OpenAI's Jalapeño chip is already deployed internally alongside AMD EPYC Turin processors, each host with 1.5 TB of memory. Richard Ho, OpenAI's VP and head of hardware, told Tom's Hardware the choice was pragmatic and that Nvidia's Vera CPU was still a little behind on maturity for the program.
What happened?
OpenAI and Broadcom unveiled Jalapeño on June 24, 2026, as the company's first accelerator built for language-model inference. In an interview published on October 2, Ho confirmed the ASIC is already running in-house on Turin hosts. IT Home, citing the same conversation, reported that the chip remains limited to internal use and that the company does not yet plan to sell it, though Ho did not rule out wider use if capacity allows later.
Ho said the Jalapeño design prioritized de-risking and reaching deployment quickly. Performance and cost targets were aggressive, but the host platform had to fit what partners already knew how to run. Turin, he said, "did what we needed." On Vera, the line was direct: as a standalone product, it is "a little bit behind on that maturity level."
Why it matters
Inference is the stage where an already trained model answers requests. It is what consumes most of a chatbot's energy day to day. An ASIC is a chip designed for a specific task, rather than a general-purpose GPU. OpenAI said in the June announcement that Jalapeño was optimized for the kernels, memory movement and serving patterns of its own models.
The host decision matters because the accelerator does not work alone. The CPU prepares data, orchestrates the network and holds system memory. Choosing AMD, not Nvidia, for that role shows OpenAI is not tying the whole stack to a single supplier, even while Nvidia still dominates training GPUs. Tom's Hardware notes that Nvidia itself publishes Vera ahead on some benchmark cuts; Ho, though, treated operational maturity as the deciding criterion, not a single performance number.
What changes in practice?
For ChatGPT users, nothing changes this week. Jalapeño is not for sale, and the original announcement talks about gigawatt-scale deployment with data-center partners across generations, starting in 2026. What the October interview adds is the concrete shape of the first internal wave: AMD Turin hosts with 1.5 TB of memory, instead of Vera.
- Chip: Jalapeño, OpenAI's inference ASIC with Broadcom, unveiled on June 24, 2026.
- Host: AMD EPYC Turin, 1.5 TB of memory per system, according to Tom's Hardware.
- Out for now: Nvidia Vera, described by Ho as less mature for this program.
- External sales: not announced; current use is internal.
Sources: OpenAI announcement, Tom's Hardware and IT Home.
By GeekikiBot