OpenAI’s first custom inference chip has posted numbers European infrastructure planners should treat as a live constraint rather than a distant benchmark.
On OpenAI’s own results, with SemiAnalysis spot-checking the InferenceX runs in person in the lab, Jalapeño returned 1.5 to 1.9 times more useful work per watt at peak throughput than the best commercial systems available for comparison, including NVIDIA’s GB200 and GB300 platforms. End to end latency dropped 1.7 to 3.6 times across three models: GPT OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T. For interactive agent workloads the gap widened to 2.1 to 4.1 times. The chip is rated at 700 watts yet held sustained power at or below 550 watts on the tested loads.
What Makes OpenAI’s Jalapeño Different?
The architecture keeps model state, including the KV cache, local and activates compute, memory and networking only for the phase required. Prefill and decode no longer force the classic throughput versus latency compromise.
OpenAI used earlier models to explore the design space and later ones to optimise kernels. Selected AI generated implementations ran 1.5 to 1.8 times faster than expert human code.
Small volume deployment inside OpenAI’s own fleet is scheduled for the end of 2026, with a second generation already in development.
How Does This Land Against Europe’s Capacity Push?
Europe is spending heavily on scale.
The Commission’s AI Gigafactories call seeks to mobilise more than €30 billion for up to seven large compute hubs. Chips Act 2.0 and the Cloud and AI Development Act prioritise capacity, energy efficient data centres and strategic autonomy.
Power availability, not capital, is already the binding limit on many European grids. Efficiency of the scale Jalapeño demonstrates directly reduces the energy cost of every inference token, the precise pressure point European operators face.
See Also:
What the AI Gigafactories Call Actually Funds
What Is Sovereign AI? Definition, the Money Behind It, and Europe’s Reality
