 208 — Innovation Fund. This section creates a $200 million per year Innovation Fund (which will need to be funded via a separate appropriation), which will be awarded to jurisdictions which show “objective improvement in housing supply growth.” Awards can be up to $10 million. This is a great provision, directly incentivizing the thing we want (increased housing supply). The only question, like with so many provisions in this bill, is how strong this incentive will actually be in practice: how likely is a jurisdiction to try and improve its housing delivery in the hopes of winning one of these grants down the road? I would be surprised if this is enough to make most jurisdictions make major changes to their housing or zoning regulations — the amount at stake is probably just too small, and too uncertain — but, like with Section 207, it might be enough to shift the marginal jurisdiction.

Section 209 — Accelerating Home Building Act. This section allows (but does not require) HUD to award grants to jurisdictions to implement pre-reviewed housing designs, often known as “pattern books.” Instead of coming up with the architectural design for some new housing project themselves, or hiring an architect to do it, the developer can just select one of these pre-reviewed designs. The benefit of this is twofold. One, the developer saves some money on design effort (since the design has already been done), and two, the permitting process becomes smoother, since the design has already been approved by the jurisdiction.

Pre-approved housing designs have been tried in a variety of jurisdictions, and they appear to be something of a mixed bag. In some cases, they’re used enough to make up a single-digit percentage of new housing construction; in other cases, they basically don’t get used at all.

Overall, I’d classify this as another “positive but marginal” section: assuming HUD decides to fund these, it probably means the marginal jurisdiction adopts pre-reviewed designs that otherwise wouldn’t, and in some fraction of those a meaningful (but not substantial) chunk of new housing gets built using them. Money and construction time are saved, which maybe means a few projects pencil out that otherwise wouldn’t, but probably not very many. My prior is that real estate developers want a lot of control over the specifics of their projects — exactly how the buildings are laid out, exactly what the floor plans are, exactly what finishes are used where — which makes me somewhat bearish on pattern books in general.

Section 210 — Revitalizing Empty Structures Into Desirable Environments (RESIDE) Act. This section authorizes (but does not require) HUD to create a grant program for converting existing vacant buildings into housing. There’s no set amount of money for it (it will depend on what’s appropriated), but the text of the law suggests it might be in the $100 million range.

Converting existing buildings into housing is a perennially popular idea that is often blunted by the fact that these sorts of conversions are expensive: it typically costs as much or more to convert an existing building into housing than it is to just build a new building from scratch. So a grant program that can provide money for these conversions might theoretically make more of these projects pencil out.

Other than funding limitations ($100 million does not buy all that many conversion projects), the impact of this is likely limited by the fact that the given vacant building must have been deemed abandoned or unsafe for occupation to be eligible for funding. The typical conversion targets are things like low-occupancy office buildings that are still in use and safe to occupy. But this is specifically aimed at abandoned, derelict buildings. So the universe of potential projects isn’t all that large. And it’s further narrowed by the fact that this section requires the resulting projects to have the majority of their units affordable to low-income households.

Overall I think this could potentially result in a few more conversion projects, but probably a pretty small number of them.

Section 211 — Housing Affordability Act. This section raises the per-unit limits for FHA insurance for multifamily mortgages. As we noted previously, the Federal Housing Administration is basically an insurance agency: it insures mortgages from the borrower defaulting, to incentivize the lender to make loans they otherwise wouldn’t. Most of these loans are for single-family homes, but it also does a few multifamily mortgages (around 17,000 units per year).

As I understand it, FHA insurance basically makes multifamily projects charging lower rents pencil out that otherwise wouldn’t, so it’s useful for building projects targeting lower-income occupants. But FHA insurance has limits on how much it can lend for a specific project, which is calculated based on the cost of the individual units. This section raises those individual unit limits, and so increases the limit on how large an FHA insured multifamily mortgage can be. It also changes the formula for how these limits are calculated.

It’s not amazingly clear to me how much this section will affect housing supply (even understanding how FHA multifamily insurance works at all was a struggle), but I suspect it will be fairly minor. For one, this is just a very small program, and this shift probably won’t change that. For another, the FHA has overall limits on the total amount the FHA can insure, and this doesn’t change those limits. Also, these limits have been repeatedly raised in the past, and don’t seem to have changed the trajectory of multifamily housing construction. This could be another case where this is pre-emptively reducing a constraint that otherwise would have become binding, but it’s hard to know. Overall I don’t expect much from this change.

