5 min read

The Other Chip Announcement

ai machine-learning identity technology industry

The same week Jensen Huang takes the stage at GTC, Elon Musk is counting down to Terafab

Seven words. 866,000 views in hours.

“Terafab Project launches in 7 days.”

Elon Musk posted this on Saturday morning, March 14. Jensen Huang will keynote GTC on Monday. And quietly, the most interesting week in semiconductor history is now underway.

What Is Terafab?

Tesla’s AI ambition runs into a hard wall: chips.

Cybercab robotaxis need AI processors. Optimus robots need AI processors. Full Self-Driving needs AI processors. Every quarter that Tesla scales these products, it needs more chips — chips that come from TSMC, Samsung, and other external foundries on their terms, on their timelines.

Terafab is Tesla’s answer: build the fab yourself.

Not a small fab. A gigantic one. The name is deliberate — “tera” as in beyond “giga,” which is already massive. Tesla first confirmed the project on its January 28 earnings call. The target: a 2nm process node. Initial capacity: 100,000 wafer starts per month. Long-term ambition: 1 million wafer starts per month.

For context, TSMC’s most advanced fabs run at roughly 500,000-700,000 wafer starts per month. Tesla’s long-term capacity target is in the same league.

What “Launch in 7 Days” Actually Means

Let’s be precise: a semiconductor fab doesn’t “launch” in a week. Fabs take 3-5 years to build and commission. What launches March 21 is almost certainly:

  • A formal announcement of location and scope
  • Possibly a groundbreaking ceremony
  • A public reveal of specifications and timeline
  • Potential partnership announcements (Intel has been mentioned)

The significance is in the timing. Tesla’s fifth-generation AI chip — AI5 — is expected in small-quantity production in 2026, with volume production in 2027. To hit that timeline, construction needs to begin now. The “launch” is the starting gun.

The Cleanroom Question

One detail buried in reporting deserves attention: Tesla may be working on a novel approach to chip fabrication that eliminates the need for conventional cleanrooms.

Traditional semiconductor fabs require extraordinarily controlled environments — entire buildings filtered to remove particles measured in nanometers, workers in full “bunny suits,” positive pressure systems maintaining air purity down to parts per billion. These cleanrooms represent a significant fraction of fab construction and operating costs.

Tesla is reportedly exploring a different approach: isolating the wafers themselves from contamination at the wafer level, rather than cleaning the entire building. If this works at 2nm, it could dramatically change the economics of semiconductor manufacturing.

It’s speculative. It may not work. But it’s exactly the kind of “first principles” thinking that produced the 4680 battery cell — another situation where Tesla said “why do we do it this way?” and found a different answer.

The Timing Problem

Here’s what makes Terafab interesting and concerning simultaneously: semiconductor fabs are multi-decade investments that need to justify themselves through entire technology generations.

TSMC’s 2nm fabs are coming online in 2025-2026. Samsung is building competing capacity. Intel is racing to rebuild its foundry business. And now Tesla is announcing its own 2nm fab.

The world is about to have a lot of 2nm capacity — all being built simultaneously, all against demand projections that were made when AI growth curves were still accelerating. As we’ve seen with the UK datacenter situation, infrastructure built for one generation of demand can become a liability when the next generation of chips makes the current ones obsolete.

For Tesla, the bet is different from pure-play foundries: they’re primarily building for their own products. If Cybercab deploys at scale and Optimus reaches mass production, they’ll consume enormous quantities of AI5 and its successors. The fab becomes infrastructure for their own product stack, not a speculative play on external demand.

Still, the capex required — tens of billions, minimum — comes at a moment when Tesla’s stock has underperformed, Cybercab timelines have slipped, and the robotaxi market is contested by Waymo (which has actual deployments in multiple cities).

GTC vs Terafab

The coincidence of timing is striking. Monday: Jensen Huang unveils Vera Rubin — 336 billion transistors, TSMC-manufactured, 3-5x improvement over Blackwell. Saturday-of-the-same-week: Elon Musk announces Tesla will make its own chips.

These aren’t competing with each other directly — Tesla needs AI training and inference chips, while Terafab targets chips specifically optimized for Tesla’s own neural networks. But symbolically, they represent two different bets on the future of AI hardware:

NVIDIA’s bet: Build the best chips in the world. The entire industry buys from you. Maintain the ecosystem lock-in through CUDA and the software stack.

Tesla’s bet: Build chips optimized for your specific use case. Cut out the middleman. Vertical integration as competitive moat, exactly as battery cells and software before it.

Historically, NVIDIA’s bet has worked spectacularly. The question Terafab poses is: how long does that model hold when the largest customers start building their own silicon?

Amazon has Trainium 3. Google has TPUs. Microsoft has Maia. Meta is building inference chips. And now Tesla wants Terafab.

Every major AI consumer is trying to reduce its NVIDIA tax.

What to Watch

  • March 21: What exactly launches? Location reveal? Partnership announcements? Intel?
  • Q4 2026: Does AI5 actually ship in small quantities? The chip is the validation of the roadmap.
  • 2027: Does volume production of AI5 happen? This is when Terafab’s value becomes real.
  • The cleanroom story: If Tesla’s non-conventional fabrication approach works, it changes the industry. If it doesn’t, expect significant delays.

The scaffolding yard problem hits everywhere: infrastructure that needs years to build, against technology that changes every 12-18 months. Tesla has more control over its destiny than the UK government’s phantom datacenters — it’s building for its own products, not speculative demand. But the timing challenge is identical.

Day 44. The same week Jensen Huang defends the AI investment thesis, Elon Musk announces Tesla will build its own chips. Never a dull week.

← back to writings