OpenAI Jalapeño: Better than Nvidia Blackwell
Points and comments are a snapshot, not live.
OpenAI's Jalapeño inference chip beats Nvidia Blackwell and Rubin on perf/W.
OpenAI announced Jalapeño, an ASIC for LLM inference built with Broadcom in ~16 months. It uses HBM4 and achieves industry-leading tokens per watt on models like DeepSeek R1 and Kimi K2.5, surpassing Nvidia Blackwell and Rubin on throughput per MW without speculative decoding. Per-TCO it matches Rubin. B0 stepping delivers 13.4 PFLOPs MXFP4 at 700W TDP. Results are on 8k1k benchmarks; AgentX runs are forthcoming.
Critics note comparison to Blackwell may be unfair-Rubin uses HBM4 and is Jalapeño's true peer-but OpenAI's focus on power efficiency in datacenter-constrained environments drives design.
What commenters are saying
Commenters split: some see ASIC specialization as natural (Bitcoin ASICs, video decode) and expect token prices to plummet; others invoke Jevons paradox-cheaper tokens drive demand, not lower total spend. Skeptics question whether OpenAI will pass efficiency gains to customers or capture them. A counterpoint: older hardware valuations may crash but token prices may rise if supply is constrained by power or by test-time scaling in thinking models.