ENVIZNThe Bottleneck Moves
Terafab is betting that the next constraint on AI will be industrial. History suggests a correct bet on that constraint can still fail if it is made too early.
Fabrication capacity is the constraint Terafab is aimed at. The rest of this list is what is likely to run short once more chips exist.
For most of two decades, computing has kept its physical infrastructure largely out of view. A browser opens and information appears. An app responds to a tap. A model answers a question in a few seconds. The chips, substations, transmission lines, cooling plants, fiber, and fabrication facilities behind that experience stay out of frame. AI is making that concealment harder to sustain.
On August 6, SpaceX and Tesla confirmed Terafab, a vertically integrated semiconductor complex in Grimes County, Texas, about an hour northwest of Houston. The announced first phase carries $16.8 billion of investment, and Texas incentive filings describe a four-phase project that could reach $119 billion. The finished campus is planned at roughly 100 million square feet, combining logic, memory, and advanced packaging on a single site. The companies say it will employ at least 3,000 people.
Those figures are large enough to be hard to evaluate on their own. The more interesting part of Terafab is the assumption underneath the building: that we are still badly underestimating how much compute the next decade will require. I think that assumption is correct, but an investment of this size wagers on direction and on timing at once, and those two can come apart far enough to ruin the company that got the direction right.
Capacity arrives before anyone knows what it is for
Whenever a large block of AI infrastructure is announced, a reasonable question follows: what is all of it going to do? The question assumes an order that technology rarely follows. Capability usually arrives first, and the applications are discovered afterward.
Early computers were not built because someone had mapped out cloud computing, autonomous vehicles, or language models. The internet was not deployed because its designers had already imagined streaming video or an economy organized around instant global connectivity. The smartphone shipped without a plan for ride-hailing, mobile banking, or short-form video. In each case the capability came first and the market spent a decade working out what it was.
There is a corollary to that pattern, and it is the dangerous one. Builders sometimes identify the infrastructure the future will need, build it correctly, and lose enormous sums anyway. The applications arrive and the capacity eventually gets used, but too late to pay back the companies that built it. That distinction does real work later in this essay, so it may be worth holding onto.
That should not be read as an accusation of aimlessness. Tesla wants autonomy and robotics. SpaceX wants compute in orbit. OpenAI, Anthropic, Google, and Meta are pursuing more capable systems. The aims keep moving because each increase in capability changes what is worth aiming at.
A more useful word is headroom, the term power engineers use for the spare capacity on a circuit before it becomes the limiting factor. Every increase in available compute buys another round of questions that were previously too expensive to ask. What happens if a model reasons for an hour instead of a second? If it sees and hears continuously rather than in samples? If an autonomous system interprets its environment in real time, and millions of them do it at once? If AI systems generate the training environments, simulations, and code that other AI systems learn from?
Nobody knows where that limit sits, because nobody has reached it. Terafab is a bet that it sits considerably farther out than the current consensus assumes.
Software has always been downstream of hardware
The software industry sometimes treats hardware as plumbing: necessary, uninteresting, and someone else's problem. That inverts the dependency. There are software ideas that are technically possible today and economically ridiculous, because the compute they require is too expensive, too slow, too power-hungry, or unavailable at the volume the idea needs.
A capability that exists in a laboratory has not yet become infrastructure. The first computers filled rooms. Then they sat on desks, then in pockets, then inside cars, cameras, thermostats, and industrial controllers. A restaurant ordering tablet now carries more computing capability than machines that once defined the frontier of computer science.
Frontier AI is still in the metered phase. Tokens cost money, inference consumes accelerators, and more sophisticated reasoning consumes more of both. Running an intelligent system continuously is a different economic proposition from asking it a question occasionally, and that constraint quietly decides what gets built.
Consider what changes if the cost of useful inference keeps falling. AI stops being an application you open and becomes a property of the systems around you: a vehicle perceiving the road without pause, a robot perceiving a factory floor, software reasoning continuously over a company's operations, research instruments running their own experiments. The interface between people and computers would shift again, away from opening applications and toward stating intentions.
None of that follows automatically from manufacturing more chips. Algorithmic efficiency, memory bandwidth, networking, energy prices, and model architecture all bind.
With enough compute, memory bandwidth tends to become the ceiling; with enough memory, firm power; with enough power, data worth learning from. That sequence is how technical progress usually runs, since each advance exposes a limit that had been hidden behind a more urgent one, and what counts as scarce keeps changing.
