Tesla's AI5 Is Reportedly in Trial Production at Samsung. Musk Says the Cars Do Not Need It.

Samsung has started putting Tesla's AI5 through its fab in Taylor, Texas. South Korea's Seoul Economic Daily reported the trial run on 16 September 2026 and Digitimes and the Tesla press followed within two days. It is the first work on Samsung's 2-nanometer process at that plant, and it is running earlier than the schedule called for.

What actually started

Trial production is not production. Samsung is making trial wafers and verifying yield — establishing what proportion of chips come off the line working — before it commits a line to volume. Samsung is aiming to finish that verification by the end of 2026, with supply at scale during 2027.

The pull-forward is the genuinely new part. Taylor had been due to enter full operation around November; demand for AI silicon from large technology buyers moved the timetable up.

Tesla is not betting on one foundry. Musk has confirmed that AI5 and its successor will be built both at Samsung's Taylor plant and at TSMC's in Arizona.

The chip is not for your car, or any owner's

This is the part that gets reported backwards, so it is worth being blunt about. Musk has said AI5 is not going into Tesla's vehicles. It is for Optimus and for the company's training clusters. His stated reason is that AI4 — the computer already fitted to every Model 3, Model Y, Model S, Model X and Cybertruck being sold — is enough to achieve much better than human safety for FSD.

The Cybercab, the car designed around having no driver at all, is also launching on AI4.

So a European owner who reads that Tesla's next AI chip has entered production should not read it as news about their car, or as a hint that the hardware in their driveway is about to be superseded. Tesla's public position is the opposite one. The retrofit question that owners of older HW3 cars have been chasing is a separate matter, and AI5 does not touch it.

What it is actually for

AI5 is quoted at roughly five times AI4's performance — Tesla's framing is that one AI5 delivers about five times the compute of two AI4 chips combined. A gain on that scale matters most where Tesla is limited by compute rather than by sensors: training the models, and running a humanoid robot that has to manipulate the world rather than drive through it. It is the same destination JPMorgan's analysts sketched out in August, and the same race XPeng joined when it opened a humanoid production line this month.

The timing deserves the same caution. Tesla has said it needs several hundred thousand finished AI5 boards line-side before it switches production over, and does not expect that volume before mid-2027. Trial wafers in September 2026 are a long way short of that — and AI5 taped out in April 2026, which Electrek noted at the time was close to two years later than Tesla originally planned.