OpenAI debuts Jalapeño inference chip, claims it outperforms Nvidia GB300
OpenAI's first custom silicon, named Jalapeño, posted benchmark results on August 26 that the company says beat Nvidia's GB300 rack systems on throughput and latency.
Key points
- OpenAI has unveiled its first custom AI inference chip, called Jalapeño.
- The company published benchmarks it says show Jalapeño outperforming Nvidia's GB300 in throughput and latency.
- Samsung is reportedly supplying HBM4 memory for the chip.
- The announcement marks OpenAI's first significant move into custom silicon for inference workloads.
Why it mattersCustom inference silicon gives AI companies direct control over performance and cost at scale. If OpenAI's benchmark claims hold up under independent scrutiny, the chip could reduce the company's reliance on Nvidia hardware and reshape procurement decisions across the industry.
OpenAI on August 26, 2026 unveiled its first custom inference chip, named Jalapeño, publishing benchmark results the company says demonstrate superior throughput and latency compared with Nvidia's GB300 rack systems, according to reports from Tom's Hardware and TrendForce.
Tom's Hardware reported that OpenAI's published benchmarks show Jalapeño beating Nvidia's GB300 in first-party performance tests. The publication did not indicate that the results had been independently verified, and the benchmarks as of the reporting date came solely from OpenAI.
On the supply side, TrendForce reported that Samsung is reportedly providing HBM4 memory for the Jalapeño chip. TrendForce noted the arrangement as reported rather than confirmed, and neither source provided additional detail on the volume or terms of any supply agreement.
The debut marks OpenAI's first publicly announced move into custom silicon. The company has until now relied heavily on Nvidia hardware for its inference infrastructure. Developing a proprietary chip represents a significant operational and engineering undertaking, and the announcement positions OpenAI alongside other large technology companies that have pursued custom accelerators for AI workloads.
What happens next remains to be seen. Independent benchmarking of the Jalapeño chip has not yet been reported, and it is not clear from the available sources when or whether the chip will be deployed at scale or made available beyond OpenAI's own infrastructure. Both TrendForce and Tom's Hardware published their reports on August 26, 2026, and the story was still developing as of that date.
Reported by
- TrendForce — News OpenAI Debuts Jalapeño AI Inference Chip, with Samsung Reportedly Supplying HBM4 · 2026-08-26
- Tom's Hardware — OpenAI says its Jalapeño chip beats Nvidia's GB300 in first published benchmarks · 2026-08-26