- Jalapeño led Nvidia GB300 in response speed and AI work per watt during public tests.
- The 700-watt processor targets AI inference rather than model training.
- OpenAI plans deployment this year while developing generations two and three.
OpenAI plans to deploy the OpenAI Jalapeño chip across its AI infrastructure by the end of 2026. In public tests, the inference processor led Nvidia’s GB300 in work per watt and response speed. Richard Ho, OpenAI’s chip chief, said the design combines high throughput with low latency. The results apply to inference, not model training.
OpenAI Jalapeño Chip Leads GB300 in Two Inference Tests
The OpenAI Jalapeño chip operates at 700 watts. That power level can reduce data-center electricity, cooling, and distribution costs. Ho said the processor can serve more users cheaply while returning faster responses for latency-sensitive workloads.
OpenAI tested the hardware with a smaller open-source model and third-party models from DeepSeek and Moonshot AI. Moonshot’s Kimi, the largest public workload tested, showed the widest gains. Internal tests also covered unreleased OpenAI models, according to the company.
Still, Jalapeño did not face Nvidia’s new Vera Rubin platform, which has begun shipping. SemiAnalysis called the GB300 comparison incomplete since Jalapeño uses HBM4 memory. Nvidia’s Rubin also uses HBM4, making it a closer technical comparison.
OpenAI will decide which models run on the processor. Customers could then select options aimed at lower costs or faster performance. Ho said the approach should improve response times and service availability as model demand grows over time.
Broadcom Partnership Pushes Custom AI Inference Chips Forward
The OpenAI Jalapeño chip was developed with Broadcom, a supplier of custom silicon for large technology companies. OpenAI said its AI models helped accelerate the design process. A second generation is nearing tape-out, while engineers have started concepts for a third.
Ho said Jalapeño can handle larger models than some specialized low-latency systems. Still, OpenAI expects to keep using Cerebras and other compute providers. Demand across its services remains too large for one chip program.
The OpenAI Jalapeño chip could lower inference unit costs, but it cannot train frontier models. Nvidia stays important for training, broad programmability, and CUDA support. Ho described Nvidia as a key partner and said OpenAI will continue buying its hardware.
Disclaimer: This article is for informational purposes only and does not constitute financial advice. CoinCryptoNewz is not responsible for any losses incurred. Readers should do their own research before making financial decisions.
