Jalapeño Is Writing Token Prices Into Silicon

你刚刷到这条消息，本来准备顺手划走，但又怕自己错过了真正会影响下一步判断的那一点。

最容易做错的，是OpenAI and Broadcom unveil LLM-optimized inference chip；代价往往是如果只盯表面热闹，你很容易在错误方向上花掉时间、预算和注意力。；我先给一个保守判断：推理成本一降，敢烧token的产品会先爆发。

The easy mistake is to file "OpenAI and Broadcom unveil LLM-optimized inference chip" under semiconductor news. My conservative read is different: Jalapeño is writing token prices into silicon, and when inference gets cheaper, the first breakout products will be the ones that can afford to burn more tokens.

Why I think that: OpenAI did not frame the payoff as "we built a chip." It framed the payoff as faster ChatGPT, more steps for Codex, and cheaper API products [S001]. That is a product-experience signal, not infrastructure optics.

The other useful tell is workload shape. OpenAI says engineering samples already ran AI workloads including GPT-5.3-Codex-Spark at production target frequency and power, and that the chip taped out in 9 months with OpenAI models helping the design cycle [S001]. Axios adds the practical boundary: OpenAI aims to start routing customer queries to these chips this year [S002].

So I would not read this as "API prices drop tomorrow." This read is based on the June 24, 2026 OpenAI announcement and the June 24, 2026 Axios follow-up; there is still no public cost curve, only early efficiency claims.

A news update is worth tracking not because it lists more features, but because it changes your next decision. If token-heavy reasoning gets materially cheaper over the next 12 months, which part of your 工作流程（workflow） becomes viable first?

真正该讨论的是：OpenAI and Broadcom unveil LLM-optimized inference chip