OpenAI Drops Jalapeño ASIC: Crushing Nvidia Chips In Inference Speed
OpenAI got tired of paying the Nvidia tax. The company revealed custom in-house silicon designed specifically for running AI models, claiming it leaves top-tier accelerators in the dust.
OpenAI revealed the first benchmark figures for its custom ASIC, codenamed Jalapeño, testing it head-to-head against commercial Nvidia GB200 and GB300 setups across the InferenceX benchmark suite.
The in-house silicon delivered a brutal performance beat across three massive open-weight models. On DeepSeek R1 670B, the hardware pushed roughly 700 tokens per second per user versus 169 on Nvidia hardware. On Kimi K2.5, it achieved 694 tokens against 182, while blasting 1459 tokens per second on GPT-OSS 120B compared to 535 on rival gear. Across peak workloads, the chip achieved 1.5 to 1.9 times more compute per watt alongside a 1.7 to 3.6 times reduction in latency.
Rather than crafting a general-purpose processor, engineers built the chip strictly around modern agentic workflows, keeping model context and KV-cache packed as close to the compute units as physics allows. Despite being rated for 700W, actual testing pulled a steady 550W under sustained load.
The deployment pipeline also leaned entirely on automation: engineers used Codex and GPT-Astra to port non-production models in under two months, with AI-generated attention and Mixture-of-Experts kernels running up to 1.8 times faster than handwritten human code. OpenAI plans to push the first production units into data centers before the year ends, with subsequent generations already in development.
The era of single-vendor domination in AI computing is crashing straight into custom hyperscale silicon. When software builders successfully fabricate specialized chips that run circles around general hardware, the balance of power across entire data center supply chains changes forever.
Source: openai.com
Comments
This is where the magic happens: AI reads your discussion and rewrites the article based on the most interesting comments. Each strong comment adds points to the meter below. Once the meter is full, the article updates live — no page reload needed.