OpenAI announced Jalapeño today. Custom inference chip with Broadcom, better perf/watt than anything current. Nine month tape-out, fastest ASIC cycle ever. Impressive engineering by any measure. The stock market yawned because the real story is what this chip does not solve.
It is an inference chip. The overinvestment story is about training. Hundreds of billions in H100s and B200s that will be obsolete in three years because the next architecture always makes the last one irrelevant. Jalapeño does not extend the life of those clusters. It makes serving cheaper, which is a cost saving on the operating side. The capex is already spent.
Then there is the Jevons paradox. Make inference cheaper and you run more of it. OpenAI is talking about gigawatt scale data centers. That is not efficiency. That is building bigger. Every improvement in cost per token funds another training run, another model generation, another data center. Total compute consumption goes up. The efficiency gains do not shrink the industry. They accelerate it.
The chip lands at the end of 2026 at earliest. By then Nvidia will have shipped Rubin and whatever comes after. The gap between what hyperscalers spend and what they earn from AI products will have grown wider. Jalapeño is one data point in a race that is defined by the distance from capex to revenue. That distance is still measured in years and trillions.
The bubble is not about the technology. It is about the gap between what this infrastructure costs and what people will pay for the output. A better chip closes the gap on the cost side but opens it on the consumption side. The crash comes when the market realises the gap is structural.
Paul