The site plans its own power and water supply
A fab at this scale cannot be separated from the systems required to run it. According to agreements released by Grimes County, SpaceX intends to power the site itself and not interconnect with ERCOT, the grid operator covering most of Texas. The plan calls for on-site natural gas generation paired with battery storage, a closed-loop water system for recycling and reuse, and water drawn from the nearby Gibbons Creek Reservoir rather than local groundwater.
In engineering terms, to build the machines that produce compute, SpaceX and Tesla have to become, at the same time, a generation developer, a water utility, and an industrial construction firm. Much of what will make this project hard sits outside the fab itself.
The decision to self-supply is the most revealing line in the filings. Choosing to build generation rather than wait for an interconnection is a statement about how long the queue takes relative to how fast these companies intend to move. It converts a scheduling risk they cannot control into a capital cost they can.
The same logic appears one level up, in the decision to build a fab at all. If Tesla and SpaceX cannot secure enough chips on somebody else's manufacturing schedule, they build manufacturing. If they cannot secure enough electricity on somebody else's grid development schedule, they build generation. Both decisions apply one strategy at two layers: identify an external dependency that can delay the company, and move as much of it as possible inside the fence. Read that way, Terafab is aimed at the companies' dependence on outside suppliers, TSMC included.
This is the pattern I keep finding underneath AI stories once you follow them far enough. A discussion of AI chips leads to fabs, fabs lead to electricity and water at industrial volumes, and AI data centers lead to generation, transmission, cooling, and land. Scaling intelligence indefinitely also means confronting its thermodynamic cost: however weightless it looks on a screen, computation consumes electricity and produces heat that has to be removed from the building.
Correct infrastructure, built too early
There is a problem with everything argued so far. History contains companies that correctly identified the infrastructure the future would need, built enormous quantities of it, and destroyed themselves anyway.
In the late 1990s, telecommunications companies made what looked like an obvious bet. Internet traffic was growing quickly, fiber optic technology was improving faster still, and the future would plainly require far more bandwidth than the network then carried. Billions of dollars went into long haul fiber, capacity expanded faster than demand could economically absorb, prices collapsed, and the debt stayed. By 2001, most estimates put the share of installed fiber actually carrying traffic in the single digits. The unused remainder became known as dark fiber.
Global Crossing became the clearest casualty. Founded in 1997, it built a network of more than 100,000 miles reaching over 200 cities in 27 countries, on a thesis that was correct in the sense that matters most. The world really did go on to consume extraordinary quantities of bandwidth. In January 2002 the company filed for bankruptcy listing roughly $12.4 billion of debt, at the time among the largest failures in United States history. The traffic did arrive, but only after the company's capital structure had failed.
The comparison needs one honest qualification, because Global Crossing did not fail from bad timing alone. It was also inflating revenue through capacity swaps with other carriers, and the accounting was fraudulent. The case is still instructive, because the carriers that kept clean books mostly failed too. The glut looks structural: once the trench is dug, the marginal cost of laying another strand is close to nothing, which may be exactly the property that turns a good thesis into an overbuild.
I think that distinction matters more than anything else in this story. The infrastructure thesis can be right while the investment thesis is wrong, and the two are easy to confuse because they tend to run on the same evidence. Musk can be entirely correct that the future requires extraordinary semiconductor capacity and still build too much of it too early, finance it too aggressively, choose the wrong architecture, or find that the binding constraint has moved before the asset has earned its cost.
Doubting the execution while accepting the thesis
Any project attached to Elon Musk becomes hard to discuss at normal volume. For some readers his involvement is proof that something will change the world, and for others it is proof that the announcement can be ignored. Neither reaction assesses the project itself.
The record supports real caution about timelines. Musk's companies have a long history of announcing schedules and targets that arrive late, change substantially, or do not arrive, and SpaceX itself acknowledges no assurance that Terafab will meet its objectives on the expected schedule or at all. Semiconductor fabrication may also be the hardest industrial process on earth to vertically integrate. TSMC, Samsung, and Intel hold decades of accumulated process engineering, supplier relationships, and institutional memory that a larger building does not replace. The first stage is expected to run on Intel's 14A process, which is a useful reminder that even this project depends on capability it did not build.
The stated output target deserves particular care. The companies describe Terafab as eventually producing more than one terawatt of compute per year. That is an unusual unit. Compute is not normally measured in watts of annual production, and the figure appears to describe the aggregate power draw of the silicon shipped rather than any conventional measure of throughput.
