When can I actually use Jalapeño?

Jalapeño is not something developers buy or rent directly yet. OpenAI plans to deploy the chip in very small volumes by the end of 2026 and scale it up meaningfully during 2027. OpenAI describes Jalapeño as the start of a multigenerational platform, and says a second and third generation are already in development.

How does Jalapeño compare to Nvidia Blackwell?

TechCrunch reports that Jalapeño was measured against Nvidia Blackwell systems on the InferenceX benchmark. OpenAI's own post is less specific and refers to "leading commercially available AI systems" as the comparison. SemiAnalysis, whose engineers ran the workloads themselves, published its analysis under the headline that Jalapeño is better than Nvidia Blackwell.

Which models did OpenAI test Jalapeño on?

OpenAI ran Jalapeño on three public open-weight models: GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T. On Kimi K2.5, the largest of the three, Jalapeño delivered roughly 1.5 times higher peak performance per watt and 3.4 times lower end-to-end latency. OpenAI says its advantage widened further on frontier OpenAI models in internal testing.

Did AI really help design the Jalapeño chip?

OpenAI says its own models moved Jalapeño from initial design to tapeout in nine months by exploring implementations and shortening design, measurement and verification loops. AI also helped optimize the chip's arithmetic circuits. OpenAI further designed Jalapeño as a predictable programming target, using local tensors and explicit communication, so models can map and schedule work across the system.

OpenAI · 2026-08-25 · major

Jalapeño first results — OpenAI's inference chip posts 1.9x more work per watt

OpenAI published the first measured results for Jalapeño, its custom inference chip. On SemiAnalysis's InferenceX benchmark it delivered 1.5–1.9x more work per watt and 1.7–3.6x lower end-to-end latency than the comparison systems.

OpenAI card announcing Jalapeño inference chip first results

OpenAI's first custom inference chip now has measured numbers, and they favour tokens per watt over tokens per chip.

Key specs

Work per watt	1.5–1.9x
End to end latency	1.7–3.6x lower

Quick facts

Maker	OpenAI, co-designed with Broadcom
Chip power rating	700 W
Measured sustained power	550 W or below
Benchmark	InferenceX, a public benchmark from SemiAnalysis
Models tested	GPT-OSS 120B, DeepSeek R1 670B, Kimi K2.5 1T
Interactive workloads	2.1–4.1x higher performance
Availability	Small volumes by end of 2026, scaling in 2027

What is it?

The first measured performance numbers for Jalapeño are now public. OpenAI announced the chip earlier and has been testing it, and the system built around it, ever since; the results landed at the Hot Chips conference. Jalapeño is an accelerator built for one job — serving language models, especially interactive agents — and OpenAI ran it on InferenceX, a public benchmark from SemiAnalysis that measures the whole process of serving a request.

How does it work?

Language-model inference moves through phases with different bottlenecks: prefill, which processes the prompt, is compute-heavy, while decode, which generates tokens one at a time, is limited by memory bandwidth. Jalapeño is designed to minimise the data movement between them, so model state and the KV cache can be explicitly placed and kept local while the chip activates the right mix of compute, memory and networking for each phase. A large network domain keeps a whole workload inside one connected system.

Why does it matter?

OpenAI argues that performance per unit of power, not per chip, is the more useful standard — and Jalapeño's numbers are built on that framing, normalised against each accelerator's published power rating. If it holds at scale, OpenAI gets more useful work out of the same megawatts, which matters most for agents that run many steps in a row where every delay compounds. The chip also went from design to tapeout in nine months, with OpenAI's own models helping optimize its arithmetic circuits.

Who is it for?

infrastructure and inference engineers

Frequently asked questions

When can I actually use Jalapeño?: Jalapeño is not something developers buy or rent directly yet. OpenAI plans to deploy the chip in very small volumes by the end of 2026 and scale it up meaningfully during 2027. OpenAI describes Jalapeño as the start of a multigenerational platform, and says a second and third generation are already in development.
How does Jalapeño compare to Nvidia Blackwell?: TechCrunch reports that Jalapeño was measured against Nvidia Blackwell systems on the InferenceX benchmark. OpenAI's own post is less specific and refers to "leading commercially available AI systems" as the comparison. SemiAnalysis, whose engineers ran the workloads themselves, published its analysis under the headline that Jalapeño is better than Nvidia Blackwell.
Which models did OpenAI test Jalapeño on?: OpenAI ran Jalapeño on three public open-weight models: GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T. On Kimi K2.5, the largest of the three, Jalapeño delivered roughly 1.5 times higher peak performance per watt and 3.4 times lower end-to-end latency. OpenAI says its advantage widened further on frontier OpenAI models in internal testing.
Did AI really help design the Jalapeño chip?: OpenAI says its own models moved Jalapeño from initial design to tapeout in nine months by exploring implementations and shortening design, measurement and verification loops. AI also helped optimize the chip's arithmetic circuits. OpenAI further designed Jalapeño as a predictable programming target, using local tensors and explicit communication, so models can map and schedule work across the system.