OpenAI's Jalapeño ASIC is already being deployed internally alongside servers using AMD EPYC Turin CPUs with 1.5 TB of memory each, according to an interview with Richard Ho, the company's VP and head of hardware, published by Tom's Hardware on October 2, 2026 and picked up by IT Home on October 3.
What happened?
Ho said the Turin choice was “pragmatic.” Nvidia's Vera CPU, he told the outlet, is “a little bit behind… on that maturity level” as a standalone host. Turin, he said, already did what the program needed, and OpenAI's partners had experience with the platform, which sped deployment.
The Jalapeño ASIC design, Ho said, prioritized de-risking and getting to silicon quickly. Performance and cost goals were aggressive, but the company did not want unnecessary risk in the CPU that hosts the accelerator. SemiAnalysis had described the rack-scale deployment after the chip was revealed; the interview confirms those racks already use Turin machines with 1.5 TB of memory.
The ASIC remains limited to OpenAI's internal use. Ho said there is no current plan to sell it externally, though he did not rule out broader use later if resources allow.
Why it matters
The Jalapeño ASIC is the inference accelerator OpenAI and Broadcom unveiled on June 24, 2026. In the official post, OpenAI said the chip went from design to tape-out in nine months, accelerated by its own models, and that engineering samples were already running machine-learning workloads at target frequency and power, including GPT-5.3-Codex-Spark. Broadcom CEO Hock Tan spoke of gigawatt-scale deployment with data-center partners beginning in 2026.
The October news is not the chip announcement itself. It is the host CPU next to it. In an AI rack, the CPU does not run the heavy model math: it manages memory, networking, preprocessing and the software that feeds the accelerators. Picking a mature EPYC Turin instead of waiting for Vera shows OpenAI preferred a shorter path to internal production over Nvidia's newer host platform.
What changes in practice?
For ChatGPT or API users, nothing changes immediately. The Jalapeño ASIC is not for sale, and OpenAI has not published a schedule for moving public traffic onto the ASIC. The interview shows private infrastructure already switched on, with a large memory pool per host node — 1.5 TB is server-class, not a desktop figure.
- Current use: internal only, according to Richard Ho.
- Host CPU: AMD EPYC Turin, with 1.5 TB of memory per machine.
- Alternative set aside on maturity grounds: Nvidia Vera, in OpenAI's assessment.
- Silicon partners already named: Broadcom for implementation and Celestica for board, rack and system.
Sources: Tom's Hardware, IT Home and OpenAI's announcement.
By GeekikiBot