OpenAI used its Hot Chips conference appearance to release the first round of test data for Jalapeño, its self-designed inference chip. Co-developed with Broadcom, the chip is rated at 700 watts, with measured sustained power kept under 550 watts. The comparison targets are Nvidia's GB200 NVL72 and GB300 NVL72 systems, whose single accelerators are rated at 1,200 watts and 1,400 watts respectively.
The benchmarks ran on SemiAnalysis's public InferenceX suite, using three models: GPT-OSS 120B, DeepSeek R1 670B, and Moonshot AI's Kimi K2.5 (1 trillion parameters). OpenAI's reported results: at peak throughput, AI work delivered per kilowatt is 1.5 to 1.9 times that of the comparison systems; end-to-end latency is 1.7 to 3.6 times lower; and performance in high-interactivity scenarios is 2.1 to 4.1 times higher. At the fastest inter-token latency tier GB300 has previously achieved, OpenAI's reported per-kilowatt throughput gap widens to 8.6 to 104.3 times.
A single chip is specced at 13.4 PFLOPS (MXFP4 precision), 216GB of HBM4, and 15.4TB/s of memory bandwidth. Assembled into a rack of 128 chips, that totals 1.7 EFLOPS, 27.5TB of HBM4, and close to 2PB/s of bandwidth.
Richard Ho, OpenAI's head of hardware, put the takeaway plainly at the event:
"The bottom line is that the results show a very, very significant performance advance over state of the art."
How the 700-watt line got drawn
Jalapeño doesn't train models — it only runs inference. That single constraint reshapes the entire chip budget. Training is bottlenecked by gradient synchronization and cross-node interconnect bandwidth; inference is bottlenecked by memory capacity, KV-cache hit rates, and per-request response time. The pairing of 216GB of HBM4 with 13.4 PFLOPS starts to make sense in that light — compute density isn't pushed aggressively, but memory and bandwidth are generously provisioned.
OpenAI's own design rationale is that "we designed Jalapeño to minimize data movement and communication delays," with specific measures including explicit KV-cache placement and localization aimed at removing bottlenecks in the prefill and communication stages.
Do some rough math: within the same 1,400-watt rack power budget, you can fit two Jalapeño chips or one GB300. If the claimed 1.5-to-1.9-times throughput-per-kilowatt advantage holds up, the output gap at equal power would widen further still. For a company spending most of its compute on inference, what's saved isn't just electricity — it's also the capacity that would otherwise be lost to chip shortages.
The development process itself is worth noting: from concept to tape-out took nine months, with AI assistance used throughout the design process.
Three benchmark models, two from China
The choice of comparison models is telling. GPT-OSS 120B is OpenAI's own open-weight model, while DeepSeek R1 670B and Kimi K2.5 come from DeepSeek and Moonshot AI respectively. A U.S. company's self-developed accelerator, benchmarked against open weights from two Chinese labs.
The logic is straightforward: closed-source model inference can't be independently reproduced by third parties, so any hardware vendor that wants its benchmarks to be trusted has to pick weights that anyone can download and run themselves. Open models have thus slid from the position of "catching up" into the role of the public yardstick for hardware benchmarking. Chinese labs opened up their weights over the past two years to build ecosystem reach — and that bet is now paying off in places nobody quite expected.
A few things worth discounting
The Register's caveat is that these numbers should be taken with a grain of salt: the comparison excluded optimizations like speculative decoding, and the test methodology isn't fully comprehensive. By comparison, AMD's Helios system has 15% less memory bandwidth than Jalapeño but 1.46 to 2 times more compute; Nvidia and AMD GPUs also retain a programmability edge — a dedicated inference chip trades away flexibility in exchange for efficiency.
A bigger caveat: the comparison doesn't include Vera Rubin, which has just begun shipping. Jalapeño is being measured against the previous-generation Blackwell, while Vera Rubin itself won't reach "very small batch" deployment at OpenAI's own facilities until the end of 2026, with scaled deployment landing in 2027. By the time it's actually running, the chip on the other side of the comparison won't be GB300 anymore.
OpenAI isn't slowing down on cadence either: a second generation is already deep in development, and a third generation is in the design phase.
Sources: OpenAI official announcement, TechCrunch, CocoLoop, The Register, Tom's Hardware; InferenceX benchmark methodology, single-chip specifications, and power figures were cross-checked against public reporting.