OpenAI has unveiled benchmark results for its custom-designed AI inference chip, codenamed Jalapeño. Developed in collaboration with Broadcom and manufactured by TSMC, the ASIC delivered 1.5 to 1.9 times more work per watt and reduced latency by 1.7 to 3.6 times compared to Nvidia's GB300 processors on the InferenceX platform. OpenAI plans a limited deployment by the end of 2026 and a full rollout in 2027, though it will continue to purchase training hardware from partners like Nvidia.
Jalapeño chip benchmark results
- ▪On the Moonshot AI Kimi K2.5 1T model, OpenAI's Jalapeño chip achieved approximately 1.5 times the peak performance per watt and 3.4 times lower latency than the comparison systems
- ▪OpenAI's Jalapeño chip reduced end-to-end latency by 1.7 to 3.6 times across several AI models, including GPT-OSS 120B and DeepSeek R1, compared to Nvidia's GB300 processors
- ▪OpenAI's custom-designed AI chip, codenamed Jalapeño, delivered 1.5 to 1.9 times more AI work per watt compared to Nvidia's GB300 processors during testing on the InferenceX platform
Custom AI chip development strategy
- ▪OpenAI hardware vice president Richard Ho stated that the company has no plans to sell or rent the Jalapeño chip to others because it is struggling to secure enough compute for its own needs
- ▪OpenAI used its own AI models to assist in the chip design process, allowing the company to move from schematic to tape-out in approximately nine months
- ▪OpenAI's Jalapeño chip is an application-specific integrated circuit (ASIC) designed exclusively for AI inference rather than model training
Broadcom manufacturing partnership
- ▪OpenAI designed the Jalapeño chip in collaboration with Broadcom, and the chip is manufactured by TSMC
- ▪The Jalapeño chip fits into an October 2025 agreement between OpenAI and Broadcom to deploy 10 gigawatts of OpenAI-designed accelerators through 2029
Deployment timeline through 2027
- ▪OpenAI plans to deploy the Jalapeño chip in racks of 128, with a full pod running 2,048 ASICs to deliver 1.7 exaflops of 4-bit compute
- ▪OpenAI is already deep into the development of a second-generation version of the Jalapeño chip, with a third-generation version also underway
- ▪OpenAI plans a limited deployment of the Jalapeño chip by the end of 2026, with a full-scale rollout expected to ramp through 2027
AI hardware market competition
- ▪OpenAI will continue to purchase hardware from Nvidia, AMD, and Cerebras to meet its computing needs, particularly for model training
- ▪Broadcom CEO Hock Tan claimed that the Jalapeño chip matches the performance of Nvidia's Blackwell architecture and Google's TPU while offering a 50% cost advantage on a per-token and per-kilowatt basis
Story comments
Loading comments…