Elon Musk first floated the idea of a Tesla chip factory at the company's November 2025 annual shareholder meeting. He made it official on January 28, 2026, during the Q4 2025 earnings call: "In order to remove the probable constraint in three or four years," he told investors, "we are going to have to build a Tesla TeraFab." Then, last Saturday, he posted on X that the "Terafab Project launches in 7 days."
That puts the date at March 21, 2026 — this coming Saturday.
What Musk is describing, if it ever reaches full scale, would be the largest in-house AI chip manufacturing facility ever built by a private company outside of Taiwan and South Korea. Bigger, in wafer output terms, than most of what TSMC has built across its entire global footprint over forty years.
The semiconductor industry is skeptical. Musk has been here before with timelines. But the urgency is real, the investment is staggering, and the chips in question power everything that makes Tesla's next decade possible.
The Supply Constraint That Drove the Decision
Tesla is not building chips for sport. The company has a problem that is only going to get worse.
Full Self-Driving, the Cybercab robotaxi, and the Optimus humanoid robot all run on custom AI silicon. Tesla's fifth-generation chip, AI5, is in its final design stages and is expected to be dual-sourced from TSMC and Samsung for initial production while Terafab comes online. But even combining both suppliers, Musk's team projects a hard ceiling on chip supply within three to four years — a ceiling that no amount of purchasing power can fix, because advanced node capacity is globally constrained.
NVIDIA's H200s and GB200s are already sold out through 2026. TSMC's N2 node is oversubscribed — confirmed first customers include Apple, NVIDIA, AMD, Qualcomm, and MediaTek, with no capacity to spare. Samsung's SF2 is not yet at full yield. There is simply no slack in the system.
The only way out, from Musk's perspective, is to become the system.
What Terafab Actually Is
Terafab is not a data center. It is not a packaging house. It is a full-stack semiconductor fabrication facility targeting 2 nanometer process technology — the same node that TSMC calls N2 and Samsung calls SF2, and which only those two companies currently operate in volume.
The stated targets are aggressive by any measure:
- $20–25 billion investment, making it one of the largest private-sector chip bets in history
- 2nm process node for the AI5 chip and subsequent generations
- 1 million wafer starts per month at full scale — roughly 70% of TSMC's total current monthly output, in a single facility
- 100 to 200 billion chips per year in annual production volume
- Initial output of 100,000 wafer starts per month, ramping over several years
The facility is planned for the North Campus of Giga Texas in Austin, adjacent to the existing Gigafactory. Site preparation is already underway, with earth-moving and foundation work covering a footprint that rivals the original Giga Texas building itself — which, at roughly 930,000 square meters (10 million square feet), is one of the largest structures in the world by volume.
Tesla's plan is to produce chips not just for its own vehicles and robots, but also for xAI's Dojo training infrastructure and Grok model inference — effectively turning Terafab into an internal foundry for the entire Musk technology stack.
| Tesla Terafab (Target) | TSMC (Current) | Samsung Foundry (Current) | |
|---|---|---|---|
| Most advanced node | 2nm (N2 equivalent) | 2nm (N2) | 2nm (SF2) |
| Monthly capacity (wafers) | 1M (full scale) | ~1.4M (all nodes combined) | ~350K (all nodes combined) |
| Annual chip volume | 100–200B | Not publicly stated | Not publicly stated |
| Investment | $20–25B | Spent $100B+ over decades | Spent $50B+ over decades |
| Timeline to volume production | 2027–2029 (projected) | Operational now | Operational now |
| Location | Austin, Texas | Taiwan, Arizona, Japan | South Korea, Texas |
What March 21 Actually Means
Semiconductor engineers will tell you the same thing: you do not open a 2nm fab in five days. You do not open one in five months. A greenfield advanced-node facility, from land prep to first wafer out, takes four to five years under ideal conditions. TSMC's Arizona fab, which broke ground in 2021, only began producing first chips in late 2024 — and it still required flying hundreds of Taiwanese engineers to Phoenix to make it work.
So what is happening on March 21?
The most likely scenarios are a formal project reveal with detailed specifications, a site groundbreaking ceremony at the Austin North Campus, or the public confirmation of a construction contract. Some reports suggest it could also be the announcement of a partnership — possibly with Intel.
