OpenAI runs Jalapeño ASICs on AMD EPYC Turin hosts
OpenAI is deploying its Jalapeño inference ASICs beside AMD EPYC Turin CPUs with about 1.5 TB of host memory. Cloud buyers watch whether custom silicon undercuts Nvidia’s pricing power. Hardware VP Richard Ho called Nvidia’s Vera standalone “a little bit behind” on maturity for this rollout.
OpenAI has put its first custom Jalapeño inference ASICs into rack-scale deployment hosted by AMD EPYC Turin CPUs, each with about 1.5 TB of memory, Tom’s Hardware reported after interviewing Hardware VP Richard Ho. Yahoo Tech republished the same interview. Ho called the Turin pick “pragmatic” de-risking: the team wanted aggressive performance and cost goals without unnecessary platform risk.
Asked about Nvidia’s Arm-based Vera CPU — which Nvidia markets for agentic workloads and which OpenAI is also exploring — Ho said Vera standalone is “a little bit behind… on that maturity level” versus Turin. Partners already had Turin experience, and socketed x86 hosts are easier to service than board-mounted Vera in early fleets. OpenAI’s June Broadcom unveil framed Jalapeño as an LLM-optimized Intelligence Processor aimed at gigawatt-scale partner deployments. August first-results posts claimed 1.5–1.9× more AI work per watt and lower latency than comparison systems on public models, with a fuller technical report still promised.
The choice matters for every company renting AI compute. If frontier labs can field mature custom ASICs on proven hosts, Nvidia’s pricing power softens even while GPU demand stays huge. OpenAI remains an Nvidia customer and AGI deployment partner; Jalapeño-plus-Turin is a portfolio move, not a divorce.
Backstory: hyperscalers from Google to Amazon already run custom accelerators; OpenAI’s production ASIC puts a model lab in that club. Engage: Tom’s Hardware interview, OpenAI’s Broadcom unveil, and the first-results note.
What we know:
Jalapeño ASICs are deployed with Turin hosts; Ho cited maturity and partner experience; Vera was judged riskier for this program; OpenAI still explores other CPUs.
What we don’t:
Public third-party benchmarks matching OpenAI’s efficiency claims at fleet scale; exact fleet size; when Jalapeño might pair with Vera or Arm AGI instead.
Agree (full): Tom’s Hardware and Yahoo Tech agree OpenAI’s Jalapeño ASIC is live in deployment with AMD EPYC Turin hosts.
Differ (full): OpenAI frames a maturity-driven Turin pick; efficiency claims that Jalapeño beats Nvidia remain stronger in company and trade coverage than in independent third-party audits.
Where outlets agree Tom’s Hardware and Yahoo Tech agree OpenAI has deployed Jalapeño inference ASICs alongside AMD EPYC Turin hosts.
Where they differ OpenAI calls Turin the mature, low-risk host while its claims that Jalapeño beats Nvidia on efficiency remain stronger in company posts than in public third-party audits.