Jalapeño's first results show industry-leading speed and efficiency in AI inference is an engineering post published by OpenAI on August 25, 2026 (Source: openai.com). It is the first publication of measured benchmark figures for Jalapeño, the custom inference accelerator OpenAI developed with Broadcom and unveiled on June 24, 2026, and it accompanied the company's presentation at the Hot Chips conference the same day.
The post is a vendor self-report of results on a public third-party benchmark. OpenAI ran the measurements itself, and no independent reproduction accompanied publication.
Measurement basis
The results are reported on InferenceX, a public benchmark from SemiAnalysis that measures the full process of serving an inference request. OpenAI states its evaluation standard explicitly: performance is assessed at a matched user experience, measuring how much useful work each system completes per unit of power while meeting a latency target, and that although performance is sometimes reported per chip, "we believe the more useful standard is performance per unit of power."
Every headline ratio in the post is therefore normalized by each accelerator's published chip power rating, not stated per chip. The appendix gives the ratings used: Jalapeño 700 W, Nvidia GB200 1,200 W for the GPT-OSS test, and Nvidia GB300 1,400 W for the DeepSeek R1 and Kimi K2.5 tests. OpenAI adds that Jalapeño's measured sustained power stayed at or below 550 W on the workloads tested, below its 700 W rating.
Three public models were tested — GPT-OSS 120B, DeepSeek R1 670B (MXFP4), and Kimi K2.5 1T (MXFP4) — all at a nominal 8k input / 1k output configuration. Comparison systems are named only in the appendix, per model.
Headline results
Across all three models OpenAI reports Jalapeño delivering 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, and, for what it calls highly interactive workloads, 2.1 to 4.1 times higher performance. It states that Jalapeño sits on the Pareto frontier of performance per watt against latency across the tested operating range, where existing hardware systems often trade throughput against latency.
The appendix gives the per-model figures:
| Model (comparison part) | Peak mixed TPS per kW | End-to-end latency | Minimum time-between-tokens | Throughput at the previous-best TBT |
|---|---|---|---|---|
| GPT-OSS 120B (GB200, 1,200 W) | ≈1.9× — 85,448 vs. 44,960 | ≈1.7× lower — 1.03 s vs. 1.80 s | ≈2.7× lower — 0.69 vs. 1.87 ms (1,459 vs. 535 tok/s/user) | ≈53.7× — 22,935 vs. 427 mixed/kW at 535.28 tok/s/user |
| DeepSeek R1 670B MXFP4 (GB300, 1,400 W) | ≈1.7× — 19,641 vs. 11,781 | ≈3.6× lower — 1.65 s vs. 5.99 s | ≈4.1× lower — 1.43 vs. 5.90 ms (700 vs. 169 tok/s/user) | ≈104.3× — 12,258 vs. 118 mixed/kW at 169.41 tok/s/user |
| Kimi K2.5 1T MXFP4 (GB300, 1,400 W) | ≈1.5× — 18,195 vs. 11,862 | ≈3.4× lower — 1.56 s vs. 5.31 s | ≈3.8× lower — 1.44 vs. 5.48 ms (694 vs. 182 tok/s/user) | ≈56.1× — 6,744 vs. 120 mixed/kW at 182.46 tok/s/user |
The very large final-column ratios are throughput measured at the operating point where the comparison system reached its own minimum time-between-tokens; they are not general throughput comparisons.
OpenAI singles out Kimi K2.5, the largest public model tested, as delivering approximately 1.5 times higher peak performance per watt and 3.4 times lower end-to-end latency than the comparison system, and states that in internal testing Jalapeño's advantage widened further on frontier OpenAI models — a claim made without figures and not reproducible from the post.
Architecture as described
OpenAI describes the design question as what hardware it would build if the primary job were serving modern and future language models, particularly interactive agents. The stated approach is co-design of chip, memory, network, software, and rack-scale system around language-model workloads, aimed at the phase structure of inference: prefill is compute-intensive, decode is constrained by memory bandwidth, and communication between cores and chips adds latency that can leave processing units idle.
The design response given is to minimize data movement and communication delay. Model state, including the KV cache used during generation, can be explicitly placed and kept local, while the system activates the combination of compute, memory, and networking appropriate to each inference phase. OpenAI describes the network as integral rather than peripheral: its large domain keeps an entire workload inside one connected system. The company characterizes the result as a balanced and fungible accelerator able to handle both prefill and decode and to adapt as the balance between them shifts, which it identifies as a defining feature of agentic workloads.
AI in the design and programming loop
The post makes two claims about AI's role in Jalapeño's own development. On design, OpenAI states that AI enabled the team to move from initial design to tapeout in nine months by exploring implementations and shortening design, measurement and verification loops, and that AI also helped optimize the chip's arithmetic circuits.
On programming, OpenAI states that Jalapeño was designed as a predictable programming target — work described through local tensors, explicit communication, and predictable synchronization — so that AI can optimize how that work is mapped, placed, scheduled and coordinated. Using Codex with GPT-Astra, the team is reported to have brought three open-weight models that were not in Jalapeño's original production plan to high performance within two months. For selected GPT-OSS attention and mixture-of-experts blocks, AI-generated implementations are reported to have run 1.5 to 1.8 times faster than existing implementations written by human experts. The post states the scope of that figure itself: it applies to the selected blocks, not to a full model.
Supporting each new model family is described as still requiring new kernels and model-specific optimizations.
Deployment and roadmap
OpenAI states that it plans to begin deploying Jalapeño within its own compute infrastructure by the end of 2026, and describes the chip as the first generation of a multigenerational platform, with Gen 2 "deep in development" and Gen 3 "taking shape." As of publication the company reported continuing production qualification, maturing the software, preparing to operate the chip at scale, and validating performance across more models.
The post states that OpenAI will continue to deploy accelerators from Nvidia and other partners widely, for both training and inference.
The stated economic rationale is operating leverage: producing more useful work from the same power and hardware lowers the cost of delivering a successful result, which OpenAI frames as allowing useful work and revenue to grow faster than the cost to serve. It lists three capability shifts it expects — ultra-fast-mode inference at efficiencies previously available only in fast mode, fast-mode inference at efficiencies previously available only in batched mode, and higher efficiency in batched mode.
Dating
The post itself carries no announcement date for Jalapeño, opening only with "since announcing Jalapeño." The vault's dated record of the unveiling is June 24, 2026, from the joint OpenAI–Broadcom announcement (Source: openai.com).
Provenance
Published at openai.com on August 25, 2026 under OpenAI's Engineering and Company categories. The captured text is complete; the post's charts are represented by their headings and axis captions, so no chart value is recoverable beyond the numeric tables reproduced in the appendix. Figures were corroborated independently by The Verge and TechCrunch of the same date (Source: theverge.com; techcrunch.com).
Relationships
- supports: OpenAI, Broadcom — the primary record for the chip's first published performance figures.
- instance-of: Inference Economics and Token Pricing — a cost-and-throughput-per-watt claim about serving, not training.
- related: Nvidia & TSMC — AI Compute Infrastructure — the comparison systems are Nvidia GB200 and GB300 parts.
- related: gpt-oss (OpenAI open-weight models), DeepSeek-R1, Kimi K2, Astra — the models tested, and the model used to generate kernels.
- related: Semiconductor Supply Chain, Compute Governance — custom inference silicon as a route around general-purpose accelerator supply.
- related: AI for Science — the post's claim that AI shortened the chip's own design cycle.