Musk has already floated the Intel angle publicly. At Tesla's November 2025 annual shareholder meeting, he told the audience: "Maybe we'll do something with Intel. We haven't signed any deal, but it's probably worth having discussions with Intel." Intel's stock jumped in after-hours trading after that comment. An Intel involvement would give Terafab something it desperately needs: process engineering expertise. Intel's Foundry business, which is building out its 18A node (comparable to 2nm), has the institutional knowledge of running a domestic fab that Tesla completely lacks. The partnership would give Intel a critical anchor customer, and give Musk a shortcut around the decade-long learning curve.
Whether that deal gets announced Saturday is unknown. What March 21 is not is an operational fab.
Jensen Huang Has Thoughts
The CEO of NVIDIA, the company that currently profits most from Tesla and every other AI company's inability to make its own chips, was asked about Musk's chip factory ambitions during a November 2025 visit to Taiwan — and his response has only grown more relevant since.
He did not hold back.
"Building advanced chip manufacturing is extremely hard," Huang told reporters. "It's not just build the plant, but the engineering, the science, and the artistry of doing what TSMC does for a living is extremely hard." He went further: matching TSMC's capabilities, he said, is "virtually impossible."
This is not just competitive trash talk. Huang is describing something that anyone who has worked in semiconductor manufacturing understands viscerally. A modern 2nm fab requires mastery of extreme ultraviolet lithography, materials science at the atomic level, yield management across hundreds of interdependent process steps, and a trained workforce that simply does not exist at scale outside of Taiwan and South Korea. TSMC has spent forty years and hundreds of billions of dollars building that expertise. It still loses yield on new nodes for the first six to twelve months of production.
Tesla has none of that institutional knowledge. What it has is Elon Musk's conviction that enough money and first-principles thinking can compress any timeline.
The semiconductor industry's history of outsiders attempting vertical integration is not encouraging. Intel, which does have that expertise, has spent $100 billion trying to rebuild process leadership and is only now approaching competitive parity with TSMC's leading edge. The notion that Tesla — a car company that recently exited a significant revenue slump — can replicate a generation of fab engineering in two to three years is, to put it charitably, ambitious.
Who Gets Hurt, and Who Benefits
If Terafab works — even partially — the consequences for the chip industry are significant.
TSMC stands to lose one of its most valuable forward-looking customers. Tesla's AI5 orders currently run through TSMC's N3 and will shift to N2. If Terafab achieves even half its stated throughput, those orders disappear. More damagingly, a successful Tesla fab would validate the thesis that major AI companies should bring chip production in-house — a trend that would eventually hollow out foundry revenue across the board.
NVIDIA faces a different kind of threat. Tesla's AI5 is purpose-built for vehicle inference workloads and is projected to have a meaningful efficiency advantage over general-purpose NVIDIA silicon in that specific application. Tesla is already not buying H-series or B-series GPUs for vehicle compute. If Terafab gives Musk the production independence he needs, NVIDIA's remaining leverage — data center chips for Dojo training — also comes under pressure as xAI builds out its own silicon roadmap.
Intel is in a genuinely interesting position. A Tesla partnership would give Intel Foundry the anchor customer its business has been begging for. Intel badly needs someone credible to commit volume to its 18A node. Tesla badly needs someone who knows how to run a fab. The incentives align in a way that most foundry customer relationships do not.
Applied Materials, Lam Research, ASML would all benefit directly from a $20–25 billion greenfield facility. A new fab of this scale is a windfall for equipment suppliers regardless of whose name is on the door.
The American context matters too. CHIPS Act funding has been flowing to TSMC Arizona, Samsung Austin, and Intel Ohio. A Tesla fab in Austin would represent a different kind of investment: not a foreign company building American capacity under political pressure, but an American company betting its strategic future on domestic semiconductor production. That story has real resonance in Washington, and it would not be surprising if federal support — formal or informal — is part of what gets announced Saturday.
The Optimus Wildcard
There is a number buried in recent Tesla earnings disclosures that puts the Terafab ambition in sharper context.
