OpenAI debuted its 700W Jalapeño ASIC at Hot Chips 2026, claiming 1.9x higher throughput per kilowatt and 3.6x lower latency than the 1,400W Nvidia GB300, with Broadcom as co-developer.
Just over a week after Nvidia agreed to backstop up to $105 billion in financing for its data centers, OpenAI took the stage at Hot Chips on Tuesday with benchmarks claiming its first in-house silicon outperforms Nvidia’s latest rack systems. Co-developed with Broadcom, the Jalapeño inference ASIC delivered 1.5 to 1.9 times greater throughput per kilowatt and reduced end-to-end latency by 1.7 to 3.6 times compared to Nvidia’s GB200 and GB300 configurations. Running on a 700W part, Jalapeño was pitted against accelerators rated at 1,200W and 1,400W using SemiAnalysis’s public InferenceX suite. OpenAI plans to begin deploying the chip in its own data centers later this year.
The evaluation covered three open-weight models: GPT-OSS 120B, DeepSeek R1 670B, and Moonshot AI’s 1-trillion-parameter Kimi K2.5. OpenAI reported its most significant performance advantages at low-latency operating points, where it claims throughput per kilowatt surged between 8.6 and 104.3 times faster than the GB300’s fastest previous time-between-tokens settings.
While OpenAI normalized results to each accelerator’s published package thermal design power (TDP), the company noted that Jalapeño’s measured sustained power remained at or below 550W during testing. An appendix analysis comparing all-in utility power per accelerator—1.18kW for Jalapeño versus 2.55kW for the GB300—yields narrower performance gaps. Similarly, when benchmarking Jalapeño against a GB300 configured for multi-token prediction, OpenAI’s peak efficiency advantage contracts to approximately 1.5 times.
Several important caveats accompany the results. Jalapeño was not tested against Vera Rubin, the Nvidia platform slated to power the first gigawatt of systems OpenAI has agreed to deploy in the second half of 2026. Additionally, the chip does not support training workloads, an area where Nvidia’s hardware remains dominant. The primary comparisons also matched Jalapeño’s single-token prediction against GB300 setups running the same task, despite production deployments commonly utilizing multi-token prediction. SemiAnalysis, which conducted the InferenceX tests alongside OpenAI engineers in the company’s lab, described the silicon as “beating every Nvidia, AMD, and Google chip we have been able to test.”
Each Jalapeño package, unveiled in June following a nine-month RTL-to-tapeout cycle, pairs its compute die with six HBM4 stacks, delivering 216 GiB of memory at 15.4 TB/s. By contrast, the GB300 carries 288GB of HBM3E under a 1,400W rating, meaning Jalapeño offers roughly 50% more memory per watt of rated power. According to OpenAI’s Hot Chips presentation, the architecture’s primary bottleneck targets exposing aggregate HBM bandwidth rather than increasing raw capacity.
High-bandwidth memory remains the semiconductor industry’s tightest commodity. Samsung, SK hynix, and Micron have already sold their HBM capacity through 2027. The shortage is severe enough that Nvidia is reportedly testing downgraded Rubin Ultra configurations with as little as 192GB, and SK hynix CEO Kwak Noh-jung has warned that 2027 will mark the peak of the crunch. At the same Hot Chips conference on August 23, Micron noted that HBM consumes roughly three times the wafer area of DDR5 for equivalent capacity—a penalty that widens with each generation. Scaling Jalapeño across the 10GW deployment agreement OpenAI signed with Broadcom last October would position the company as a substantial new claimant to HBM4 supply, currently dominated by Nvidia through multi-year allocation deals with SK hynix.
Development is already accelerating beyond the initial release. A second-generation chip is approaching tapeout within months, according to Bloomberg, with concept work on a third generation underway. The first iteration reportedly utilizes a TSMC 3nm-class process, keeping OpenAI in the same wafer, memory, and advanced packaging queues as Blackwell and Rubin for the foreseeable future.
OpenAI is securing these critical inputs while simultaneously deepening its financial reliance on the very company it just benchmarked. On August 17, Nvidia agreed to provide up to $105 billion in financing for an OpenAI-leased data center campus in Ohio. “Nvidia is a really good partner, and we continue to need a lot of Nvidia,” Richard Ho, OpenAI’s vice president of hardware, told Bloomberg following the announcement.