They made a chip. That's not the interesting part.

Altman posted it the way this industry posts everything now: we made a chip and it is fast. Yesterday at Hot Chips, OpenAI walked through Jalapeño, its first inference ASIC, built with Broadcom. Nine months from design to tape-out. The models helped design the silicon that will run the next models. Codex wrote kernels that beat the human-expert versions on selected attention and MoE blocks.

The numbers they want remembered: 1.5x to 1.9x more work per watt than the comparison systems, 1.7x to 3.6x lower end-to-end latency, across GPT-OSS 120B, DeepSeek R1, and Kimi K2.5. Power-normalized InferenceX, measured in their lab, with SemiAnalysis in the room. It is still a lab result, not a 100-megawatt fleet.

The architecture argument is the real one. Inference is three different machines pretending to be one request: compute-bound prefill, a tiny draft model for speculation, then memory-bound decode with bursty MoE traffic. A heterogeneous GPU fleet pays idle power for the phases it isn't in. Jalapeño keeps the KV cache local and darkens the silicon it isn't using. One chip, three regimes.

Nvidia printed: $96.2B revenue, $89.0B data center, non-GAAP $2.22, Q3 guided at $108B. Huang's line was compute is revenue. OpenAI will keep buying their chips for training. Nobody serious thinks GB300 is finished. What changed is that the biggest buyer of inference just became a vendor, and they did it on a timeline that used to belong to software.

Gen 2 is already in development. Deployment inside OpenAI starts by year end. The rest of us will feel it as cheaper tokens, or we won't. That's the test.