OpenAI pairs its Jalapeño ASIC with AMD Turin to avoid first-generation CPU risk

"Vera, as a standalone, is a little bit behind on that maturity level," Ho told Tom's Hardware ^[4]. Vera is Nvidia's first custom CPU core for the data center ^[8]. Tom's Hardware reads the remark as a comparison of Turin with Vera, and not of AArch64 with x86 ^[7]. The quote supports that reading ^[4].

Ho framed the design around schedule. "The way we approached that design was really in terms of de-risking and being able to do that design fast," he said ^[3]. Of Turin, he said: "It did what we needed to do, and partly our partners had some experience with it" ^[5]. I'd expect that second point to weigh more than any benchmark. SemiAnalysis described the deployment as rack-scale ^[16], with 1.5TB of memory on every host ^[1]. Someone has to bring that configuration up and support it, and partners with Turin experience remove one unknown from the schedule.

Nvidia has promoted Vera with a 1.8x improvement over a Turin chip, a number pulled from a suite in which Vera finishes 3% ahead of AMD's EPYC 9755 overall ^[9]. The headline figure is an 80% lead ^[1]. For it to transfer to a Jalapeño host, the host's work would have to resemble the specific test that produced it. Of the two numbers, the 3% is the one I would put in a capacity plan. Arm's claim that its AGI CPU delivers more than twice the performance of modern x86 platforms rests on internal estimates ^[10]. Both chips are pitched as "agentic" CPUs that purportedly speed up reasoning loops ^[12]. Vera has yet to be shown across a wide variety of workloads, or against AMD's Venice and Intel's Diamond Rapids ^[11].

Serviceability points the same way. Vera is mounted on the board and Turin sits in a socket ^[13]. Server CPUs are rarely swapped, but when one has to be, Turin makes the job easier ^[13]. Tom's Hardware connects that difference to the partner experience Ho cited ^[13].

OpenAI still has Arm CPUs in its plans. It is one of Arm's deployment partners for AGI and, according to Nvidia, among the customers "exploring the Vera CPU" ^[14]. Frontier labs and hyperscalers keep a wide variety of hardware in their fleets ^[15]. The evidence supports a narrow conclusion. A team building a custom accelerator on a tight schedule can take the established host for that program and keep Arm parts in the plan. It does not show OpenAI judging x86 the safer architecture. Ho scoped his answer to one program. "For the Jalapeño program, we were trying to make very pragmatic decisions. We wanted to be aggressive on the goals of the performance and the cost, but we didn't want to take unnecessary risks," he said ^[6].