OpenAI isn’t building a chip. It’s reclaiming the profit margin on every token from Nvidia. OpenAI just announced the first round of testing for its in-house inference chip, Jalapeño: - AI workload per watt increased by 1.5–1.9x - End-to-end latency reduced by 1.7–3.6x - Internal deployment planned by the end of this year This doesn’t mean Nvidia will be replaced immediately. Jalapeño is currently focused on inference, not training; and its initial deployment scale won’t be anywhere near Nvidia’s global supply chain. But the AI business model is already shifting: In the past, model companies bought compute power from chip companies and sold it to users per token. In the future, leading model companies will control: Models + inference software + chips + networks + racks. Every time they lower the cost per token, they can choose to: - Lower API prices; - Increase free usage limits; - Enable agents to run more steps; - Or turn the savings into profit. Five years from now, the biggest competition in the AI industry might not just be about “whose model is the smartest.” It’ll be about: Who can produce a useful token at the lowest cost.