OpenAI named its first custom chip after a pepper, and judging by the benchmark numbers it just published, that's less a cute branding choice and more a warning label for Nvidia. The company that spends more on Nvidia GPUs than most nations spend on infrastructure just said, politely, "we built our own."
Meet Jalapeño
Jalapeño is OpenAI's first custom inference chip — it doesn't train models, it runs them, functioning as a general-purpose LLM accelerator rather than something tuned narrowly to OpenAI's own stack. Built with Broadcom, design work started in mid-2024 and the fabrication blueprint was finalized in November 2025, with OpenAI saying it used its own AI models to speed up the chip-design process itself.
Per benchmarks verified by SemiAnalysis, OpenAI claims Jalapeño delivers 1.5x to 1.9x more AI work per watt at peak throughput than the best commercially available systems, with 1.7x to 3.6x lower latency and up to 4.1x higher performance on interactive workloads. On GPT-OSS 120B, it reportedly hit around 1,400 tokens per second.
Impressive, With an Asterisk the Size of Texas
The chip hasn't been tested against newer models like DeepSeek V4 Pro or Kimi K3 that Nvidia and AMD have already benchmarked, and Jalapeño is still sitting at the engineering-sample stage while Nvidia's Vera Rubin platform is already shipping to real customers. Some of the comparison numbers also didn't account for optimization tricks like multi-token prediction, which Jalapeño hasn't adopted yet.
The bigger signal isn't the specific multiplier — it's that OpenAI, Google, Amazon, and Meta are all now racing to own silicon instead of just renting it, because at OpenAI's scale, a few percentage points of efficiency per chip compounds into billions of dollars. Nvidia's moat was never just performance; it was everyone needing Nvidia. That's the part actually under threat here.
Jalapeño won't be shipping commercial workloads anytime soon, but the message to Nvidia landed exactly on schedule for earnings season.
Chasing the latest chip news is fun, but most businesses win more by making sure their AI tools are actually integrated into how their teams work — if that's the gap you're staring at, let's fix it.
Source: The Decoder