Tesla Shifts 2nm AI Chip Production to U.S., Splits AI6 to Samsung Texas, AI6.5 to TSMC

Tesla’s next‑gen AI silicon will be a true “Made‑in‑America” effort, with the company carving the 2 nm process between two rivals on U.S. soil. According to Wccftech, the AI6 design will be fabricated by Samsung at its new Austin fab, while the slightly tweaked AI6.5 will land at TSMC’s Arizona plant. The split mirrors Musk’s broader strategy of diversifying supply chains after the AI5 tape‑out, which was a single‑sourced Samsung run‑off. By spreading production across the two most advanced 2 nm lines in the United States, Tesla hopes to hedge against capacity bottlenecks and keep its autonomous‑driving stack humming.

The two chips aren’t just copies of each other; each carries a distinct feature set. AI6, Samsung’s version, will pair the 2 nm node with LPDDR6 memory, a move that should boost bandwidth for the massive neural nets powering Full Self‑Driving (FSD). Wccftech notes that the AI6.5, built by TSMC, will “improve performance” – a vague but telling phrase that likely points to higher clock speeds or better power efficiency, given TSMC’s recent gains in yield and transistor density at the same node. Musk has hinted that the AI6.5 will feed the upcoming Dojo3 supercomputer, the next iteration of Tesla’s in‑house training cluster, suggesting a tighter integration between training and inference hardware.

Putting both chips in the United States also aligns with the company’s political calculus. Tesla has repeatedly warned that reliance on overseas fabs could expose its autonomous‑driving roadmap to geopolitical risk. By anchoring production in Texas and Arizona, the automaker not only shortens the logistics chain but also taps into the U.S. government’s push for domestic semiconductor capacity. The decision arrives just months after the U.S. Inflation Reduction Act opened new tax credits for chips made on American soil, a benefit Tesla can now claim for both Samsung’s and TSMC’s U.S. fabs, according to the same Wccftech report.

Musk’s roadmap doesn’t stop at AI6 and AI6.5. He has already teased Dojo3, a next‑generation training system that will lean heavily on the new silicon. The AI6 chip will likely serve as the inference workhorse in Tesla’s fleet, while AI6.5 powers the training backbone, creating a closed loop where data collected from cars feeds the supercomputer, which in turn refines the models that run on the vehicles. By splitting production between two foundries, Tesla can iterate faster, swapping in incremental improvements from either Samsung or TSMC without being locked into a single vendor’s roadmap. The move underscores a broader industry trend: AI‑centric companies are betting on multi‑foundry strategies to stay ahead in the race for ever‑more efficient, on‑device intelligence.