OpenAI deploys Jalapeño ASIC processors alongside AMD EPYC 'Turin' CPUs in rack-scale configurations, bypassing Nvidia's high-performance agentic chips (Vera, Arm AGI).

OpenAI is deploying its new Jalapeño ASIC internally alongside AMD EPYC Turin processors as host CPUs, each configured with 1.5TB of memory. According to Richard Ho, VP and Head of Hardware at OpenAI, the decision to pair Jalapeño with Turin was driven by pragmatism and risk management.

"For the Jalapeño program, we were trying to make very pragmatic decisions," Ho explained in an interview with Tom's Hardware Premium. "We wanted to be aggressive on the goals of performance and cost, but we didn't want to take unnecessary risks. Vera, as a standalone, is a little bit behind on that maturity level. The Turin device is strong. It did what we needed to do, and partly our partners had some experience with it."

The choice reflects OpenAI's cautious approach despite Nvidia's close relationship with the company and the emergence of Arm-based CPUs designed specifically for agentic workloads. Vera and Arm's own AGI CPU are positioned as accelerators for complex reasoning in agentic loops compared to x86 alternatives from AMD and Intel. Other hyperscalers deploy Arm-based processors internally—Google uses Axiom, AWS runs Graviton—but the new Vera and AGI represent a shift toward peak performance messaging rather than efficiency optimization.

Arm has claimed that AGI delivers over twice the performance of modern x86 platforms, though this figure is based on internal estimates rather than independent benchmarks. Despite $2 billion in commitments, analysts project the market penetration will remain in the low single digits within two years. Nvidia, meanwhile, has marketed Vera aggressively, claiming up to 1.8x improvements over competing AMD Turin systems. However, this headline figure comes from dividing per-core scores of SPEC CPU 2026; on overall benchmark performance, Vera shows only a 3% advantage.

The Turin-versus-Vera decision ultimately came down to platform maturity. x86_64 has a far deeper foundation than AArch64, with decades of datacenter deployment experience behind it. Arm's 64-bit extension arrived 12 years after x86's introduction, and despite growing datacenter presence over two decades, the x86 ecosystem remains more established.

Practical considerations also favored Turin. Vera is board-mounted while Turin processors use standard sockets, making chip replacement simpler should service be required—a distinction rarely relevant in production server environments but still a point in Turin's favor. Ho's reference to partner experience with Turin underscores that OpenAI's hardware team prioritized proven operational patterns over cutting-edge performance claims.

As OpenAI's hardware fleet evolves, Jalapeño deployments with alternative host systems—potentially Vera or AGI—remain possible. For the current rack-scale rollout, Turin proved the lower-risk choice.