OpenAI is running its custom Jalapeño inference chips alongside AMD EPYC "Turin" processors as host CPUs, and its hardware chief said Nvidia's competing Vera CPU was not mature enough for the first deployment, Tom's Hardware reported Friday. Richard Ho, OpenAI's vice president and head of hardware, described the choice as a pragmatic way to reduce risk on a fast-moving program.

What Happened

Research firm SemiAnalysis described the rack-scale setup after OpenAI unveiled Jalapeño: racks of the inference accelerators paired with racks of Turin servers, each host fitted with 1.5TB of memory. Tom's Hardware asked Ho why OpenAI did not use Vera, given its long working relationship with Nvidia.

Ho said the design was approached with de-risking and speed in mind. He said Vera, taken as a standalone part, is "a little bit behind" on maturity, while Turin did what OpenAI needed and the company's partners already had experience with it. He added that the Jalapeño team wanted aggressive performance and cost targets but did not want to take unnecessary risks.

RelatedNvidia Plans to Invest in Inference Chip Startup d-Matrix, The Information Reports

The choice is notable because OpenAI appears on Nvidia's list of customers exploring Vera and is also a deployment partner for Arm's own data center processor, the AGI CPU. Both chips are Arm-based and have been marketed as CPUs for agentic AI, meant to speed the reasoning loops that agents run.

Why It Matters

Custom accelerators get most of the attention, but every rack still needs general-purpose host processors to feed data, run orchestration and manage memory. That host market has become contested this year as server CPU demand rose with agentic workloads. Nvidia and Arm are trying to take share from AMD and Intel x86 chips, and a marquee AI lab picking AMD for its first in-house inference deployment undercuts their pitch.

The performance claims in that fight are disputed. Nvidia has said Vera delivers a 1.8x improvement over a competing AMD Turin chip, but Tom's Hardware noted that the figure comes from a benchmark suite in which Vera finished only about 3% ahead of AMD's EPYC 9755 overall. Arm says its AGI CPU offers more than twice the performance of current x86 platforms, based on internal estimates rather than published benchmarks. Arm has about $2 billion in commitments for the chip, though analysts expect its market share to remain in the low single digits after two years, the publication said.

There are also practical reasons for the choice. Vera is mounted directly on the board, while Turin chips sit in sockets, which makes replacing a failed processor easier, Tom's Hardware noted. Ho's comments, the publication said, were about Vera as Nvidia's first custom data center CPU core rather than a broader verdict on Arm versus x86.

Context

Jalapeño is OpenAI's first custom chip and was detailed at the Hot Chips conference this year. OpenAI has said it is for internal use, though Ho told Tom's Hardware in a separate interview that it could be used elsewhere, with OpenAI's own compute needs taking priority. AI labs and cloud providers are building their own accelerators to lower the cost of serving models, but most still buy their host CPUs from the established vendors.

What to Watch

Tom's Hardware said OpenAI keeps a wide range of hardware and could pair Jalapeño with a different host later, including Vera or Arm's AGI CPU. Later Jalapeño racks will show whether Nvidia closes the maturity gap Ho described. Benchmarks against AMD's next-generation Venice processors and Intel's Diamond Rapids, neither yet widely tested against Vera, will shape the next round of host decisions.