30 August 2026

OpenAI uses its own models to design custom chip

• OpenAI built a chip called Jalapeno with assistance from its own AI models, which wrote code and helped reduce design costs by roughly 10%.
• In 16 months, Jalapeno achieved 1.5-1.9x better efficiency at generating tokens (text output) per unit of power compared to Nvidia's comparable chips.
• Exponential View argues this shows AI can speed up hardware design cycles and may lead to more specialized chips from different makers rather than one dominant standard.

How it was covered

Exponential ViewAzeem Azhar

OpenAI designed a new chip called Jalapeno with help from its own models, which wrote kernels and reduced compute costs by roughly 10%, achieving 1.5-1.9x better tokens per megawatt than comparable Nvidia silicon in 16 months. The newsletter argues this demonstrates AI's ability to compress chip design cycles and will create a highly segmented, heterogeneous compute market with room for specialist firms.