Section 212 — Rental Assistance Demonstration (RAD) Program. The RAD Program restructures how certain public housing projects are administered, converting them to the Housing Choice Voucher Program, also known as Section 8 Housing. The purpose of this is to fund the maintenance of public housing: for abstruse financial reasons, this conversion basically makes it possible for housing projects to get loans against the value of their future subsidy income streams and use those loans to fund maintenance and repair projects. This section raises the cap on the number of housing units that can be converted by 100,000, from 455,000 to 555,000 units, and also makes the program permanent (it was previously set to expire in 2029).

This will potentially let 100,000 units of public housing fund maintenance projects. Some of those units might otherwise gradually decline to the point of being unoccupiable. So this is another provision that could keep some large chunk of housing units from being removed from the housing stock. And it also can result in the construction of new housing: some chunk of RAD projects involve tearing down an existing public housing project and building a new one, also building some non-public housing units along with it. The Fulton and Elliot-Chelsea redevelopment, for instance, done through New York’s RAD implementation, is replacing 2056 public housing units with 2056 new public housing units, and also adding another 3454 new housing units. So this is potentially a big deal.

Section 213 — BUILD Now Act. This section tethers a chunk of certain high-cost jurisdictions’ CDBG funding to their level of housing growth. If the rate of increase in housing production is lower than the median of these high-cost jurisdictions, you lose a chunk of your CDBG funding (10%), and that money gets reallocated to jurisdictions where housing production rose faster than it did historically.

At a high level, it’s good to have another fairly direct incentive for jurisdictions to try to increase housing production. And the threat of taking away federal funding has been successfully used before, such as when the federal government got states to adopt a national speed limit of 55 and a drinking age of 21 by threatening to take away highway funding. This is one of the first times the federal government has been willing to penalize high-demand places that don’t build housing. But there’s a few issues with this provision. One is that the way that it’s calculated is weird: per the text, a jurisdiction that went from 0.2% growth to 0.25% growth would score higher than a jurisdiction that went from 2% growth to 2.4% growth, even though the latter is a much larger increase in housing supply.

The big question here is whether this loss is enough to incentivize jurisdictions to make jurisdictions change their behavior. 10% isn’t necessarily a huge amount (the average CDBG is only like $2 million, though for the jurisdictions in question it’s probably substantially higher), but there may be loss aversion and a reluctance to lose money that had already been allocated that might be quite motivating in practice. And small funding reductions can be motivating: the threat of highway spending reduction to induce states to adopt a 21 year drinking age was only 5% of federal highway dollars. So this could be potentially quite meaningful, though it’s hard to tell.

—

That covers all of Title II. Next up is Title III, “Manufactured Housing For America.”

Title III — Manufactured Housing For America

This title has four sections, each of which are aimed at making it easier to build manufactured homes — homes built under the national HUD code for manufactured homes instead of local building codes, and formerly known as mobile homes or trailers.

Section 301 — Housing Supply Expansion Act. This section eliminates the requirement that manufactured homes have a permanent steel chassis. It also makes some adjustments around manufactured home energy regulations.

The permanent chassis removal is a pretty big deal: It’s something that manufactured home advocates have been requesting for decades. Historically manufactured homes were required to be built on a steel chassis to maintain the fiction that these were fundamentally portable trailers, but almost all manufactured homes are placed once and then never moved. The steel chassis adds substantial cost (on the order of 5-10% of the total cost of a manufactured home), so removing it could mean a potentiallyarge savings (though the steel chassis also acts as a structural support, so a home without one might require additional structural elements that will eat into this savings).

But I’m moderately skeptical that removing this requirement will increase the number of manufactured home purchases. For one, manufactured homes are already substantially cheaper than conventional construction, so it would be surprising to me if another few percentage points of cost reduction resulted in a big uptick in their adoption. For another, it’s often claimed that the steel chassis is what makes it possible for jurisdictions to discriminate against manufactured homes, but there’s no real reason to think that this is true: It would be trivial for jurisdictions to just restrict HUD-code homes, regardless of whether they had a steel chassis or not. So I don’t think this will make it easier to place these homes than it was previously.

