OpenAI's Answer to Nvidia is Jalapeño. First Numbers Released

Custom hardware begins crushing the standard

During the weekend's Last Week in AI podcast, a crucial mention was made of Jalapeño, the first generation of custom AI hardware from OpenAI. According to the released numbers, the new inference chip beats current leading systems in both speed and overall energy consumption per token. Sam Altman's team is thus paving the way to gradually reducing dependence on third-party hardware supplies and plans to deploy Jalapeño within its own data centers by the end of this year.

The end of endless Nvidia tax payments

Jalapeño represents a strategic turning point for cloud providers. While training still relies on heavy machinery and GPU farms, the inference run (when a model actually answers millions of queries) is becoming cheaper and optimized via dedicated chips. OpenAI utilizes deep hardware-software co-design here. If it shifts the bulk of ChatGPT traffic to its own silicon, it will save billions of dollars in margins. For Nvidia, this means a loss of revenue from a client that dictates the market pace.

Cheap operation as a necessity, not a luxury

With the price per token dropping for third and fourth-generation models (including Google's Gemini 3.7 Flash announcement), a hardware emergency brake is a necessity. Whoever offers the cheapest inference pipeline will retain customers running long and iterative agentic workflows. The lingering question remains whether the first generation can scale without teething problems in cooling or networking interconnects.

When will we see the impact on pricing?

The proof that Jalapeño actually works under full load won't just be press releases. It will be another deep cut in API pricing at the end of 2026. Once OpenAI cuts the output price per million tokens across the board below the level that standard H100 or B200 infrastructure can generate with a reasonable margin, Jalapeño has won.

Lilith's verdict

It's not the end of the world for Nvidia, but it is a wake-up call. Whoever pays the bill for tokens will eventually build their own pub.