Tesla's AI5 chip has cleared a key production step at Samsung's Texas fab, several months after TSMC. The rare dual foundry setup signals very high expected production volumes.
James Kim, a principal engineer at Samsung Foundry, posted on LinkedIn on Monday that Tesla's AI5 chip has reached tape-out at Samsung's Taylor, Texas fab on a 2nm-class process, the first on-record confirmation that Tesla's flagship AI silicon has cleared the same engineering gate at a second foundry (Tom's Hardware). The Samsung tape-out lands "several months" after AI5 taped out at TSMC, consistent with Elon Musk's earlier statement that the chip would be produced concurrently at both foundries.
Dual-foundry qualification is reserved for chips expected to ship at very high scale. When a company pays to qualify the same custom AI silicon at two of the world's most advanced fabs on the same process class, it is signaling silicon volumes that no single foundry could comfortably absorb, and building insurance against the kind of capacity crunch that has derailed AI hardware rollouts in the past. Samsung's Taylor, Texas facility, the first U.S.-based leading-edge foundry Samsung has built, is also getting its first public leading-edge production win on home soil.
AI5 is Tesla's in-house AI accelerator, the successor to the silicon that currently powers its self-driving hardware. Musk demonstrated the first AI5 sample in mid-April 2026 and has said the chip will run in Tesla cars, the Optimus humanoid robot, and Tesla's own data centers. Tape-out is the design-to-manufacturing handoff: the point at which a chip's schematic is frozen and sent to a fab to print physical wafers. Reaching it means the design is mature enough to risk real silicon; volume production follows.
Samsung's framing leans confident. Kim's LinkedIn post called the tape-out a milestone for Samsung's 2nm process and said AI5 "will soon be integrated into Tesla's newest products." Some coverage has read that as a signal Samsung has put yield concerns behind it at Taylor, a question that has dogged the fab since construction began (Wccftech framed the news as "silencing yield doubters"), but no actual yield figures have been disclosed.
Two specific performance claims attached to AI5 belong to Musk, not to benchmarks. He has said AI5 can deliver up to 40 times the performance of its predecessor in certain workloads, and described it as one of the most-produced chips ever, the stated justification for the dual-foundry plan. Neither has been independently verified, and the chip's peak compute, memory bus width, and whether the SK hynix memory packages on the module are GDDR6 or GDDR7 all remain officially undisclosed (Tom's Hardware infers a 384-bit bus and 768 GB/s to 1.536 TB/s of bandwidth from the visible packaging, but that is engineering texture, not a confirmed spec sheet).
The production-start timeline is also thinner than the headlines suggest. A single LinkedIn post by a Samsung engineer is the entire public basis for the "production starts soon" claim; on-the-record confirmation from either Samsung's corporate communications or Tesla would tighten the story. The post does establish that AI5 has cleared a real engineering gate at a named fab on a named process, and that Tesla has paid to do that work twice on the leading edge. The next milestone to watch is when the first Samsung-made AI5 wafers actually come off the Taylor line.