Research2026-08-25
An independent benchmark by research firm SemiAnalysis found OpenAI's custom inference chip, Jalapeno, delivered up to 1.9 times more work per watt than Nvidia's Blackwell systems, the first public outside test of an AI lab's own chip against Nvidia hardware. Against Nvidia's newer Vera Rubin platform the two were described as roughly tied on cost per token. Volume deployment only begins at the end of 2026, and OpenAI has not said it will lower prices.
What changed
No AI lab's own inference chip had been independently benchmarked against Nvidia's production hardware in public.
What it unlocks
A concrete outside reference point for how far custom inference silicon has closed the gap with Nvidia.
- 1.9x more work per watt vs GB200/GB300
- 700W vs 1,400W per chip
- $96.2B Nvidia quarterly revenue, +106%
- 12 GW Nvidia compute through 2030
Sources