OpenAI built its own chip — and it beats Blackwell on power

OpenAI built a custom AI inference chip in 16 months, and early tests show it outpaces Nvidia's Blackwell on power efficiency.

At the Hot Chips conference, OpenAI detailed "Jalapeño," a custom chip designed from a blank slate with Broadcom. SemiAnalysis visited OpenAI's lab to verify benchmark runs, reporting that the hardware hit over 700 tokens per second per user on DeepSeek R1 and 1,400 on GPT-OSS. It beat Nvidia's Blackwell on tokens per watt across almost all test scenarios—and even ran Doom using prompts from Codex.

Why it matters: Today's data centers are constrained by power, not cash. Maximizing token output per megawatt directly lowers serving costs and lets OpenAI run fast reasoning models at scale without relying entirely on Nvidia's margins.

Know this: These tests used engineering samples provided by OpenAI, not full production silicon. Additionally, Nvidia's next-gen Rubin chips—which also use high-bandwidth HBM4 memory—are already starting to ship to customers, offering a tougher benchmark for final production hardware.

First-generation custom chips usually flop, but AI-assisted design timelines are clearly changing the math.