OpenAI's Jalapeño inference chip beats Nvidia Blackwell on throughput and latency — and it's going into production

OpenAI's Jalapeño Chip Enters Production — Outperforming Nvidia on Inference

OpenAI has built an inference chip that beats Nvidia's Blackwell systems on the metrics that matter most for serving AI models at scale. The Jalapeño chip delivers 1.5x to 1.9x more useful inference throughput per watt compared to Nvidia's GB 200 and GB 300 systems, while simultaneously cutting end-to-end latency by 1.7x to 3.6x. The design is purpose-built for inference, not training, and it's already moving into production.

The scale of the bet is enormous. OpenAI has secured a deal with Broadcom for 10 gigawatts of chip capacity deployed from the second half of 2026 through 2029. For context, that's roughly five times the company's current operational compute footprint — all added within roughly three and a half years. The speed of execution is notable; industry wisdom holds that custom chips take years to move from design to production. Jalapeño is on track to begin deployment by mid-2026.

The commercial logic is straightforward. Lower latency and higher throughput per watt means fewer data centers running the same inference workload, which improves gross margins and capital efficiency. It also reduces the narrative risk that OpenAI remains dependent on Nvidia's hardware roadmap.

The fact that Broadcom is manufacturing raises questions about whether Broadcom is the foundry partner, the chipmaker, or both — the transcript doesn't clarify the manufacturing relationship. What is clear is that OpenAI designed the chip and is contracting for production capacity at scale.