This change could theoretically make it easier to build manufactured housing projects from multiple modules, taking advantage the 2024 change that allows the HUD code to be used for structures of up to four housing units (HUD has specifically noted that the chassis requirement makes it harder to build multistory manufactured housing structures). But I similarly suspect that this wasn’t enough of a constraint to meaningfully alter their uptake for this sort of structure. My prior here is that in the current environment it’s pretty hard to meaningfully increase manufactured home shipments via regulatory changes: folks have been trying various ways of doing that for the last two decades but don’t seem to have had all that much success.

What is true, however, is that jurisdictions often define manufactured homes by referring to their steel chassis, at least somewhere in their regulations. Because of that, implementing this will require states to modify their regulations, so this implementation will take some time: jurisdictions have a year to make the necessary regulatory changes. So it’ll be a while before we see chassis-less manufactured homes rolling down the road. Overall, I expect this change to possibly reduce the price of manufactured homes by a few percentage points, but probably not result in a huge uptick in their use.

Section 302 — Modular Housing Production Act. This section instructs HUD to study FHA construction financing, determine whether there are financing barriers that might be preventing the adoption of modular housing, and to issue a rule addressing the issue if it seems like one is justified.

This is looking at a real problem: construction projects are financed with construction loans, and these loans typically have a “draw schedule”: the funding gets issued a little bit at a time, just enough to complete the next phase of the project. But these draw schedules often don’t work well with modular construction, which requires a lot of funding upfront (before the project even begins) to fabricate the modules. So fixing this would be meaningful.

But there’s a few issues with this section. One is that it doesn’t mandate making any change: All HUD is required to do is study the problem, and there’s no deadline for completing the process. Two is that HUD (which oversees the FHA) already had the authority to modify FHA loan draw schedules for modular projects, and has since 1974. So it’s not clear why this section was actually needed. Three, making it easier to use modules doesn’t really impact supply at all. It just makes it possible for developers to use modules instead of conventional construction. (Theoretically, there could be projects that only make sense with modular, and this could make those projects happen that otherwise wouldn’t, but I suspect the number of these sorts of projects is very near zero.) Fourth, this probably only applies to FHA-insured multifamily projects, which as we’ve noted is a very small housing universe. So while this section is targeting a real problem, I expect the impact of this section to be basically nil, though the resulting HUD report will probably make for interesting reading.

Section 303 — Property Improvement and Manufactured Housing Loan Modernization Act. This section increases FHA-insured loan limits and the allowable duration for manufactured home loans, as well as improvements/renovations to single family homes and manufactured homes. It also allows loans for Accessory Dwelling Units (ADUs) to be FHA insured.

These are changes to a pretty small program — FHA insurance for manufactured home loans was a substantial lending program in the 1980s, but today it’s basically not used at all (there were three such loans in 2021). Property improvement loans are now just a few thousand a year. Possibly these changes increase those numbers somewhat, but probably not by much. And only some fraction of these loans will have an impact on overall housing supply (specifically, property improvement loans which keep a manufactured home or park in operation that would otherwise deteriorate and be unoccupiable, and loans for ADUs that allow ADUs to get built that otherwise wouldn’t). So overall I see a pretty small impact here.

This section also requires HUD to conduct a study on the cost-effectiveness of offsite construction, which I suspect will make for interesting reading but won’t impact housing supply.

Section 304 — Preservation and Reinvestment for Community Enhancement (PRICE) Act. The PRICE program funds maintenance and repair of infrastructure at manufactured home parks — sewers, plumbing or electrical lines, roads, and so on. PRICE grants were first issued in 2024, and this section basically makes the program “official,” instead of needing to be reapproved every year. The first round of PRICE grants was for $225 million, allocated to 17 projects.

This is a potentially useful thing to allocate funding towards, and is another “conservation” section: keeping units in the housing market that might otherwise deteriorate to the point that they’re unoccupiable. To me the bigger question is how much of a difference it makes to be an official, authorized program. So the counterfactual impact of this section, compared to what would occur if the program was just reapproved (or not) on a yearly basis, is plausibly pretty low.

Conclusion

Titles II and III mirrored what we saw in Title I: The biggest potential impacts are concentrated in a small number of sections (the Innovation Fund, the RAD limit increase, the manufactured home chassis rule), but their actual impact will be a function of how much the constraint it targets was actually binding and how strong the incentive is, which in general is difficult to know. And there are a bunch of sections that target real problems (commercial to residential conversions being expensive, construction draws not aligning well with modular building, jurisdictions with overly restrictive housing regulations), but probably don’t have enough juice to make much of a dent in them.

