'Tesla's chip game is no joke': Elon Musk confirms it has restarted work on its biggest supercomputer yet - but what will it actually be used for?

Tesla restarts Dojo 3 after earlier supercomputer projects failed to meet expectationsAI5 chip performance will reportedly rival Nvidia Hopper while consuming less powerFuture chips, AI6 and AI7, are planned for incremental technical evolution

Tesla has restarted development of its Dojo 3 supercomputer project after shelving or abandoning earlier versions.

Elon Musk confirmed the move on X, linking the restart directly to progress on Tesla’s in-house AI5 chip.

Previous Dojo efforts failed to meet expectations, with Dojo 1 losing relevance quickly against Nvidia systems and Dojo 2 being cancelled before completion.

Tesla reboots Dojo 3 with ambitious in-house AI chips

Dojo 3 is framed as a recovery attempt rather than a clean breakthrough, as Tesla claims the technical foundation is now strong enough to justify reallocating engineers and capital back to the project.

Dojo 3 is expected to be Tesla’s first supercomputer built entirely on internal hardware, without relying on Nvidia components.

Earlier Dojo designs mixed Tesla silicon with external GPU products, limiting differentiation and control, whereas the new approach aligns chip design, system architecture, and software under one roof.

Tesla has openly recruited engineers to scale chip production, signalling ambitions for high-volume manufacturing.

Central to Dojo 3 is Tesla’s plan to release custom AI chips every nine months, although this will probably test the company’s resolve.

In terms of application, the AI4 and AI5 chips are linked to self-driving development and humanoid robotics, and AI6 is tied to Optimus and large-scale data center deployments.

Future iterations, including AI7, are already mapped out, although expectations point to incremental evolution rather than radical redesigns.

Beyond vehicles and robots, the supercomputer could support Tesla’s broader AI tools ecosystem, including training models that compete with established cloud providers.

These claims place Dojo 3 in direct competition with mature AI infrastructure vendors.

According to social commentator Nic Cruz Patane, “Tesla’s chip game is no joke,” noting that AI5’s performance is roughly comparable to Nvidia’s Hopper on a single chip, approaches Blackwell levels when paired, and runs at approximately 250W compared to H100’s 700W or Blackwell’s 1,000W+ at full specification.

Tesla claims its chip designs deliver similar output at lower wattage, but maintaining the planned release cycle will test its discipline and execution consistency.

Its technical promises are ambitious and financially driven, especially given the rising cost of external AI hardware.

Dojo 3 may reduce Tesla’s dependence on third-party silicon, but success will require consistency that earlier projects lacked.