Tesla to widen hardware upgrade plan as AI chip work advances

Tesla is looking at a wide hardware upgrade for its vehicle fleet and moving forward with a new lineup of custom chips, according to comments from Chief Executive Elon Musk on the company’s most recent earnings call. The plan would touch nearly every car built with camera-based driver assistance, and it comes as Tesla works through a chip roadmap that includes AI4, AI5 and AI6.

Musk lays out the plan for older cars

Musk told analysts on the call that Tesla wants to bring older vehicles up to its newest computing hardware, rather than stopping at a single hardware tier. In his own words:

“I think it will make sense to upgrade all cars that have less than HW4. Any cars that have cameras basically. I think we want to upgrade them to the next-gen AI board. We have an upgraded AI4 chip around the middle of next year. Then AI5, which hopefully will be in volume production in middle of next year. That will initially go into Optimus. AI6 will be the best edge computing in the world. It’s just so good. Things are going really well on the chip end.”

That statement lands months after Musk confirmed Hardware 3 cannot reach Unsupervised FSD, and after Tesla began rolling out a scaled-back FSD v14 Lite build to US HW3 owners in June, a fix that still falls short of the Unsupervised capability many of those owners originally paid for.

Where AI4 and the coming AI4+ stand today

AI4 is the computer now running in Model 3, Model Y and Cybertruck, and it handles the Supervised version of full self-driving on current production cars. Musk has said before that AI4 alone can push safety results well past human driving averages, so the newer chips are framed as gains in speed and headroom rather than a fix for something broken.

An in-between chip, sometimes called AI4+, is set for production around the middle of next year, and it doubles memory over AI4 with a smaller bump in raw compute and bandwidth. This chip sits between the AI4 systems shipping now and the far more capable AI5 platform built for both cars and Optimus.

Tesla taped out the AI5 chip in April, marking the point where a design is locked and handed to manufacturing partners. Samsung and TSMC are both working on production, giving Tesla two sources for the chip. Musk has compared a single AI5 package to Nvidia’s H100 in Tesla-specific workloads, and he has said a dual-chip AI5 setup can match newer Blackwell-class hardware at a fraction of the cost and power draw.

Even so, getting AI5 into cars at scale is a longer process than the tape-out headline might imply. Musk called the design “finished” in mid-2025, then said it was “almost done” again in January 2026, and volume production for vehicles has now slipped to around mid-2027. Tesla has said it needs several hundred thousand completed AI5 boards ready on the line before it can switch over production, so the Cybercab is launching on AI4 hardware instead of waiting for the newer chip.

Once AI5 does reach scale, Musk has said it will go first into Optimus, the humanoid robot program, before it becomes standard equipment across the vehicle lineup. Tesla previously said AI5 could offer somewhere between ten and forty times the useful performance of AI4, though those figures have shifted some as the design matured.

AI6 moves from cars to robots and data centers

Musk’s remarks on the roadmap have changed a bit over time. He once described AI6 as a chip destined for vehicles as well as data centers, but by January he was saying AI6 would go mainly into Optimus and Tesla’s data center work, not into cars. Under that framing, AI5 becomes the last major hardware jump planned for vehicles in the near term, and later chips push forward the software and robotics side instead.

Tape-out for AI6 could land around December, Musk said in March, and Samsung has signed on to build the chip using 2-nanometer process technology, in a deal Reuters put at roughly 16.5 billion dollars. That timeline has already slipped some, with reports pointing to yield issues on Samsung’s 2nm line pushing AI6 mass production toward the end of 2027. Musk has said a single AI6 chip, built on the same footprint, could match the output of a dual-chip AI5 system, which would double edge compute without a larger package.

Musk has talked about compressing Tesla’s chip design cycle down to roughly 9 months per generation, well ahead of the eighteen to twenty-four month pace typical for major chipmakers. Beyond AI6, he has pointed to AI7, AI8 and even AI9 on the horizon, with AI7 floated as a candidate for space-based compute tied to Starlink and Starship. Musk’s stated goal is for Tesla’s chips to end up as the highest-volume AI silicon in the world, spread across millions of cars and, eventually, a large number of Optimus units.

Backing all of this is Terafab, a large in-house chip fabrication effort near Tesla’s Austin campus, built with SpaceX and, more recently, Intel as a manufacturing partner. The project is described as bigger in scope than Tesla’s existing battery plants and will need tens of billions of dollars in spending over the next several years.