Tesla’s New AI Plan post Dojo
A few years ago, Elon Musk bet big on Dojo , a custom-built AI supercomputer. The idea was bold: instead of relying on Nvidia like everyone else, Tesla would design its own D1 chips , its own training cabinets, its own networking, even its own compiler. It was a full-stack vertical integration play — from the data collected by millions of Teslas, all the way down to the silicon.
The goal? Out-compute everyone else on self-driving video data and make Full Self-Driving smarter, faster.
But the world didn’t stand still. Nvidia kept releasing faster, more efficient GPUs like the H100 and B100. They improved so quickly — and came with a huge software ecosystem (CUDA) — that Dojo’s advantages started to evaporate. Dojo was powerful, but also narrowly specialized and slower to evolve.
By mid-2025, it was clear the economics didn’t work. Maintaining two separate AI chip architectures — one for cars, one for Dojo — was expensive. Talent drifted away. In August 2025, Tesla shut the Dojo team down.
Musk didn’t abandon custom chips altogether — he refocused. Tesla is now building AI5 and AI6 : new in-house chip designs that can handle both inference (real-time driving decisions) and some training . These chips are more flexible, cheaper to scale, and will be used in vehicles, humanoid robots, and data centers.
Instead of fabricating them at TSMC (as with Dojo’s D1), Tesla signed a $16.5 billion deal with Samsung to manufacture AI6 in Texas. Samsung offered something TSMC couldn’t: guaranteed capacity, U.S.-based production, competitive pricing, and a faster path to large-scale manufacturing.
For massive AI training, Tesla will now lean heavily on Nvidia and AMD clusters — the same hardware most of the industry uses — and possibly tap into Musk’s other big compute asset, xAI’s Colossus , a gigantic Nvidia-powered supercomputer.
Dojo was a classic case of vertical integration going a bridge too far: Tesla tried to own the full AI stack, but the speed of the external market — and Nvidia’s dominance — made it unsustainable. The pivot to AI6 keeps Tesla in control of the chips inside its products while letting the heavy training work run on the best external hardware available.
It’s no longer about out-Nvidia-ing Nvidia. It’s about owning the parts of the stack where Tesla can truly differentiate , and buying the rest.
