OpenAI built its own chip, and now there is real data to back up why it bothered. New independent benchmarks published this week put OpenAI’s Jalapeno inference chip, made together with Broadcom, ahead of Nvidia’s best Blackwell systems on the exact workload every major lab depends on daily: serving AI models to millions of people at once, fast and without melting the power bill.
The numbers are honestly kind of wild. Jalapeno pulled in up to 1.9 times more throughput per kilowatt and cut latency by as much as 3.6 times, all while drawing noticeably less power than Nvidia’s top rack. It is the kind of gap that makes a multi-year hardware bet start to look worth it.
To be fair, the comparison has some wrinkles. Jalapeno uses newer memory than the Blackwell chips it was tested against, so it is not a perfectly clean fight, and it is still headed for low volume production later this year, not a full replacement for Nvidia hardware anytime soon.
Still, the ripple goes beyond OpenAI’s own data centers. When one of Nvidia’s biggest customers proves it can build something competitive, every other lab watching gets a little more leverage in its own next negotiation. Nvidia is not going anywhere, its manufacturing scale and software ecosystem are still enormous. But the idea that AI compute has to run on one company’s chips just got a lot shakier.
For the rest of us, cheaper and more efficient inference eventually shows up as cheaper AI tools, faster responses, and more compute to go around instead of everyone fighting over the same scarce GPUs. Competition at the hardware level is good news even if you never think about chips at all.
——
Follow: @Ali Demi
Book your free AI clarity call, NOW!
https://buff.ly/TpWy277
——
Sources:
https://www.tomshardware.com/tech-industry/semiconductors/openai-says-its-jalapeno-chip-beats-nvidias-gb300-in-first-published-benchmarks
https://newsletter.semianalysis.com/p/openai-jalapeno-better-than-nvidia
https://www.cnbc.com/2026/08/26/openai-jalapeno-ai-chip-nvidia.html
Repost this. Thanks.