Welcome to the reading list, a weekly roundup of news and links related to buildings, infrastructure, and industrial technology. This week we look at the Endangered Species Act and housing production, why it’s hard for LEGO to use 3D printing, robotic welding and shipbuilding, modular data centers, and more. Roughly 2/3rds of the reading list is paywalled, so for full access become a paid subscriber.

Housing

A paper by Max Tabarrok finds that the Endangered Species Act reduces housing construction by 10% in places near species’ listed habitats. Across the US, this works out to about a 4% reduction of total US housing stock. “In 1980, for example, 200 plant and animal species were listed as endangered or threatened and in 2025 there were 1,465 such listings. Each listed species has a distinct habitat. Thus, the growing list of protected habitats is a source of detailed variation in the intensity of land use regulation. This paper studies the effect of these regulations on housing construction. In a staggered difference-in-difference framework, exploiting 1,200+ species listings across the U.S. from 1980–2024, I find that the average treated place reduces its annual permit flows by 0.52 permits per thousand 1980 residents after a species listing, about 10% of the average annual flow.” [MTabarrok]

It’s been widely noted that people often pay a premium for access to good school districts. Elizabeth Warren, before she went full slopulist, wrote a book about this, and I noted in my look at build to rent housing that one benefit of build to rent is that it allows people to gain access to school districts that they otherwise couldn’t afford to live in. This cuts both ways though: if a good school near you closes, the value of your house might decline. “More than 1,000 public schools across 700 communities closed last year, according to new data—shuttering a prized neighborhood amenity and potentially putting nearby home values at risk.” [Realtor.com]

Manufacturing

Claims about why it would be hard for Lego to replace injection molding with 3D printing. “Even if you can get bricks that will mate the tolerances for clutch force, color accuracy, part durability, longevity, and more are so unbelievably tight…Polymer manufacturing has hard mode, expert mode, and then there is legendary god mode where LEGO delivers from on high. It is literally the pinnacle of what injection molding is capable of when held to standards like this.” [X]

Earlier this year we noted the risk of a chemical explosion at a California manufacturing facility. That manufacturer made commercial aircraft windows, and as a result there’s apparently now a looming shortage of them. “GKN Aerospace, one of the world’s biggest suppliers of cockpit and passenger windows, halted production in late May at its factory in suburban Los Angeles after an overheating tank raised fears of an explosion and prompted the evacuation of 50,000 nearby residents for several days.” [Reuters]

Automated factory startup Hadrian raises $1.37 billion in funding. [X]

Thanks to the on again, off again war with Iran, the US is apparently running low on long-range precision missiles. [Reuters]

Dylan Matthews argues that Trump’s suspension of the Jones Act (the much-reviled law that requires goods carried between US ports must be done on US-built ships) has created the opportunity to reform it. “But that a waiver actually went through this year, and has continued to be renewed, suggests the law’s critics are winning the intra-administration fight this time around. This makes 2026 perhaps the single most promising moment for meaningful Jones Act reform in decades. Durable reform, though, has to go through Congress, and will have to be bipartisan. It will also have to grapple seriously with what transitioning to a new regime looks like.” [Abundance & Growth]

As I note in my book, the best-quality iron ore in the US was mined out many years ago, and iron ore mined in the US needs to go through a process of “beneficiation,” which produces higher iron percentage pellets that can then be processed in a blast furnace. The US is apparently getting its first new iron pellet plant in almost 50 years, to produce pellets not for blast furnaces but for direct reduction in electric arc furnaces. [Industrial Sage]

Mariana Minerals raises $310 million in funding after restarting a shuttered US copper mine. [PRNewswire]

Read more

A few weeks ago the 21st Century ROAD to Housing Act became law. This is an enormous housing bill that’s been in the works for roughly a year — I first discussed it on Statecraft back in 2025 — and folks are optimistic that it will, per its description, “increase the supply of housing in America.”

But actually understanding what, specifically, the law will do is tricky. Like many large pieces of legislation, this is an amalgamation of many smaller laws: the Act has 59 individual sections, each of which does something different. Some of these sections create new housing programs; some of these sections modify existing government housing programs in complex ways; and others relax or change various housing laws but have complicated implementation details.

And once you’ve parsed what the law will actually do, you still need to predict how the various participants in the housing market will actually respond to the changes. (As we covered last month, the actual effects of laws can diverge significantly from lawmakers’ intentions.) Much of this law is aimed at changing the incentives around housing production — stimulating the production of housing by relaxing various constraints on producing it. But it’s not always obvious how binding the particular constraints that the law targets are, and how much housing we can expect to be unlocked by easing them.