That should change how the rest of Terafab's numbers are read. One terawatt of compute works as an engineering specification and also as part of the argument SpaceX and Tesla are making about what the facility is: a factory whose product is computational capacity, described in different terms from a conventional fab measured in wafer starts, yields, and process nodes. The framing is revealing, and it also means the headline figure should not be treated as a comparable measure of foundry output. It probably says more about the scale of Musk's expectation for future demand than about Terafab's eventual throughput.
Dismissing the project because it sounds absurd would be a different error. Reusable orbital rockets and vertical booster landings sounded absurd, and so, to most of the automotive industry, did building electric vehicles at mass-market volume. Some ideas that sound absurd fail and others become industries, so how a claim sounds is no evidence either way.
My own position sits between the two reflexes and is narrower than either. I expect the schedule to slip and the design to change substantially from what is being described this month. I also think the reasoning that produced the decision is sound, and those are separate judgments that deserve to be held separately.
National underinvestment, industry wide underinvestment, and a single company's inability to secure supply are three different problems supported by three different kinds of evidence, and they are worth keeping apart. Terafab does not seem to require the first two in order to make sense. It requires only that Tesla, SpaceX, and xAI eventually need more semiconductor capacity than outside suppliers will guarantee on the schedule Musk wants. That is a much smaller claim, and probably a much harder one to argue with.
The bet is larger than any model
Terafab's demand case extends well beyond chatbots: autonomous vehicles, humanoid robots, industrial automation, scientific simulation, computer vision, and satellite autonomy all need compute, and whatever succeeds today's language models will need it too. Tesla, SpaceX, and xAI are betting that demand generated across those systems eventually grows large enough that depending on the existing global supply chain becomes a strategic liability rather than a procurement decision.
The bet does not assume the current semiconductor industry disappears: Musk has credited TSMC, Samsung, and Micron while arguing that his companies' expected demand will outrun what those suppliers can allocate to them. What has changed is posture: a very large buyer looked at its own forecast and concluded that buying chips may not secure enough of them.
AI competition began as a software race and became a GPU race, then a data center race, then a power race. The largest participants are now reaching further down the stack toward the industrial machinery that produces compute itself.
The bear case: AI succeeds and the constraint moves
The strongest argument against Terafab assumes artificial intelligence succeeds, just in a way that makes Terafab's assumptions obsolete.
Suppose algorithmic efficiency keeps compounding. Models complete today's workloads on a fraction of the computation. Inference migrates onto specialized silicon built for narrower jobs. Memory or networking rather than logic fabrication becomes the binding constraint. Existing foundries expand faster than anyone expected. Software learns to extract far more useful work from every transistor. Terafab might then be adding capacity at very large scale against a constraint that no longer binds.
I think that is the serious risk, and it is serious because a fab is not an asset that can be cheaply repurposed each time the architecture underneath it changes. The future almost certainly needs more compute. What matters for Terafab is whether logic fabrication stays the binding constraint long enough for Terafab to earn back the cost of moving it.
Efficiency, though, cuts both ways. If useful inference becomes ten times cheaper, people may not consume a tenth as much compute. They may find ten times as many things worth doing with it. Economists call this a rebound effect, and computing has produced it repeatedly. Cheaper machines produced more computers rather than less computing. Cheaper bandwidth produced streaming, cloud software, and video heavy applications rather than a society content with the data it already used. Terafab is therefore also betting that making useful computation cheaper expands the market for computation itself. I find that bet persuasive, though the rebound effect is a historical pattern that may not repeat.
The most consequential technology of the next decade may not resemble what we are using now. The smartphone was not a smaller desktop, and the next generation of AI may not be a better assistant. We might address computers differently, robots might act in the physical world differently, and research might become partly autonomous. Categories of software may appear that make current applications look primitive. We do not know which, and that uncertainty is part of the case for building capacity before its uses are known.
My guess is that within five years the need for the capacity will be settled, and the interesting question about Terafab will be who else started building theirs early enough to matter.
Terafab may prove too large or too early, and its economics could fail even if its forecast of demand proves right.
That combination of a sound thesis and an uncertain investment is why it is worth studying closely. The future will almost certainly require more infrastructure; the open questions are whether Musk has correctly identified the next binding constraint and whether he can afford to be early.
Terafab's first phase commits $16.8 billion, against up to $119 billion across four phases in Texas incentive filings.
All insights