Wes Roth · 2026-08-26 · notable
Wes Roth — 'OpenAI BROKE the Industry Overnight' on the Jalapeño report
Wes Roth's August 26 episode lists one source in its description: the SemiAnalysis report on OpenAI's Jalapeño inference chip, which argues the chip beats Nvidia Blackwell on throughput per megawatt.
Wes Roth's August 26 episode points at the SemiAnalysis report that puts OpenAI's first custom inference chip ahead of Nvidia Blackwell.
What is it?
'OpenAI BROKE the Industry Overnight....' went up on the Wes Roth channel on August 26, 2026, and its description names a single source: the SemiAnalysis report 'OpenAI Jalapeño: Better Than Nvidia Blackwell', published August 25. Jalapeño is OpenAI's first custom inference chip.
How does it work?
The SemiAnalysis report behind the episode is an efficiency analysis of a first-generation ASIC. It says Jalapeño leads every other chip on throughput per megawatt without the speculative decoding rivals rely on, reaching over 700 tokens per second per user on complex models and roughly 1,400 on some. The B0 stepping adds about 25% performance per watt over the earlier A0 silicon, HBM4 supplies 15.4TB/s of bandwidth per package, and a dual-rack system draws around 160kW.
Why does it matter?
Those numbers are the argument that a lab can build its own inference silicon and beat the incumbent on power, which is what decides serving cost at scale. The report also states its own caveats: the performance data came from OpenAI with little independent verification, and the tests ran at 8k-token context rather than the longer production workloads that stress infrastructure differently.
Who is it for?
people following AI inference hardware