As this is the biggest US housing bill in living memory, aimed at addressing one of the biggest problems in the US, it’s worth looking closely at the law and trying to tease out what the likely effects on housing supply will actually be.

Because there are so many individual sections, each of which requires drilling somewhat deep on a particular government program or aspect of the housing market, I’m not going to try to do this all in one shot. This will be the first in a series of posts (probably 4–5 total), each of which looks at one chunk of the housing bill.

A high-level look

At a high level, we can sort the individual provisions of the ROAD to Housing Act into three categories:

• Provisions that target the construction of new housing, either directly or indirectly.

• Provisions that target the preservation of existing housing.

• Provisions that do something else.

Despite the ostensible overarching goal of the law being increasing housing supply, most provisions of it actually fall into the third category and won’t realistically have any impact on housing supply. Some of these third-category provisions are totally unrelated to housing: there’s one that prevents the Federal Reserve from establishing a digital currency through 2030. Others are about housing, but won’t do anything to actually touch housing supply, such as a provision that provides grants to install temperature sensors in public housing. So while the law is a sprawling mass of changes, the biggest potential effects on housing supply are concentrated in a smaller number of provisions.

Another key aspect of this law is that no funding is allocated for any of its programs. The very last section, 1202, specifically states that “[n]o additional funds are authorized to be appropriated to carry out the requirements of this Act.” That means that any changes that require money to implement will need to get it from a separate congressional appropriation. For instance, section 208 of the Act creates an Innovation Fund of up to $200 million a year to award to jurisdictions that show increases in the housing supply, but the actual money for this fund will need to be awarded separately. This doesn’t mean these changes won’t be implemented, but it does create another barrier that many of the provisions will need to clear before they can have any effect. (For reference, the science half of the 2022 “CHIPS and Science Act” similarly passed with no funding specifically allocated to it, and anywhere from 60 to 90% of its authorized funding was actually appropriated depending on the year.)

The ROAD to Housing Act is broken into 12 “titles,” each one of which contains a bundle of provisions, but the most important changes are concentrated in titles 1, 2, 3, 5, and 10. In this post I’ll just look at Title I, “Opportunities for Housing.”

Title I — Opportunities for Housing

Title I of the ROAD to Housing Act contains seven provisions and is basically a grab bag of new programs and modifications to existing programs.

The headline items in this section are:

102 — Federal Guidelines for Point-Access Block Buildings. This instructs the Department of Housing and Urban Development (HUD) to create federal guidelines for point access blocks, multifamily apartment buildings with a single stairway and no central corridor. It also allows HUD to set up a grant program for jurisdictions that want to evaluate point access blocks.

103 — Exemption on Construction or Modification of Residential Housing on an Infill Site. This provision exempts infill housing constructed as part of the United States Department of Agriculture’s (USDA) Rural Development program from having to undergo an environmental review (such as required by NEPA).

105 — FHA Small-Dollar Mortgages. This allows (but does not require) HUD to create a program to expand access to inexpensive mortgages (less than $100,000) backed by the Federal Housing Administration (FHA).

These are all positive changes. Point access blocks, also known as single-stair buildings, are widely used in Europe, but disallowed by most modern US building codes, and they’re something that urbanists have been advocating for years. The Center for Building in North America, for instance, has devoted a lot of effort to advocating for single-stair buildings at the state level. Eliminating the requirement for a second stairway means that, for certain building sizes, a lot less of a building’s floorplate is devoted to non-rentable stairs and corridors, effectively reducing construction costs by an estimated 6–13%. I once spoke to a developer who was willing to cover the cost of a new fire truck if a jurisdiction would allow the use of single-stair buildings, such was the magnitude of the projected savings.

Expanding access to small-dollar mortgages is also tackling a real problem. The fundamental issue with small-dollar mortgages is a tension between what a mortgage originator earns on a mortgage and what it costs them to originate. A large chunk of origination costs is fixed — it takes a similar amount of effort to process the paperwork whether the mortgage is for $100,000 or $1,000,000. But earnings on a mortgage are a function of how much is being loaned, so profits are much smaller on small mortgages and originators have less incentive to originate them.

Likewise, it’s been noted over and over again how burdensome federal environmental review requirements can be: reducing them for a federal homebuilding program is a real win.

