Jalapeño’s first results show industry-leading speed and efficiency in AI inference

OpenAI reports testing Jalapeño on InferenceX with GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T. Across the three, it claims 1.5 to 1.9 times more AI work per watt and 1.7 to 3.6 times lower end-to-end latency than comparison systems.

Jalapeño is rated at 700 watts, while measured sustained power was at or below 550 watts in the tested workloads. OpenAI plans to deploy it in its infrastructure by the end of 2026, with second and third generations in development.

The architecture is built around inference’s different phases: compute-heavy prefill, memory-bound decode and communication delays. Keeping model state local and integrating the network into the accelerator is intended to make one system effective across interactive and batched serving. The source gives a plausible mechanism, but not a complete proof of generality. The likelier read is that deployment conditions will determine the real value.

Signal Desk's read: The architecture is built around inference’s different phases: compute-heavy prefill, memory-bound decode and communication delays. Keeping model state local and integrating the network into the accelerator is intended to make one system effective across interactive and batched serving. This is a meaningful direction, but the strongest headline claim should be treated as a hypothesis until controls, broader workloads or outside measurements arrive.

This is a vendor or team-authored post, so its claims are vendor-reported and require independent validation. The practical question is who can reproduce the result, at what cost, and under which constraints. That is where the announcement either becomes infrastructure or remains a persuasive case study.

This is OpenAI’s own engineering report, so the benchmark selection, normalization and performance claims are vendor-reported. The comparison uses published chip power ratings and a public benchmark, but the source does not provide independent replication.

OpenAI says AI helped design and program the chip, including selected attention and mixture-of-experts blocks. Those blocks ran 1.5 to 1.8 times faster in its tests, which is not a full-model result.