SemiAnalysis on Jalapeno: OpenAI's chip, measured in tokens per megawatt
- Source
- x.com
- Date
OpenAI's secret weapon just dropped. Jalapeno, the AI-designed chip that beats Rubin on tokens per megawatt, taped out in nine months, and runs Doom. Myron and Bryan join @JordanNanos on SemiAnalysis Weekly to break it all down: - Jalapeno vs Rubin and GB300 on InferenceX - Why tokens per megawatt is the metric that matters - Samsung's HBM4 comeback - AI-written RTL and kernels - Is the CUDA moat draining? - Anthropic's ASIC is next Watch Now: https://t.co/C1p4oguqlt
Context
OpenAI's Jalapeño chip is being judged on power efficiency. SemiAnalysis uses generated-text throughput per megawatt, a measure of how much output a system produces for a given electrical power draw. Its August 25 report says Jalapeño's results on that measure exceed NVIDIA's Vera Rubin figures published in July.
The report limits what this shows. OpenAI supplied the numbers, and the analysts say they watched the runs in OpenAI's lab without running the full InferenceX suite or seeing results on AgentX, which uses long multi-turn workloads that stress components such as request routing and prefix caching. The comparison covers shorter single-turn tests, so it is an attributed efficiency result on those workloads, not evidence about sustained work.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
- token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
Checking sign-in…
Loading comments…