Hardware
OpenAI Jalapeño Inference Chip: First Measured Results
OpenAI reports Jalapeño delivers 1.5–1.9× more AI work per watt and 1.7–3.6× lower latency vs comparison systems on InferenceX.
OpenAI Jalapeño Inference Chip: First Measured Results
What was released
OpenAI published first measured results for Jalapeño, its custom inference ASIC (with Broadcom), tested on SemiAnalysis InferenceX across GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 1T.
Key specs / results
| Claim | Reported range |
|---|---|
| AI work per watt @ peak throughput | 1.5–1.9× vs comparison systems |
| End-to-end latency | 1.7–3.6× lower |
| Interactive workloads | 2.1–4.1× higher performance |
| Architecture pitch | Higher throughput and lower latency without the usual tradeoff |
OpenAI frames evaluation as useful work per unit power at matched user-experience latency—especially relevant for multi-step agents where delays compound.
Why it matters
First-party silicon with public cross-model numbers strengthens the “full-stack” lab narrative (models + serving + chips) and sets a competitive marker against Nvidia Blackwell/Rubin-class systems.
Related
- openai-jalapeo-better-than-nvidia-blackwell
- openai-and-broadcom-unveil-jalapeo-an-llm-optimized-inferenc
- a-detailed-look-at-jalapeo-openais-asic-developed-with-broad
Source: https://openai.com/index/jalapeno-first-results | 2026-08-25