OpenAI runs Jalapeño ASICs on AMD Turin hosts
OpenAI's Jalapeño custom inference chips are deployed in racks with AMD EPYC Turin CPUs, hardware VP Richard Ho told Tom's Hardware. He called Nvidia's Vera CPU a bit behind on maturity for that design. The custom-chip race is now a live equity and CapEx story beyond a single GPU vendor.
OpenAI has begun deploying its Jalapeño ASICs alongside AMD EPYC Turin host CPUs with 1.5TB of memory per host, Hardware VP Richard Ho told Tom's Hardware in an interview published Oct. 2. Ho called the Turin choice pragmatic, saying Nvidia's new Vera CPU was "a little bit behind… on that maturity level" for a fast, de-risked rack design even though OpenAI also works closely with Nvidia. Weekend equity futures are quiet; the CapEx narrative carries into Monday.
The comments underline how frontier labs now mix custom accelerators with mature x86 hosts rather than waiting for every new Arm server CPU. SemiAnalysis had described the rack-scale Jalapeño layout after the chip's reveal; Ho's on-record interview confirms Turin is in the live design. Partners' prior Turin experience and socketed serviceability also factored in, he said.
OpenAI and Broadcom unveiled Jalapeño earlier as an LLM-optimized inference processor built for performance-per-watt gains and gigawatt-scale partner deployments beginning in 2026. Early testing claims substantially better performance per watt than current state-of-the-art, with a detailed technical report still pending. Equity traders still treat Nvidia as the bellwether, but host-CPU and ASIC choices increasingly decide who can ship capacity on time and at what inference cost.
Backstory: hyperscalers and labs have spent 2025–26 pushing custom silicon to diversify beyond a single GPU vendor. Jalapeño's Turin pairing is a maturity-and-speed decision, not a public break with Nvidia.
Where to watch: any follow-on comments on Vera or Arm AGI hosts, and partner deployment timelines for Jalapeño racks.
What we know:
OpenAI confirmed Jalapeño racks with AMD Turin hosts. Ho framed Turin as the lower-risk host versus Vera for this generation.
What we don't:
Production volumes, power and cost-per-token claims versus Nvidia GPUs, and whether Jalapeño later moves onto Vera or other Arm hosts.
Where outlets agree Tom's Hardware and OpenAI's Jalapeño materials agree OpenAI is fielding a Broadcom-built custom inference ASIC as part of a multi-generation platform.
Where they differ OpenAI frames Turin as the mature host choice while Nvidia still markets Vera as a performance leap for agentic workloads.