Elon Musk on Wednesday showed one of the first samples of Tesla’s AI5 hardware, giving the clearest public look yet at a chip he says will power the next wave of the company’s cars, Optimus robots and possibly xAI data centers.
Musk said the AI5 processor is about half the reticle size and uses industry-standard memory. He also said the chip can be up to 40 times faster than AI4 in certain scenarios, a claim he tied to Tesla’s push to move more of its AI work onto custom silicon. In an X post, he congratulated the Tesla_AI chip design team on taping out AI5 and said AI6, Dojo 3 and other chips were in the works.
The sample he displayed appears to be an already fabricated processor marked “KR 2613,” a label that suggests packaging in the 13th week of 2026. The module shows a fairly small ASIC die surrounded by 12 memory packages from SK hynix on an organic substrate, with the memory parts marked like standard DRAM products. If the module uses 12 GDDR6 or GDDR7 memory ICs, the AI5 ASIC may have a 384-bit memory interface, putting memory bandwidth somewhere between 768 GB/s and 1.536 GB/s depending on the memory type.
That matters because Musk has repeatedly framed AI5 as a major step up from the company’s current hardware. During Tesla’s Q3 2025 earnings call, he said by some metrics the new chip would be 40 times better than AI4. On Wednesday, he repeated that view and thanked Taiwan Semiconductor and Samsung for support in bringing the chip to production, saying it would be one of the most produced AI chips ever.
The back end of the story is less tidy than the announcement. Last August, there were reports that Tesla’s Dojo wafer-level processor effort had been abandoned and the team dismantled, and Peter Bannon, who led the project, listed his retirement on LinkedIn around the same time. Musk’s latest comments suggest Tesla has not given up on Dojo entirely: he said Dojo 3 is still in the works, even as he points investors toward AI5 and the broader custom-chip program.
Tesla previously said AI5 would be made by both TSMC and Samsung Foundry, but the current sample shown by Musk does not identify which company made it. If Tesla received the chip in March or early April and no re-spin is needed, the processor could reach deployment in 2027. For a company that has spent years promising more from its in-house chips than the market expects, that timetable may be the real test.