But we should also be realistic about the magnitude of the impact these changes are likely to have.

For section 102, point access blocks, this is a change that’s part of a much larger trend towards legalizing them. As of 2025, 19 states had introduced bills to study or allow point access blocks, and in May of this year the International Code Council voted to adopt provisions that would allow for point access blocks up to four stories tall in the 2027 International Building Code (a model building code adopted in some form by every US state). So even in the absence of this provision in the new federal law, jurisdictions were already moving in this direction. And on its own, section 102 has little in the way of teeth. It just mandates the creation of federal guidelines regarding point access blocks; it doesn’t require any jurisdictions to actually follow them.

Which isn’t to say that there’s no potential impact. A federal set of guidelines could potentially provide enough legitimacy to push jurisdictions to adopt point access blocks that otherwise wouldn’t. If a grant program materializes, that could also accelerate their adoption in certain jurisdictions. And if the generated federal guidelines allow for buildings with a higher number of stories (the provision allows up to six, rather than the four adopted by the IBC), this could theoretically push states to set their height limits higher than they otherwise would. But overall I expect the impact of this provision to be fairly marginal.

For section 103, environmental review of the USDA’s Rural Development infill housing, the impact of this is blunted by two things. One is that much of the program’s housing construction is already subject to a NEPA categorical exclusion; the other is that the program currently builds very little housing, and this section probably won’t change that.

The USDA’s Rural Development agency administers a collection of programs intended to “improve the economy and quality of life in rural parts of the US.” Many of these efforts are devoted to housing: since 1950 the USDA has funded the “construction, purchase, or repair of over 5.5 million rural housing units.” But currently, this program doesn’t build much housing: it’s hard to get clean numbers because the USDA doesn’t break down which loans go to new construction and which go to purchasing existing housing, but it’s probably building less than 5,000 units per year across single-family and multifamily housing, and possibly less than 1,000. Housing built on infill sites specifically (which are sites already served by infrastructure such as electricity, water, and sewer lines), which are the only projects covered by this provision, will be some smaller fraction of this.

What’s more, much of this construction was already effectively exempt from NEPA’s requirement to produce an environmental report. Repairs to existing housing, construction of single-family homes, and small multifamily projects already had a NEPA categorical exclusion, which means they don’t require a more thorough Environmental Assessment or Environmental Impact Statement. And this provision wouldn’t affect any state-level environmental study requirements, such as California’s CEQA. So in practice removing the requirement for an environmental study will only affect a narrow sliver of an already small universe of housing. It’s conceivable that removing a burdensome federal environmental review process might create an uptick in the number of projects that would fall outside the categorical exclusion, such as larger multifamily projects, but this would surprise me: I expect the impact of this change to be pretty minor.

For section 105, small-dollar mortgages, it helps to remember that the FHA is basically an insurance company: it sells mortgage lenders insurance policies against borrowers defaulting, in order to “incentivize lenders to make loans they otherwise would not.” It sells this insurance to a substantial chunk of the market: roughly 11 to 15% of US home purchases have FHA insurance. Because FHA insurance is so widely used, and because it specifically targets folks who might otherwise have difficulty getting a mortgage, a program to expand access to mortgages below $100,000 could potentially be a big deal for low-income homebuyers, enabling them to get mortgages that might otherwise not be worth a lender’s while to make.

The challenge here is that very few new homes built in the US are sold below $100,000. In 2023 there were only 1,000 new homes sold for less than $150,000 in the entire US. So currently, this would theoretically apply to at most a few hundred new conventional homes built each year.

There’s also a larger universe of manufactured homes that this could apply to: roughly 100,000 manufactured homes are sold each year, and nearly 40% of manufactured home mortgages are FHA-insured (though not all manufactured homes have mortgages). But this universe is also narrowed: a small-dollar mortgage program would only apply to homes titled as real property, which is probably about 20% (~20,000) of new manufactured home shipments. And only a fraction of these would actually fall below the $100,000 cutoff: the average price of a new manufactured home, sans land or installation, is already above $120,000. Once you include land and installation (as you would for any manufactured home titled as real property), you’re even farther above $100,000. It’s thus likely more than half of new manufactured homes titled as real property would be beyond the $100,000 cutoff.

