Semiconductor analysis firm SemiAnalysis has upgraded its outlook on AMD's chances of challenging Nvidia's software dominance from a 0% chance in its first critical report to what it now calls a great chance of success, provided AMD clears two remaining risks. The firm's review of AMD's MI455X accelerator, unveiled at Advancing AI 2026, found it is the first datacenter chip built on TSMC's 2-nanometer N2 process and packs 12 stacks of HBM4 memory for 432 gigabytes per package, more than the 8 stacks and 288 gigabytes Nvidia and Google use. Anthropic has committed to deploying 2 gigawatts of AMD chips, and Microsoft, which had dropped AMD after quality problems with an earlier chip, has reversed course and will deploy the new Helios racks, likely for OpenAI as the end customer. SemiAnalysis flagged two risks holding AMD back: a slow production ramp for the Helios rack due to reliability issues in its backplane design, and a chronic shortage of internal GPU clusters for software testing that has left AMD's automated test parity with Nvidia's CUDA behind schedule, with a target originally set for this event now pushed to October. AMD also announced a partnership with Cerebras on disaggregated inference technology meant to speed up response times. Meta ordered a scaled-down version of the MI455 that SemiAnalysis says hurts AMD's chances there, though it reports Meta has since begun reconsidering its infrastructure strategy after the firm's earlier criticism.
Why it matters · It matters that an analysis shop known for being tough on AMD's software has changed its tune, but the two risks it flags, production ramp and internal test infrastructure, are execution problems rather than chip specs, so the real test is still ahead.
Worth asking · SemiAnalysis, long skeptical of AMD's software, now calls its odds against Nvidia's CUDA moat a great chance, even as it flags production ramp and internal testing risks. Do you think AMD can actually close the software gap this cycle, or will execution problems trip it up again?