OpenAI published the first measured performance results for Jalapeno, the inference accelerator it is building with Broadcom, on August 25. Against the comparison systems it selected, OpenAI reports 1.5 to 1.9 times more AI work per watt at peak throughput, 1.7 to 3.6 times lower end-to-end latency, and 2.1 to 4.1 times higher performance on highly interactive workloads. The tests ran on three models: GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 1T.
Key facts
- 1.5-1.9x more AI work per watt at peak throughput versus the comparison accelerators, per OpenAI's own measurements.
- The chip is rated at 700 W, with measured sustained power at or below 550 W on the tested workloads; results were normalized using each accelerator's published chip power rating.
- On Kimi K2.5 1T specifically, OpenAI cites roughly 1.5x higher peak performance per watt and 3.4x lower end-to-end latency.
- Primary source: OpenAI, "Jalapeno's first results show industry-leading speed and efficiency in AI inference", August 25, 2026.
The benchmark table is the least interesting part of this announcement.
What OpenAI is actually arguing is that Jalapeno is fast because of a software decision, not a silicon one. The company frames it as full-stack co-design: it "can design models, products, serving software, chips, memory, networking, and systems together," and the result delivers higher throughput and lower latency from one architecture rather than trading one against the other.
The concrete version of that claim appeared four weeks earlier, in a companion post on GPT-5.6. There, OpenAI says GPT-5.6 Sol, working through Codex, "autonomously rewrote and optimized our production kernels." A kernel is the small, brutally hand-tuned piece of code that actually executes one operation -- a matrix multiply, an attention step -- on a specific chip. Writing good ones is among the most specialized work in computing, and the accumulated stock of them is a large part of why one chip vendor's ecosystem is hard to leave.
OpenAI says it trained GPT-5.6 specifically to write and improve kernels in Triton and Gluon, which it describes as "two open-source GPU programming languages maintained by OpenAI," and that this work together with broader kernel advances "reduced end-to-end serving costs by 20%." It also says it built verification tooling for the effort, including an open-source floating-point sanitizer called FpSan.
The mechanism worth understanding is the ordering. OpenAI did not build a chip and then ask a model to program it. It narrowed the programming model until kernels became something a model could reliably synthesize and tune, then co-designed memory movement, synchronization, and data layout around that narrower surface. The mathematical groundwork for that surface is public: arXiv:2505.23819, "Linear Layouts: Robust Code Generation of Efficient Tensor Computation Using F2" by Keren Zhou and colleagues, describes exactly the Triton-integrated layout algebra that makes machine-generated kernel code tractable rather than a combinatorial nightmare.
Analogy: general-purpose GPU programming is like writing prose, where a good writer beats a machine because the space of good sentences is enormous and unstructured. A constrained layout algebra turns it into something closer to filling in a form. Machines are excellent at forms. If that reframing holds, the durable competitive asset in accelerators stops being the silicon and starts being whoever can most cheaply generate the software layer -- and that is the part being automated. Ground Truth covered the first half of this story when OpenAI said Sol rewrote the kernels that run Sol.
One of the tested models is worth grounding, since anyone can check it. GPT-OSS 120B is open weights: the standard checkpoint set is about a 65 GB download, and OpenAI's own model card states that its MXFP4 quantization makes the model "run on a single 80GB GPU (like NVIDIA H100 or AMD MI300X)." The other two, DeepSeek R1 and Kimi K2.5, are the kind of large open-weight models that make inference efficiency a commercially interesting problem in the first place.
Now the caveats, and they are substantial.
These are vendor benchmarks, on vendor-selected workloads, against vendor-selected comparison systems. The power normalization is the part to look at hardest: OpenAI says results were normalized using each accelerator's published chip power rating, while separately noting that Jalapeno's measured sustained draw was at or below 550 W against a 700 W rating. Normalizing by rated rather than measured power flatters a part that runs well under its own ceiling.
"Highly interactive workloads," where the largest multiples appear, is also doing significant work. Low-concurrency, short-context serving is precisely the regime where a narrow, specialized part looks best and where a general-purpose accelerator is least optimized. It is a real workload -- it is what a chat interface feels like -- but it is not the regime where most tokens are served.
And no independent verification has been confirmed. Analysis attributed to third-party testing has circulated, but none of it could be retrieved against a primary source, so specific figures from it should be treated as unconfirmed. Nor is this silicon anyone can buy. OpenAI's own language is that "in the months ahead, we will ramp Jalapeno," and that it is "the beginning of a multigenerational platform" -- notably softer than the end-of-year deployment timeline that has appeared in secondhand coverage. Jalapeno also appears designed to serve OpenAI, not to be sold, which makes "beats the competition" a claim about OpenAI's cost structure rather than about anyone else's purchase options.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)