Tesla Terafab Supply Chain

Three foundries

Tesla Buys Leading-Edge Logic From All Three Makers, and Terafab Ships Nothing Before 2028

Tesla designs its own AI inference chips and buys the wafers. It is now a customer of all three companies that build leading-edge logic in volume: Samsung Foundry, which makes AI4 at 7nm in Pyeongtaek, South Korea, and holds a $16.5 billion contract for AI5 and AI6 at Taylor, Texas; TSMC, which builds a second physical version of AI5 and, per Elon Musk, AI6 at its Arizona fabs; and Intel, which joined Terafab, the chipmaking program Tesla and SpaceX announced in Austin on March 21, 2026, as process partner on April 7, 2026.

Terafab is the long-term answer to foundry dependence, but it is late in the chip calendar. On Tesla's October 21, 2026 earnings call Musk said the Research Fab will begin commissioning machinery in the first half of 2027, make a first test chip "around May or June," and aim to make "some kind of useful chip at volume around Q3 of 2028." AI5 is due in Optimus "approximately April of next year." Every Tesla AI chip until Terafab ships at volume therefore comes from Samsung or TSMC.

Supply chain

Logic Has Three Routes; Memory Has One

• Made in US
• Some US volume
• US planned
• Made abroad
• Not disclosed

Logic has three routes. Memory has one: the same three DRAM makers that sell to every AI data center. Musk thanked Micron for allocating memory to Tesla on the July 2026 earnings call and said memory prices had climbed to unusual levels. Terafab's March 2026 plan listed memory alongside logic, masks and packaging, but its first stated output is logic test chips.

Chip family

Each Generation Has Its Own Foundry Path

Foundry partners

Samsung, TSMC and Intel Each Hold a Different Piece

Samsung Taylor

Samsung Taylor Is a Dedicated Line Tesla Does Not Own

Samsung and Tesla signed the Taylor contract on July 28, 2025: about $16.5 billion of supply through December 31, 2033. Samsung had planned Taylor for 4nm and moved the fab to 2nm gate-all-around after Tesla signed. Musk wrote that the fab "will be dedicated to making Tesla's next-generation AI6 chip" and that he would walk the line himself. AI5 was added to Taylor later.

Samsung held its equipment move-in ceremony on April 24, 2026, moved core engineers from Korea to secure first yields, and completed AI5 tape-out for its 2nm line in July 2026; Seoul Economic Daily reported trial AI5 production at Taylor in September 2026. Korean press reported Samsung 2nm yields rising from about 20% in late 2025 to the high 50s in 2026, against the roughly 60% usually needed for stable volume. Ramp production starts in 2027. Tesla gets most of the control of a captive fab while Samsung carries the capital; the AustinIO Samsung Taylor Fab page covers the campus and its supplier ring.

AI5

AI5 Is One Design Built at Two Foundries

AI5 is the first Tesla chip built at two foundries at once. Samsung and TSMC each turn the same design into a slightly different physical chip, and Musk said in November 2025 that Tesla's software "works identically" on both. One AI5 replaces the dual-SoC AI4 board. Tesla has published AI5 performance only as ratios: about five times the useful compute of a full AI4 board, and, on the October 2026 call, "an order of magnitude more compute" than AI4. Capacity per board rises about ninefold if scaled from AI4's 16GB, which lands AI5 in the middle of the AI-driven shortage of LPDDR5 and high-bandwidth memory. The AI5 supply rows (Samsung and TSMC logic, SK hynix and Samsung memory, packaging not disclosed) are on the Tesla AI5 supplier ring.

AI7 and D3

The Radiation-Tolerant Chip Is the One Only Terafab Will Make

AI7, which Musk also calls Dojo3 and which appeared as D3 at the Terafab launch, is the SpaceX chip in the family: radiation-tolerant compute for 100 kW orbital server satellites. Commercial foundries rarely maintain a radiation-tolerant library and process variant for one customer at satellite volumes, and that is why D3 sits at Terafab rather than Samsung or TSMC. Until then SpaceX's first orbital data center satellite, AI1, uses NVIDIA Rubin GPUs and Vera CPUs in its Starmind payload, announced August 4, 2026. Rad-Hard and Rad-Tolerant Chips covers the design rules.

Terafab

Terafab Splits Process Development From Volume Production

Musk announced Terafab at the former Seaholm Power Plant in Austin on March 21, 2026, as one program for chip design, masks, logic, memory, advanced packaging and test, with a target of 1 terawatt of new AI compute a year. On the October 2026 call he put capital equipment at "not less than tens of billions" through early 2028 and "north of $100 billion long term"; Tesla's finance chief said most Research Fab tool orders are already placed. SpaceX's May 20, 2026 S-1 filing describes "a general framework" and states that neither Tesla nor Intel is obligated to stay in the project. Bernstein estimated in March 2026 that 1 TW of compute a year would take 140 to 360 new fabs of 50,000 wafer starts a month. SpaceX describes the Gibbons Creek production fab as one building of about 100 million square feet, at least 2.5 miles long, on more than 7,720 acres, where the retired 470 MW coal plant supplies the power rights, water and land; Governor Abbott announced the $16.8B first phase on August 6, 2026 with 3,000 jobs committed. Site, permitting, power and water detail are on AustinIO at Terafab Research Fab Giga Texas and Terafab Gibbons Creek.

Memory

Memory Is the Dependency Tesla Cannot Route Around

Tesla can split logic between Samsung and TSMC and plan a third source in Terafab, but memory has no Tesla-specific route. AI5's memory is its largest bill-of-materials exposure, and Tesla buys from SK hynix, Samsung and Micron at the same time as the hyperscalers building AI data centers. Reports conflict on AI5's memory supplier: Hankyung named SK hynix and Newsis named Samsung LPDDR5X. See DRAM and HBM for the supply picture.

US status

AI5 and AI6 Bring Tesla Logic to Texas and Arizona; Memory Stays in Asia

• Some US volume
• US planned
• Made abroad
• Not disclosed

Related coverage

Related Coverage

• Spotlight: Tesla, A Rolling Data Center
• Humanoids and Robotics
• Samsung Semiconductor
• Intel Foundry
• TSMC
• Leading-Edge Logic Fabs
• AI Inference and Edge Compute SoCs
• Chip Supplier Rings