in short
Elon Musk has committed $16.8 billion, with plans totaling as much as $119 billion across four phases through 2036, to building the largest chip factory on Earth, a joint venture across Tesla, SpaceX, and Intel, aiming for a scale of "one terawatt" of annual AI compute. TSMC took roughly fifteen years and a comparable order of magnitude in capital to become the company every advanced chip in the world quietly depends on. The question worth asking isn't whether Terafab has enough money. It's whether money was ever the thing TSMC's fifteen-year head start was actually built on.
On March 21, 2026, Musk stood inside a defunct Austin power plant and announced Terafab: a joint Tesla-SpaceX-xAI venture, with Intel brought in shortly after for manufacturing expertise, aimed at ending what Musk described as an unavoidable chip shortage across his own companies. "We either build the Terafab, or we don't have the chips, and we need the chips, so we build the Terafab," he said. By August, the project had a real site, Grimes County, Texas, outside Houston, a $16.8 billion first phase, and a stated ambition to construct the largest building on Earth, more than five times the size of the current record-holder (Fortune; TechCrunch). Site clearing began in August, and state filings list a formal construction start of December 1, 2026 and completion of the Grimes County facility by the end of 2028 (Spectrum News; Wikipedia). ENR's breakdown of the plan shows four construction phases running from 2026 through 2036, with eligible investments of $16.8 billion, $38.6 billion, $38.6 billion and $25.3 billion, for a potential total of $119.4 billion (ENR).
This is a genuinely serious project, not a keynote slide. It's also attempting something no company has done in modern chipmaking history: compress a fifteen-year institutional head start into a single, capital-driven sprint.
Why Musk Says This Is Necessary
The stated rationale is straightforward, even if the scale is not: Musk's companies believe their own future chip demand will exceed what the existing global chip supply can provide, regardless of price. SpaceX put it plainly in its own public statement: "While we are deeply appreciative of our current chip suppliers and encourage them to expand production whenever possible, this looming gulf between supply and demand is at the core of Terafab's necessity" (Fox Business). Musk has previously suggested his companies could need on the order of 100 to 200 billion AI chips per year once every project is running at scale (Yahoo Finance), a figure that dwarfs what any merchant foundry currently allocates to a single customer. At the launch event, Musk went further, saying all the fabrication facilities on Earth today produce only about 2 percent of what Tesla and SpaceX will need across their projects (Wikipedia, citing Electrek).
The demand is coming from several directions at once, which is part of why the number is so large. Tesla needs chips for full self-driving compute in every vehicle and for training the underlying models; the Optimus humanoid robot program adds an entirely separate category of edge-inference silicon at (Musk hopes) eventual mass-market volume; SpaceX's Starlink and orbital ambitions add their own compute needs, including nonterrestrial chips designed specifically for space deployment; and xAI's Grok models require frontier-scale training clusters on top of all of it (MarketWise). No single one of these programs would justify building the world's largest factory. Combined, and projected forward at the scale Musk has described, they plausibly could.
The Queue-Jumping Logic
There's also a competitive-positioning argument underneath the demand argument. TSMC, Samsung, and Intel's existing capacity is already substantially spoken for by the hyperscalers, Microsoft, Google, Amazon, Meta, who have spent years building the same kind of allocation priority TSMC itself took fifteen years to earn. A new, chip-hungry entrant competing for space in that same queue is negotiating from a weaker position than any of the incumbents already in it. Vertical integration, building your own fab rather than depending on someone else's allocation decisions, is one way to escape a queue you didn't get to join early. Whether that's a rational strategic move or an expensive way to relearn why the queue exists is precisely the question the rest of this piece takes seriously.
Colossus: The Proof Already Exists
Colossus is xAI's GPU training cluster in Memphis, and it's the clearest live evidence for the demand side of this argument. Estimates put the complex at roughly 555,000 Nvidia GPUs by February 2026, up from 100,000 in 2024, with a stated target of more than a million (Introl). By mid-2026 the combined sites provided on the order of one gigawatt of AI compute, with Musk projecting roughly two gigawatts once the cluster is fully scaled (Wikipedia; Measured AI). Since SpaceX's acquisition of xAI in February 2026, that compute pool also feeds Tesla's Full Self-Driving development directly (Basenor).
Every one of those GPUs was built on the same TSMC-fabricated, ASML-dependent supply chain this piece has been describing. That makes Colossus do double duty: it's Musk's strongest evidence that his companies' demand is real, and it's his clearest demonstration of exactly how exposed that demand currently is to someone else's factory. One industry analysis framed the two projects as intentionally complementary: "training infrastructure producing models, and manufacturing capability producing deployment hardware" (Nexairi). Terafab, in that reading, isn't a separate project, it's Musk trying to build the supply-side answer to a demand-side problem Colossus already proved exists. Notably, Colossus has also run directly into this piece's other bottleneck: when the local Memphis grid connection proved too limited, xAI deployed temporary gas turbines and battery systems to cover the gap (Wikipedia), the same energy constraint driving fab site selection everywhere else in this program's coverage, showing up one layer downstream, at the point of actually running the chips rather than making them.
