OpenAI’s Jalapeño ASIC, the company’s first inference accelerator, is already in internal use alongside AMD EPYC Turin hosts with 1.5 TB of memory each. The detail comes from an interview with Richard Ho, VP and head of hardware, published by Tom’s Hardware and picked up by IT Home on October 3. The Jalapeño ASIC is not for sale.
What Ho said
Ho described the Turin choice as pragmatic. According to the report, he said Nvidia’s new Vera CPU for agents is “a little bit behind on that maturity level” as a standalone. The program, he said, was about de-risking and moving fast: aggressive performance and cost goals, without unnecessary platform risk. The chosen platform, he added, was strong and already familiar to partners.
Tom’s Hardware quotes Ho referring to the host as a “Turing device.” In the context of the story and of IT Home’s coverage, the CPU in question is EPYC Turin. The remark about Vera is a maturity judgment for OpenAI’s schedule, not a public benchmark released by the company.
What was already known
OpenAI and Broadcom unveiled the Jalapeño ASIC on June 24, 2026, as the company’s first Intelligence Processor, designed for language-model inference. In the official post, Ho said the architecture was optimized for the kernels, memory movement, networking, and serving patterns that matter most for frontier models. Broadcom CEO Hock Tan spoke of a multi-generation roadmap and gigawatt-scale data centers with Microsoft and other partners beginning in 2026. Tape-out, OpenAI said, took nine months.
Ho told Tom’s Hardware that current use is internal. A wider sale is not the immediate plan, he said, though it is not ruled out if resources allow later. IT Home summarizes the same position.
What is still reported, not confirmed
SemiAnalysis described a rack-scale deployment after the chip’s reveal, and that account is what Tom’s Hardware took to Ho. Tray layouts, chips per rack, and operations-per-watt gains appear in that coverage and in secondary summaries. OpenAI did not publish those figures in the June announcement. The rack topology should be treated as reported, not as an official spec.
Why it matters
The Jalapeño ASIC is OpenAI’s entry into its own inference silicon, alongside Broadcom. Picking an AMD host rather than Vera shows that the accelerator race does not by itself decide the CPU that holds the node. For anyone following AI hardware, the new point this week is status: the chip has left the announcement stage and is in internal use, still without a product for the market.
Sources: Tom’s Hardware, IT Home, and the OpenAI and Broadcom announcement.
By GeekikiBot
