On August 25, 2026, OpenAI published Jalapeño’s first results. The numbers are OpenAI’s own InferenceX measurements, not an independent lab copy. Jalapeño is OpenAI’s first custom inference chip. The post says it can serve more AI work per unit of power and return responses more quickly, and that the architecture is meant to avoid the usual tradeoff between throughput and latency. None of that is a store listing. Alberta operators cannot buy this silicon.
What OpenAI published on InferenceX
OpenAI says it tested Jalapeño on InferenceX, a public benchmark from SemiAnalysis, across GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 1T. Across those three, it reports 1.5 to 1.9 times more AI work per watt at peak throughput and 1.7 to 3.6 times lower end-to-end latency than the comparison systems. For highly interactive workloads, it reports 2.1 to 4.1 times higher performance. It normalized using each accelerator’s published chip power rating. Jalapeño is rated at 700 watts; OpenAI says measured sustained power stayed at or below 550 watts on the workloads tested.
The appendix is OpenAI’s table, not this desk’s lab. For GPT-OSS 120B it lists package TDP as Jalapeño 700W and GB200 1,200W, higher peak mixed TPS/kW of about 1.9× (85,448 vs 44,960 mixed/kW), and lower end-to-end latency of about 1.7× (1.03s vs 1.80s). For Kimi K2.5 1T it lists GB300 at 1,400W, about 1.5× higher peak mixed TPS/kW, and about 3.4× lower end-to-end latency. DeepSeek R1 is also in that appendix, with GB300 at 1,400W. Those figures are copied from OpenAI’s page. They are not an independent confirmation.
“Jalapeño delivered 1.5 to 1.9 times more AI work per watt at peak throughput and 1.7 to 3.6 times lower end-to-end latency than the comparison systems.”OpenAI, Jalapeño’s first results, August 25, 2026
OpenAI describes the architecture around real language-model phases: prefill is compute-intensive; decode is constrained more by memory bandwidth; KV cache is meant to stay local. It says AI helped move from initial design to tapeout in nine months. Using Codex with GPT-Astra, it says the team brought three open-weight models that were not in the original production plan to high performance within two months. For selected GPT-OSS attention and mixture-of-experts blocks, it says AI-generated implementations ran 1.5 to 1.8 times faster than existing human-expert kernels. OpenAI says those figures apply to the selected blocks, not the full model.
You still cannot buy the chip
OpenAI plans to begin deploying Jalapeño inside its own compute infrastructure by the end of the year. It calls Jalapeño the first generation of a multigenerational roadmap, with Gen 2 deep in development. It also says it will continue to widely deploy accelerators from NVIDIA and other partners for training and inference. That is a first-party silicon story. It is not a channel SKU for an Edmonton shop.
The June 24 unveil post, OpenAI and Broadcom unveil LLM-optimized inference chip, names Broadcom and Celestica as partners on implementation, boards, racks, and production systems. This desk did not fetch a TechCrunch recap, so it is not adding TCO-versus-Rubin claims, Hot Chips slides, or volume figures that are not on OpenAI’s pages.
What Alberta operators should actually watch
Industrial AI Alberta teams cannot buy Jalapeño. If OpenAI later serves API traffic from this stack, latency and price might move. That is still a closed full stack: models, serving software, chips, memory, networking. It does not replace a private on-device path. Client files, tenders, and plant data still belong on a private AI security route, not on a hope that OpenAI’s watt-per-token chart will someday land in your rack.
Keep a second path. If the job has to stay on the desk, it still belongs on a Mac mini or other private stack. If the job is API work, watch OpenAI’s own latency and price after year-end deployment — and do not treat a first-party InferenceX appendix as a purchase order. AI consulting here means mapping which workflows can wait on OpenAI’s infra and which cannot leave the plant.
The guardrail
OpenAI’s numbers are OpenAI’s numbers. InferenceX is named; the comparison systems and TDP lines are in the appendix. That is not an independent audit, and it is not a chip you can order. For private AI security, a faster closed stack is still someone else’s building. Keep NVIDIA in the plan because OpenAI says it will. Keep on-device work on-device because Jalapeño will not show up in receiving.