What TSMC Actually Has That Money Can't Buy Overnight
TSMC didn't become the company producing the overwhelming majority of the world's advanced logic chips by simply spending the most money, though it has spent enormous sums. It became that company by being first in line, consistently, for every generation of the one machine this entire industry depends on, ASML's EUV lithography tools, for the better part of two decades, and by using that position to build an ecosystem no rival has matched: materials suppliers calibrated to its exact process, a trained workforce numbering in the tens of thousands, and yield rates on advanced nodes that took years of iteration to reach.
Terafab is attempting to build all four of those things, allocation priority, materials relationships, workforce, and yield, simultaneously, from a standing start, at the same time it's trying to construct what would be the largest building on the planet. Musk has stated Terafab will use Intel's 14A (1.4-nanometer) process, and that by the time the full-scale facility ramps, the process "will be probably fairly mature or ready for prime time" (Wikipedia, aggregating Reuters/press coverage). That single sentence contains the entire risk profile of this project: it is relying on a process node that isn't proven yet, at a facility that doesn't exist yet, to solve a chip shortage that already exists today.
The Lithography Math Doesn't Care How Rich You Are
Here is the part of Terafab's plan that collides directly with physical reality rather than financial reality. Intel's plans for the 14A process are built around High-NA EUV, the newest and scarcest tier of ASML's lithography technology, though Intel has also described a Low-NA-only option (Bits&Chips). Either way the process runs on EUV, and ASML is the only company that makes it. ASML expects to ship roughly 65 standard Low-NA EUV systems in 2026 and plans about 30 percent more capacity for 2027, but its CEO said in July that nearly all of that expanded capacity through 2027 is already booked (ASML, SEC filing; Reuters, via Euronext). The High-NA tier is far smaller still: ASML's own estimate is about ten scanners for 2027 (TechPowerUp), and fewer than a dozen such systems existed worldwide as of early 2026 (TrendForce).
Timing is the subtle part. Terafab's Grimes County facility isn't due for completion until the end of 2028, so its tool orders land in the window where ASML says it is still investigating whether to add yet another 30 percent of capacity. Terafab would be competing for those later slots against customers who have been ordering for years, and its own partner, Intel, is only now qualifying High-NA on select 18A product layers (ASML, SEC filing) and targets 14A manufacturing in 2027 to 2028 (Bits&Chips). Terafab is asking its manufacturing partner to share an already-scarce supply of tools with a brand-new, unproven third-party megaproject.
What the Scale Gap Actually Looks Like
The gap between ambition and physical capacity is not subtle. One independent analysis calculated that reaching Terafab's stated goal of one terawatt of annual compute would require something on the order of 22 million advanced logic wafers (a wafer is the thin slice of silicon that chips are printed onto, dozens or hundreds of individual chips are cut from a single one once manufacturing is complete) a year and roughly 150 fab modules (a fab module is one self-contained production line within a factory, a large facility runs many of these in parallel, and each one takes years to build and qualify on its own), a level of advanced manufacturing capacity that does not currently exist anywhere on Earth, across every chipmaker combined (Optimusk). The same analysis's own verdict is worth stating plainly: "$16.8B buys a very large first phase. It does not buy a terawatt. Treat the terawatt figure as a directional ambition, not a capacity forecast." Electrek's coverage has been openly skeptical for a related reason, the "terawatt" framing doesn't actually originate from chip-industry convention at all, but from Musk's separate orbital-compute ambitions, imported into a semiconductor announcement where it doesn't naturally belong (Optimusk, citing Electrek).
Where Terafab's Approach Might Actually Be an Advantage
Fairness requires taking the strongest version of the counter-case seriously, and there is one. Musk has a documented pattern of defying informed consensus: SpaceX succeeded at reusable rockets when the launch industry's experts said it wasn't commercially viable; Tesla became the world's most valuable automaker after most of the auto industry dismissed EVs as a niche. Doubting that pattern has been a consistently bad call for over a decade.
