Tesla is no longer just a car company that buys chips. With Terafab — a giant semiconductor complex rising in Grimes County, Texas — Elon Musk’s companies are building their own chip factory, designing their own silicon, and even rewriting how much RAM an AI chip really needs. It is one of the boldest manufacturing bets in tech, and in October 2026 it snapped into focus.
What exactly is Terafab?
Terafab is the planned Tesla and SpaceX chip complex in Texas, unveiled in March 2026. The headline numbers are pure Musk scale: an initial investment of US$16.8 billion, roughly 3,000 jobs, and a long-term build-out that Texas tax filings suggest could reach US$119 billion across four phases. The target output is just as outsized — around one terawatt of AI computing capacity per year.
The plan calls for two mass-production factories, each dedicated to a single chip design, plus a research fab where new designs can be tested fast. The silicon is destined for Tesla’s driver-assistance systems, the Cybercab robotaxi fleet, Optimus humanoid robots and xAI — and SpaceX even talks about feeding space-based data centers and a vast AI satellite constellation.
“We will build and run the fab”
In early October, speculation swirled that TSMC — the world’s dominant contract chipmaker — might end up owning or operating Terafab. Musk shut that down on October 7 in unmistakable terms: “No, we will build and run the fab. Let there be ZERO doubt about that.” The most TSMC could get, he added, is subleasing part of the site. The earlier talks with TSMC were, in his words, “just discussions.”
One partner is confirmed, though: Intel, which joined the project in April, contributes its advanced 14A manufacturing process technology. Intel CEO Lip-Bu Tan reaffirmed this month that Intel remains involved — a notable alliance between a resurgent Intel and the Musk empire.
The research fab: 10 chip designs a year
The newest piece of the puzzle is speed. Musk says the research wafer fab planned at Giga Texas aims to make and test about 10 chip designs per year — a radical pace in an industry where a single major chip typically takes around two years from design to volume production. The trick is putting everything under one roof: fabrication, testing, packaging and photomask production in the same building, so engineers can test a chip, cut new masks and spin another version without breaking the loop.
The approach is deliberately scrappy: start with off-the-shelf equipment, modify it to raise capacity, then build custom machines for the bottlenecks. And Musk is candid about the learning curve: “I have no idea how to make chips, but we’ll figure it out.”
The chips: AI5 and AI6
Terafab is years from full production, so Tesla’s next chips are being built with partners in the meantime. AI5 — the processor set to power the next generation of Tesla vehicles and Optimus — taped out 45 days ahead of schedule in April 2026. Built on a 2-nanometre process and reported to deliver 2,000–2,500 TOPS of inference performance, it is expected to enter volume production in 2027, with trial production already under way in Texas.
AI6 follows with roughly double the performance, and an optimized AI6.5 after that. Production is split across America’s two new chip strongholds: Samsung’s Taylor, Texas fab for AI6, and TSMC’s Arizona operation for AI6.5 — both on 2nm technology.
The RAM cut everyone is talking about
The most surprising move came on October 1, when Musk announced Tesla was slashing the memory on its own chips: AI5 drops from 144 GB to 72 GB of LPDDR5, and AI6 loses a third of its planned memory, landing at 144 GB of LPDDR6.
The reason is manufacturing reality. Memory is one of the hardest components to secure at the volumes Tesla needs for mass Optimus production, and halving the RAM per chip effectively doubles how many chips Tesla can ship — while cutting cost per unit. Musk insists the trade-off will have a negligible effect on Optimus performance. Tesla’s architecture also leans heavily on SRAM built into the chip itself: roughly half of the compute accelerators on these designs are dedicated to SRAM, which Musk says delivers an order of magnitude more effective bandwidth than DRAM for work that lives in cache.
Samsung’s $16.5 billion Texas bet
Underpinning all of this is the mega-deal signed in July 2025: Samsung will build Tesla’s AI chips under a contract worth about US$16.5 billion, running through the end of 2033 — a figure Musk has called a floor, not a ceiling. Samsung’s Taylor fab began moving in equipment on April 24, 2026, and its 2nm lines are being prepared first and foremost for Tesla silicon. For Samsung’s foundry division, long stuck in the red, Tesla is the anchor customer that changes the math.
Terafab at a glance
- Project: Terafab — Tesla & SpaceX semiconductor complex
- Location: Grimes County, Texas, plus a research fab at Giga Texas
- Investment: US$16.8 billion initially; filings point to up to US$119 billion over four phases
- Goal: ~1 terawatt of AI compute capacity per year; ~3,000 jobs
- Technology partner: Intel (14A process). TSMC: talks only, no operating role
- Chips: AI5 (2nm, 2,000–2,500 TOPS), AI6 (~2× AI5), AI6.5
- Memory: AI5 — 72 GB LPDDR5 · AI6 — 144 GB LPDDR6
- Manufacturing today: Samsung Taylor (Texas) and TSMC (Taiwan, then Arizona); Samsung deal ≈ US$16.5B through 2033
📚 Elon Musk by Walter Isaacson — the bestselling biography
Get it on Amazon
FAQ
Who will run Terafab — Tesla or TSMC?
Tesla and SpaceX themselves. Musk confirmed on October 7, 2026 that his companies will build and run the fab; TSMC might at most sublease part of the site.
Why did Tesla cut the RAM on AI5 and AI6?
To secure enough memory volume for mass Optimus robot production and to cut cost. AI5 goes from 144 GB to 72 GB of LPDDR5; AI6 drops to 144 GB of LPDDR6. Musk says the effect on performance should be negligible.
When will Terafab chips reach products?
Terafab itself is a multi-year build. In the meantime, AI5 enters volume production in 2027 via Samsung and TSMC, with AI6 and AI6.5 to follow — powering vehicles, Cybercab robotaxis and Optimus robots.
The verdict
Terafab is vertical integration taken to its logical extreme: the company that designs the cars, the robots and the AI models now wants to print the silicon too. The RAM cut shows the pragmatic side of the plan — Tesla will trade spec-sheet bragging rights for chips it can actually ship by the million. If the research fab hits anything close to 10 designs a year, the rest of the chip industry is about to feel very slow. Either way, Texas is becoming the center of gravity for AI hardware — and Tesla intends to own the factory floor.
