On August 25, 2026, OpenAI published measured results for Jalapeño, its first custom inference accelerator co-developed with Broadcom. The chip delivers 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 across GPT‑OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T. For highly interactive workloads, Jalapeño achieved 2.1 to 4.1 times higher performance.

OpenAI tested Jalapeño on InferenceX, a public benchmark from SemiAnalysis that measures the full request-serving pipeline. The accelerator is rated at 700 W package TDP but sustained at or below 550 W on the tested workloads. Compared with NVIDIA GB200 (1,200 W) on GPT‑OSS 120B, Jalapeño reached roughly 1.9× higher peak mixed tokens-per-second per kilowatt (85,448 vs. 44,960) and 1.7× lower end-to-end latency (1.03 s vs. 1.80 s). On DeepSeek R1 670B and Kimi K2.5 1T versus GB300 (1,400 W), it posted 1.7× and 1.5× higher peak performance per watt respectively, with latency improvements of 3.6× and 3.4×.

What's new

  • Chip: Jalapeño, OpenAI's first Intelligence Processor, designed for LLM inference
  • Power: 700 W rated TDP; measured sustained ≤550 W
  • Benchmarks: InferenceX (SemiAnalysis) across GPT‑OSS 120B, DeepSeek R1 670B, Kimi K2.5 1T
  • Efficiency gains: 1.5–1.9× peak throughput per watt; 1.7–3.6× lower end-to-end latency
  • Interactive workloads: 2.1–4.1× higher performance
  • Deployment target: Within OpenAI compute infrastructure by end of 2026
  • Roadmap: Gen 2 in deep development; Gen 3 taking shape
  • Full-stack framing: OpenAI cast progress as a compounding system across data centers, custom chips, frontier models, developer platforms, and consumer products.
  • Compute partners: Microsoft, NVIDIA, AWS, AMD, Broadcom, Cerebras, CoreWeave, Oracle, SB Energy, and SoftBank — spanning cloud, accelerators, low-latency inference, data-center buildout, and energy.
  • Project Camellia (Georgia): Workload-shaped data-center design with closed-loop water conservation, local jobs, and annual independent public audits of commitments.
  • Serving efficiency (companion context): GPT-5.6 Sol–driven serving software cuts (~20% end-to-end serving cost; >15% speculative-decoding efficiency) and ARC-AGI-3 score lift from 13.3% to 38.3% with 6× fewer output tokens via retained-reasoning settings — model weights unchanged.

Architecture and development

Jalapeño was designed from scratch around language-model inference phases — prefill (compute-intensive) and decode (memory-bandwidth constrained) — minimizing data movement and communication delays. The network is integral to the architecture, allowing the entire workload to remain within one connected domain. Model state, including the KV cache, can be explicitly placed and kept local while the system activates the right combination of compute, memory, and networking for each phase.

OpenAI models accelerated the chip's development. The team moved from initial design to tape-out in nine months, which the companies believe is the fastest ASIC development cycle in high-performance semiconductors. AI-generated kernels for selected GPT‑OSS attention and mixture-of-experts blocks ran 1.5 to 1.8× faster than human-expert implementations. Using Codex with GPT‑Astra, the team brought three open-weight models not in the original production plan to high performance within two months.

Full-stack strategy

In a companion post the same day, OpenAI published a broader full-stack compute strategy: data centers, custom chips, frontier models, developer platforms, and consumer products each reinforce the next layer. Jalapeño is framed as a credible first-party inference path alongside partner silicon from NVIDIA, AMD, Cerebras, and others — giving OpenAI more control over how models run and what it costs to serve them, while still widely deploying merchant accelerators for training and inference.

The company ties chip and software efficiency to economics: as the cost of useful intelligence falls, more work becomes practical at scale (analysis, contract review, live financial scenarios, broader engineering exploration). OpenAI also highlighted Project Camellia in Georgia — a data-center design built around customer workloads, with closed-loop water conservation, local job creation, and annual independent public audits of its commitments.

Why it matters

Jalapeño demonstrates a full-stack advantage: OpenAI designs models, serving software, chips, memory, networking, and rack-scale systems together, using real workload data to improve every layer. The efficiency gains translate to lower cost per useful AI result, which can improve operating leverage and support broader adoption. OpenAI plans to continue widely deploying NVIDIA and partner accelerators for both training and inference alongside its own silicon.

Our take

Jalapeño's measured results on public models are a credible first step, but the real test will be sustained performance on OpenAI's own frontier models at production scale. The nine-month tape-out cycle — enabled by AI-assisted design — may prove more consequential than the chip itself if it establishes a repeatable cadence for future generations.

Sources