OpenAI unveils Jalapeño chip with performance gains over Nvidia flagship GPU

Aug 25, 2026

OpenAI unveils benchmark results for its first custom AI inference chip, Jalapeño, co-developed with Broadcom. Tested on SemiAnalysis's InferenceX suite, the 700W chip delivers up to 1.9 times more throughput per kilowatt and 3.6 times lower latency than Nvidia's flagship GB300 GPU. OpenAI plans to deploy Jalapeño in its data centers starting in late 2026, even as it maintains a deep financial partnership with Nvidia, which is backstopping a $105 billion data center campus in Ohio.

Jalapeño chip performance benchmarks

▪OpenAI's Jalapeño chip achieved 1.7 to 3.6 times lower end-to-end latency than Nvidia's GB200 and GB300 systems in tests covering the GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 models.
▪In benchmarks run on the SemiAnalysis InferenceX suite, OpenAI's Jalapeño chip delivered 1.5 to 1.9 times more throughput per kilowatt than Nvidia's GB200 and GB300 rack systems.
▪At low-latency operating points, OpenAI's Jalapeño chip achieved 8.6 to 104.3 times more throughput per kilowatt than Nvidia's GB300 at the GB300's fastest previous time-between-tokens settings.
▪The peak efficiency lead of OpenAI's Jalapeño chip over Nvidia's GB300 shrinks to roughly 1.5 times when compared using all-in utility power or when evaluated against a GB300 running multi-token prediction.

Inference ASIC technical specifications

▪OpenAI's Jalapeño chip is an Application-Specific Integrated Circuit (ASIC) designed specifically for AI inference workloads rather than model training.
▪Each package of OpenAI's first-generation Jalapeño chip pairs a compute die built on a TSMC 3nm-class process with six HBM4 stacks, totaling 216 GiB of memory at 15.4 TB/s.
▪OpenAI's Jalapeño chip is rated at 700W package thermal design power, though its measured sustained power stayed at or below 550W during testing.
▪OpenAI designed the architecture of the Jalapeño chip to minimize data movement and communication delays during the prefill and communication phases of inference processing.

OpenAI-Broadcom chip development partnership

▪OpenAI co-developed the Jalapeño chip with Broadcom under a 10GW custom chip deployment agreement signed in October 2025.
▪OpenAI is currently working on a second-generation chip that is approaching tapeout and has begun concept work on a third-generation chip.
▪OpenAI used its own artificial intelligence models to assist in the development of the Jalapeño chip, which completed a nine-month register-transfer level to tapeout cycle.

Deployment timeline through 2027

▪OpenAI plans to ramp up the deployment volume of its Jalapeño chips in 2027.
▪OpenAI's custom chip development occurs alongside its deep financial dependence on Nvidia, which agreed on August 17, 2026, to backstop up to $105 billion in financing for an OpenAI-leased data center campus in Ohio.

Nvidia competitive landscape

▪OpenAI's Jalapeño chip was not benchmarked against Nvidia's upcoming Vera Rubin platform, which is scheduled to power the first gigawatt of Nvidia systems OpenAI agreed to deploy in the second half of 2026.
▪Scaling the Jalapeño chip will require OpenAI to secure high-bandwidth memory (HBM4) supply, which Nvidia currently dominates through multi-year allocation deals with SK hynix.