OpenAI has finally shown its hand on custom silicon, and the numbers are more interesting than the hardware. The company published the first measured performance results from Jalapeño, its first custom inference chip, claiming it delivered more peak throughput per kilowatt and lower token latency than the commercial systems in the comparison on a public benchmark called InferenceX running GPT-OSS 120B. The chip also performed well on DeepSeek R1 and Kimi K2, which matters because it suggests the efficiency gains aren’t just tuned for OpenAI’s own models.
But the chip is only one piece of a broader argument OpenAI is making about how it thinks about compute. The company describes an integrated system where data centers, chips, models, products, and devices reinforce each other. Better software makes hardware more productive. Better models unlock better products, which generate more usage and learning. Those signals feed back into the system. It’s a flywheel pitch, and OpenAI is now backing it with a growing list of partners that includes Microsoft, NVIDIA, AWS, AMD, Broadcom, Cerebras, CoreWeave, Oracle, SB Energy, and SoftBank. The strategy is deliberately portfolio-based: premium systems where capability matters most, efficiency-focused hardware where scale and cost dominate.
The most concrete data point in the announcement isn’t about silicon at all. On the Artificial Analysis Coding Agent Index, GPT-5.6 Sol with max reasoning hit a new high while using 54% fewer output tokens than another leading model. That’s the kind of number that makes CFOs pay attention, because token reduction translates directly into lower serving costs and faster responses. OpenAI is framing this as Jevons paradox in action: make intelligence cheaper and more work becomes economically practical, which expands consumption rather than shrinking it.
What’s striking is how much of this reads as a response to the broader market pressure on inference economics. DeepSeek’s efficiency gains last year forced every major lab to talk about cost per token, not just benchmark scores. OpenAI is now making the same argument with hardware, co-designing chips, serving software, and models as one system. The claim is credible enough on paper: first-party silicon with measured results and future generations already underway. Whether Jalapeño can hold up against the next wave of custom accelerators from Google, Amazon, and Meta is the question that will determine whether this is a genuine moat or just a well-timed press release.
