OpenAI says its first in-house AI chip went from blank slate to tape-out in nine months, and its hardware boss credits software, not silicon, for making that possible.
OpenAI VP of Hardware Richard Ho told Tom's Hardware that the Jalapeno inference ASIC, unveiled at Hot Chips in August 2026, was built chiefly for efficiency, because Ho says OpenAI expects to be limited by data center power before it runs out of compute demand. The company built the chip in-house instead of buying merchant silicon so engineers could trade off software, model, and hardware decisions with full visibility into OpenAI's own research, something Ho said a third-party vendor could not be trusted with. To counter claims the chip is tuned only for OpenAI's own models, Ho pointed to Jalapeno's Hot Chips results on SemiAnalysis's InferenceX benchmark, which uses a mix of open-source models, and said his team got those models running well within about two months of getting first silicon back. Ho said OpenAI could eventually sell or license the chip externally, but internal demand is growing so fast the company expects to consume most of its own output for a long time.
The bigger claim here is about pace, not the chip itself. Ho says AI-assisted design tools, OpenAI's own Codex and Sol models, with a newer one called Astra now coming online, took Jalapeno from a blank slate to tape-out in roughly nine months, against an industry norm of 18 months to two years even when a team reuses existing IP. If that holds up on OpenAI's next chip, it suggests AI-assisted design could compress silicon timelines industry-wide, which matters more to chip economics than any single chip's spec sheet.
Worth noting the messenger: OpenAI is also the company selling the AI models it credits for the speedup, so nine months is as much a pitch for Codex as it is a hardware milestone.