OpenAI's Jalapeño inference ASIC is being deployed internally alongside AMD EPYC Turin hosts, each with 1.5 TB of memory, according to Tom's Hardware. In an interview published Friday, October 2, 2026, Richard Ho, OpenAI's VP and head of hardware, called the choice pragmatic and said Nvidia's standalone Vera CPU was still a little behind at that maturity level.
The chip itself is not new. OpenAI and Broadcom unveiled Jalapeño on June 24, 2026, as the company's first Intelligence Processor, an accelerator designed for language-model inference, with a goal of gigawatt-scale deployment with data-center partners. What is new is the server design around it: the host CPU is a Turin-generation EPYC, not Nvidia's Vera.
What Ho said
Ho told Tom's Hardware that the Jalapeño program tried to cut risk and move the design faster. According to the interview, standalone Vera did not yet have the desired maturity, while Turin “did what we needed” and partners already had experience with the platform. The outlet also attributes the rack description to SemiAnalysis, consulted after the chip was presented.
The comment is not a verdict that Vera is inferior in final performance. It is a statement about maturity and schedule risk for this specific program. OpenAI has not, as of this article, published its own notice detailing the 1.5 TB-per-host configuration.
Why it matters
Custom accelerators still depend on a host CPU for orchestration, memory, and I/O. Choosing Turin over Vera shows that even an OpenAI-linked inference project does not have to be Nvidia end to end. For the AI server market, it is another sign that platform maturity and integrator experience weigh as much as the announced spec sheet.
Sources
Transparency: This content was created, edited, or reviewed with the help of artificial intelligence. Information was cross-checked with public posts on X and sources available on the internet. Check the original sources for the full context.
By GeekikiBot