Tesla

Tesla is the only car company that designs the inference silicon in its own vehicles, which is the part of its AI programme that has unambiguously worked. Every car it builds carries a custom NPU running vision networks in real time.

Dojo is the more instructive half. Tesla spent years building a bespoke training supercomputer around the D1 chip, then shut it down in August 2025, with Musk describing the architecture as an evolutionary dead end. The team was disbanded and training consolidated onto clustered AI5 and AI6 parts.

A programme called Dojo 3 was announced in January 2026, but it is a different thing wearing the same name, based on a future AI7 chip and framed around space-based compute. Reports that Dojo was simply cancelled, and reports that it was simply revived, are each half the story.

Silicon from Tesla

2 chips on record, newest first.

  1. 2021

    Tesla Dojo D1

    TSMC 7nm Training die, 354 processing nodes, 440 MB SRAM

    362 TFlops BF16/CFP8

    Absolute
  2. 2019

    Tesla FSD Computer (HW3)

    14nm FinFET Two neural network accelerators per chip, 96×96 MACs each

    96×96 MACs (36.8 TOPS/NNA)

    Absolute

Compare against every other vendor in the silicon tracker

Our coverage of Tesla