OpenAI Jalapeño: Better Than Nvidia Blackwell
OpenAI’s first disclosed inference ASIC, built with Broadcom, is a serious hardware/software co-design effort—but the headline performance advantage is still a vendor-involved benchmark claim, not an independent production result.
SemiAnalysis reports a generalized LLM-inference chip designed from scratch in roughly 16 months, using HBM4 and targeting throughput per megawatt across several open models. Its charts put Jalapeño ahead of Nvidia Blackwell on the tested configurations, without speculative decoding; the report says adding speculative decoding could improve Jalapeño further. OpenAI even demonstrated Doom running on the chip via Codex-generated porting work.
The Hacker News discussion challenged the comparison’s missing ISA details, the choice to omit speculative decoding on Jalapeño while including it for competitors, and the gap between lab benchmarks and months of real production traffic. Other comments debated whether custom silicon is a durable moat or mainly leverage against Nvidia, and whether weights can ever be baked into hardware quickly enough for rapidly changing frontier models.