Of course, the question for the effect on housing supply is not how many homes are currently built that cost under $100,000, but how many might be built if it were easier for folks to get mortgages for them. It’s conceivable that lendability of small-dollar mortgages is a major binding constraint, and with it relaxed we’d see a major uptick in the construction of smaller, more affordable houses for less than $100,000. (This might be especially true for manufactured homes, given the cost-reducing changes that are being made to manufactured home regulations in Title III of this legislation, which I’ll cover in a future installment in this series.) This doesn’t seem amazingly likely to me, but neither does it seem impossible. And outside of the impact on housing supply, there would probably also be some social benefits from increasing access to mortgages for existing low-priced homes.

A small-dollar mortgage program could also potentially apply to ADU construction, as long as the remaining mortgage on the primary property didn’t push the combined loan over $100,000. I suspect that’s also a pretty small universe of potential homes — if your mortgage balance is substantially less than $100,000 and you’re considering an ADU, you probably have other financing options available (such as a home equity line of credit) — but it’s another possible avenue for a provision like this to work.

Of course, this all depends on a small-dollar mortgage program actually being created, it being structured in such a way that actually incentivizes lenders to make more small-dollar mortgages, and its effectiveness not being eaten away by inflation (since there’s no provision for raising the $100,000 cutoff over time). None of these is guaranteed. But to me there’s a higher plausible upside to this provision than to the other provisions we’ve looked at.

Sections 102, 103, and 105 are what I consider the headline provisions in this section, but there are several other provisions as well.

101 — Reforms to Housing Counseling and Financial Literacy Programs. This section “clarifies the authority of HUD to evaluate housing counseling agencies” and makes some tweaks to how counseling resources are distributed. This will possibly have some impact on reducing foreclosures by giving counseling resources to help people avoid them (though I doubt it), but the effect on new housing supply is basically nothing.

104 — Database of Publicly Owned Land. This requires any jurisdiction that receives a federal Community Development Block Grant (CDBG) from HUD to maintain a publicly accessible database that identifies all undeveloped parcels of land owned by the jurisdiction.

This is, presumably, to make it easier for developers to find publicly owned parcels of land that might be suitable for development. And this affects a lot of jurisdictions — over 1,200 cities, states, and counties in the US get CDBGs from HUD. But it strikes me as unlikely that finding these parcels of land is a major bottleneck to the development process. GIS-based parcel ownership lookup tools, for instance, are already pretty common. There’s also the problem that this provision excludes a lot of potentially useful publicly owned parcels, because they’d be publicly owned by a government entity other than whoever is receiving the block grant. School districts, for instance, are technically their own special-purpose governments that own things in their own name. So an empty parcel of land owned by a school district wouldn’t necessarily be included in an undeveloped land database administered by the state or county.

It’s possible that the mere act of creating this database causes jurisdictions to look more closely at their land resources and decide to sell some of them for housing development, but that strikes me as not amazingly likely. Overall it’s hard to see this having much, if any, impact on housing supply.

106 — Temperature Sensor Pilot Program. This section requires HUD to create a pilot program to provide grants to housing agencies to install temperature sensors, to ensure public housing doesn’t get too hot or too cold. A positive development, but not one that affects housing supply at all.

107 — Housing Supply Frameworks. This section requires HUD to publish best practices for state and local zoning “to support the production of adequate housing.” This includes things like reducing or eliminating parking minimums, increasing allowed floor area ratios (FAR) to allow for bigger buildings on sites, encouraging transit-oriented development, and so on.

These are all things that urbanists like, but it’s hard to see much upside from publishing a list of best practices when information on these sorts of best practices is already widely available. I have heard, however, that the hope is that a “ROAD 2.0” housing act could include incentives for zoning best practices adoption, and for that it’s helpful if there’s already a set of federal recommendations. So the main impact of this provision might be a function of future legislation. On its own, though I doubt this will have much impact.

Conclusion

Overall the content of Title I in the 21st Century ROAD to Housing Act is fairly representative of what’s in the law. There are a few provisions that are aimed at very real problems in the housing sector — legalizing single-stair buildings, reducing the burden of environmental review, and improving the incentives for lenders to offer small-dollar mortgages — the effects of which will depend on how much that particular constraint was actually binding, and how people respond to the changed incentives. Along with these is a collection of other provisions that make various positive changes regarding housing (new resources, new or improved programs) but that probably won’t affect housing supply at all.

The next post in this series will look at Title II, a long section with lots of relevant housing supply provisions in it. And for more on what jurisdictions can do to take advantage of the various provisions and the federal dollars that will flow from them, my colleagues at IFP wrote a useful explainer.