OpenAI Unveils Jalapeno AI Chip With TSMC 3nm Process, Targets 50% Cost Reduction

OpenAI and Broadcom unveiled Jalapeno, a custom AI inference processor designed with AI model participation, employing TSMC's 3-nanometer technology. According to Broadcom CEO Chen Fu-Yang, Jalapeno's compute performance matches Nvidia's Blackwell and Google's TPU, with early tests showing significantly better power efficiency than current market alternatives and potential 50% reduction in large model inference costs.

The chip's 9-month development cycle—half the industry's typical 18-24 month timeline—marks a record driven by AI engineers' participation in circuit design and architecture optimization. Jalapeno is slated for batch deployment by end-2026, targeting 10GW-scale deployment capacity.