The Cleanroom Idea and the Vertical-Integration Edge
Musk has also floated a genuinely unconventional idea about fab design that, if it works, could sidestep part of the labor and cost bottleneck entirely: rather than keeping an entire building ultra-clean the way conventional fabs do, isolate the silicon wafers themselves in sealed environments throughout the manufacturing flow, relaxing the cleanroom requirements for the building and the people working in it (Yahoo Finance). Whether this is a genuine engineering breakthrough or an underestimation of why cleanroom-scale sterility exists in the first place is an open question, but it's exactly the kind of first-principles reconsideration that has paid off for Musk's companies before, even when experts initially dismissed it. And there is a vertical-integration logic that does differentiate Terafab from TSMC's model: TSMC builds chips for hundreds of external customers with competing priorities; Terafab, if it works, builds exclusively for Tesla, SpaceX, and xAI's own captive demand, which removes an entire layer of allocation negotiation TSMC has to manage across its customer base.
The Challenge: This Might Simply Be a Later-Stage Version of the Same Project Everyone Else Is Running
The honest skeptical read deserves equal space. Every element of Terafab's pitch, vertical integration, first-principles cleanroom redesign, an unprecedented capital commitment, has echoes of claims made about other ambitious infrastructure projects that ultimately reduced to "very large, real, and still years behind the stated ambition," which is precisely the assessment independent analysts have already reached about Terafab itself, less than a year after announcement (Optimusk). The project's own numbers have moved meaningfully within months: an initial $20-25 billion pitch in March (Tech Insider), county filings in May discussing an initial outlay of about $55 billion and up to $119 billion across all phases, and a $16.8 billion first phase announced in August (Wikipedia). Even the near-term ambition has been reset: the prototype fab's initial target of 100,000 wafer starts a month was later described by Musk as "maybe a few thousand wafers per month," intended mainly "to try out ideas" (Wikipedia, citing Reuters). SpaceX's own IPO registration said Terafab was at a very early stage, with Tesla and SpaceX having agreed only a general framework and with financial terms, intellectual property rights and the continuing participation of Intel and Tesla still subject to later agreements (Wikipedia, citing Electrek). Shifting numbers are common to ambitious infrastructure projects generally, not unique to Musk, but they are worth noting as a base rate regardless of who is running the project.
One fact also complicates the demand story. SpaceX is renting large blocks of Colossus capacity to outside customers: Anthropic agreed in May 2026 to rent essentially all of Colossus 1, and Google agreed in June to rent about 110,000 GPUs starting in October (Wikipedia). Renting GPUs to others and needing custom chips of one's own are different needs, so the two aren't contradictory. But a company that says it faces an acute chip shortage while leasing out a meaningful share of its existing compute is telling a more nuanced story than the headline suggests.
What Doesn't Transfer With a Capital Commitment
There is also a meaningful difference between what Terafab needs to prove and what TSMC already has. TSMC's advantage was never simply "TSMC has more money than everyone else", it's accumulated yield data across billions of chips, multi-decade relationships with ASML that determine allocation priority in scarce years, and an ecosystem of materials suppliers calibrated specifically to its process quirks. None of that transfers with a capital commitment, however large. Terafab could, in principle, eventually build all of it. Fifteen years is the honest timeline TSMC needed to get where it is today, and nothing about Terafab's announcement suggests that clock resets faster just because the checks are bigger.
Conclusion
Terafab is not a bluff, and treating it as one would be a mistake, the site is real, the initial capital is real, and Intel's manufacturing partnership is real. But the specific claim worth scrutinizing isn't whether Terafab will get built. It's whether the thing that actually made TSMC dominant, allocation priority with a single, scarce supplier, built over a decade and a half of consistent execution, can be compressed into a few years through capital alone. On the evidence available today, the honest answer is: probably not on the original timeline, and almost certainly not at the scale currently being promised. What Terafab is actually buying, at $16.8 billion and rising, is a real, serious first phase and a credible long-term threat to the current order. Whether it becomes more than that will be decided by ASML's shipping manifest, not by Musk's balance sheet.
Sources
source
covers
Terafab site and scale
$16.8B first phase
Groundbreaking and site status (Sept 2026)
Four-phase, $119.4B plan through 2036
Timeline; Intel partnership; 14A detail; state filings; S-1 language; Musk's 2% claim
SpaceX's stated rationale for Terafab
Musk's chip-demand figures; cleanroom concept
Demand sources; SpaceX/xAI context
Initial cost pitch (March)
Colossus GPU count (Feb 2026) and target
Colossus compute power (~1 GW; ~2 GW projected)
SpaceX-xAI merger; link to Tesla FSD
Compute capacity; grid constraints; Anthropic and Google rental agreements
Colossus/Terafab as complementary demand/supply bet
2026 Low-NA volume; 2027-2028 capacity plans; Intel High-NA qualification
EUV capacity booked through 2027
ASML's ~10 High-NA scanners planned for 2027
Fewer than a dozen High-NA systems worldwide
Intel 14A timing; Low-NA-only option
Terawatt-vs-capacity analysis; Electrek skepticism
