Terafab: The One-Terawatt Bet Rounding Up the Chip Giants

The world's biggest chipmakers are lining up at Elon Musk's door. Intel signed in April, Nvidia chief Jensen Huang gave open backing, ASML wrote Terafab into its 2027 plans, and now TSMC, the planet's largest foundry, is discussing a Texas partnership. Musk, meanwhile, moved the goalposts: SpaceX is no longer a rocket company in his telling, it is a superintelligence company . The target of one terawatt of computing hardware a year means roughly fifty times today's global AI chip output.

The video opens on a confident note: the AI agents game is won, the chapter is closed, and the comeback has started. The Grok-based agent bot answers almost instantly, a rehearsal for a future where agents exchange a thousand times more messages than humans. Costs are moving in the right direction too, with open-source models pushing token usage up and unit prices down. The picture favors infrastructure investors, but the real break will happen not in chat, but in the physical world.

From SpaceX to SpaceXSI: The Superintelligence Sign

Over the weekend Musk announced the new doctrine in one short line: the artificial intelligence era is over, and superintelligence is next. When a fan suggested the SpaceXSI name, he confirmed the change is coming. Politics plays a part; as QZ.com reported, a September 29 executive order pushes federal agencies to say superintelligence instead of artificial intelligence, and Musk is the first to embrace the language. Markets played along, with QZ.com data showing SpaceX shares trading about 18 percent above their June listing price.

The rebrand continues a giant merger from February. According to BBC, SpaceX absorbed xAI, the startup behind the Grok chatbot, valuing xAI at about 125 billion dollars and SpaceX at 1 trillion, creating the most valuable private company in history. Musk called the combination an innovation engine uniting AI, rockets, space internet and media under one roof. Tesla's 2 billion dollar investment in xAI last month, and the central role of humanoid robots in the company's future, belong to the same story.

What exactly does Terafab promise? The logic is simple: a vertically integrated giant plant putting everything from chip design and mask-making to logic, memory manufacturing and advanced packaging under one roof. The prototype will iterate fast on the Tesla campus in Austin, while the permanent plant rises in Grimes County. The scale is staggering, around 100 million square feet, roughly ten times the Texas gigafactory. The site will make chips for both model training and on-device inference.

The Bill Grows: From 55 Billion to 119 Billion Dollars

Filings attached to a tax-abatement request put the first phase at 55 billion dollars, rising to as much as 119 billion at full build-out. CNBC reports Musk unveiled the project in March as a joint venture of Tesla, SpaceX and xAI, with a research fab in Austin budgeted near 3 billion dollars and targeting a few thousand wafers a month. Chip analyst Ben Bajarin calls it a 15-year strategy. The numbers are huge, yet Musk's record whispers a familiar lesson: at sufficient scale, quantity becomes a quality of its own.

The consortium's first heavyweight is Intel. Under the April deal, Terafab will use Intel's 14A process in the full-scale plant. Datacenterdynamics reports Musk told the first-quarter earnings call that 14A should be mature by the time the plant scales. Intel brings design, fabrication and packaging know-how that neither Tesla nor SpaceX owns alone. The partnership ranks among the boldest outside-customer wins in the race to bring chip manufacturing home to America.

The biggest surprise came from TSMC. The foundry giant that only wished Musk good luck in June is now discussing a Texas role. Talks are at an early stage, TaiwanPlus reports, with chairman C.C. Wei reminding everyone that a new wafer fab takes two to three years plus another year or two for capacity. Two formulas are on the table: TSMC builds and runs the plant while SpaceX invests and buys chips, or SpaceX keeps the majority. That even the company making the world's chips is engaging says something about where demand is heading.

Nvidia's Backing and ASML Taking It Seriously

Jensen Huang spoke without hesitation at the September All-In summit: if anyone can pull this off, Musk can. The two spent hours discussing Terafab on a long flight together, Yahoo Finance reports, with Huang calling Musk's undiscourageable nature his superpower. The relationship runs deeper than words; the two firms built the 100,000-GPU Colossus supercomputer in Memphis in 122 days, and in August SpaceX committed to building exclusively with Nvidia. Huang's support lends technical legitimacy, though whether Nvidia will ever fab its own chips at Terafab remains open.

The equipment side opened up too. ASML chief Christophe Fouquet says he spoke with Musk directly and found him very serious. Demand will keep the market supply-constrained for a long while, Benzinga reports him warning, forecasting a global semiconductor market of 1.5 trillion dollars by 2030. First logic chips from next-generation High-NA EUV systems are expected within months, with Intel among the earliest users. Since Terafab is both a prospective customer and a potential long-term rival, the Dutch giant faces a double-edged opportunity.

The Real Break: The Century of Inference, Not Training

The most provocative claim is that the race between language models leapfrogging each other is nearly over. Models converge, costs fall, usage shifts to open source; what matters is no longer who chats best, but who carries intelligence into the physical world. Musk builds the infrastructure himself, keeps vehicles and soon humanoid robots ready, and leases spare capacity. While rivals source compute from a dozen clouds, this vertical integration buys speed. Forecasts of agent traffic reaching a thousand times human usage explain the hunger for inference capacity.

The numbers back the appetite. Nvidia disclosed about 1 trillion dollars of order visibility stretching into 2027, Constellationr notes, with the GTC 2026 headline being the arrival of the inference inflection. Training chips make models smarter, but volume lives in inference : cars, robots and phones will all carry inference silicon. In Huang's words, inference is both the hardest and the revenue-driving job. Terafab's one-terawatt goal plays exactly there, aiming to be the factory of the agents' world, not today's.

What remains is risk and schedule. As the guest stresses, hundreds of billions of dollars are on the table with no room to miss; stumbling at this scale is nothing like stumbling on a battery line. Musk sees the next two years as critical and political support as essential. The bulls rest their case on the track record: those who took him seriously on electric cars, the Supercharger network and self-driving were rewarded early. If Terafab works, computing gets cheaper; if not, the bill enters history among the costliest industrial bets.