Tesla has stated a target of producing 10 million Optimus humanoid robots per year at full scale. Each robot requires custom AI chips for locomotion inference, sensor fusion, and real-time decision-making. At 10 million units annually, chip demand from Optimus alone would dwarf what Tesla currently buys from TSMC for its entire vehicle fleet.
Add the Cybercab robotaxi program, the Dojo supercomputer expansion, and xAI's Grok inference infrastructure — all of which Musk has said Terafab will supply — and the demand case for a facility of this scale is not as absurd as it first sounds. The question is not whether Tesla needs the chips. The question is whether Tesla can build a fab capable of making them, on anything approaching its stated timeline.
That distinction matters enormously. A company that needs chips and a company that can manufacture them at leading-edge nodes are separated by something close to the distance between wanting to build a rocket and actually putting one in orbit.
Musk knows that distance. He has crossed it before.
Is This Different From the Other Times?
Musk has a well-documented history with announcement timelines. Full Self-Driving was "one year away" for several consecutive years. The Cybertruck was first shown in 2019 and delivered at scale in 2023. The Roadster 2.0 remains, as of this writing, undelivered.
The pattern is real. But the Terafab context has a few features that separate it from those examples.
First, the investment is not hypothetical. $20–25 billion is not a sketch on a whiteboard. Site preparation is already underway in Austin. Construction contracts, if not already signed, are imminent.
Second, the timeline pressure is structural. Musk is not announcing Terafab for marketing reasons. He is announcing it because TSMC's capacity allocation is a genuine ceiling on Tesla's AI roadmap, and that ceiling arrives on a fixed schedule regardless of what Musk does or does not build.
Third, the Intel angle — if confirmed — would provide something Musk's other moonshots have not always had: an experienced operational partner with skin in the game.
None of that means Terafab will hit one million wafer starts per month by 2028. It probably will not. But the more relevant question is whether it achieves enough capacity, fast enough, to give Tesla meaningful independence from external suppliers before the supply constraint arrives. That bar is considerably lower than the stated targets, and considerably more achievable.
The NVIDIA GTC 2026 conference — running this week in San Jose — has been dominated by Jensen Huang's own chip roadmap. The contrast between Huang's posture (we will supply the world's AI compute) and Musk's posture (we will stop relying on you) tells you more about the direction of the industry than any single product announcement.
The Bottom Line
What launches Saturday is almost certainly not a chip factory. It is a commitment, a groundbreaking, a line in the sand.
What it represents is more significant than what it literally is. Tesla is declaring that the model of AI companies depending on a single Taiwanese foundry for the silicon that runs their most critical systems is not sustainable — and that it intends to be the first major AI-hardware company to build a real alternative on American soil.
The semiconductor industry has heard ambitious promises before. It has also watched, more than once, as a company everyone wrote off as overreaching quietly learned to do something no one believed it could.
Musk has earned the benefit of the doubt on exactly that kind of underestimation, and almost nowhere else.
The chips will not ship Saturday. But the direction is clear, the capital is real, and the need is urgent. Whether Tesla can execute on the engineering is the only question that matters now, and March 21 will not answer it.
That answer comes in about three years.
Sources
- Musk Says Tesla Needs to Build 'TeraFab' to Manufacture Chips — Bloomberg
- Musk Says Tesla's Mega AI Chip Fab Project to Launch in Seven Days — Ground News
- Tesla Terafab Set for Launch: Inside the $20B AI Chip Factory — Teslarati
- NVIDIA CEO Jensen Huang Warns Against 'Extremely Hard' Challenge — Tom's Hardware
- Inside Tesla's 'Terafab': Musk's Bold AI Chip Plan and Possible Intel Partnership — TechReport
- Musk: Tesla to Launch 'Terafab' AI Chip Factory Project Next Week — Not a Tesla App
- Tesla Terafab Project: Elon Musk Confirms Launch in Seven Days — FinTech Weekly
- Tesla's $25B Terafab AI Chip Project — Phemex Academy
- Tesla Terafab Project Launches in 7 Days: What We Know — Basenor
- Elon Musk Announces 'Terafab Project' Chipmaking Venture to Launch in Seven Days — Creati.ai