OpenAI’s Jalapeño is a custom application-specific integrated circuit (ASIC) built with Broadcom primarily to serve trained language models. In selected InferenceX tests, OpenAI reports higher performance per watt and lower latency than Nvidia GB200 and GB300 systems. The important qualification is hiding in plain sight: these are comparisons between specific configurations and workloads, not proof that Jalapeño replaces Nvidia across AI computing.
What OpenAI built
Inference is the part of AI computing that turns a trained model into an answer. Every chatbot response, generated token, and agent step consumes inference capacity. Training creates the model; inference serves it, often repeatedly and at enormous scale.
Jalapeño is designed for that serving job. OpenAI developed it with Broadcom as a specialized inference ASIC rather than as a broad, general-purpose accelerator. That gives OpenAI more control over the path from model software to silicon, while accepting a narrower target than a platform intended to handle many different workloads.
The distinction matters. A chip tuned for serving language models can prioritize token delivery, predictable memory movement, latency, and power use instead of trying to cover every possible AI workload. It is closer to building a highly optimized delivery van than another all-purpose vehicle: less flexible, potentially much better at the route it was designed to run.
OpenAI’s stated strategy is not to discard Nvidia. The company says Nvidia and other partner accelerators will continue to be widely deployed for training and inference. Jalapeño is therefore best understood as an additional tool in a mixed infrastructure fleet.
The architecture is about moving data predictably
Jalapeño uses a memory-sliced, NUMA-style architecture. In practical terms, the chip organizes compute resources alongside local memory regions so that software can manage where data lives and reduce unnecessary movement. For language-model inference, that movement can become a major source of delay and power consumption.
The design also separates a low-latency collective network from a more general network-on-chip. That division reflects the workload: large language models often need many accelerator elements to exchange data quickly, while the rest of the system still needs general communication paths.
OpenAI has described Jalapeño as a full-stack project, not merely a piece of silicon. The company worked on the processor, memory behavior, networking, serving software, and model workloads together. That approach can remove bottlenecks that would be harder to address when the hardware and software come from separate design priorities.
It also creates a trade-off. Nvidia’s CUDA ecosystem is broadly portable across many applications. A specialized OpenAI stack can be highly efficient for known workloads, but that efficiency depends on the software and models being mapped well to the chip.
Reading the InferenceX results without the hype
OpenAI’s published InferenceX results cover three models: GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T. Against selected Nvidia GB200 and GB300 systems, OpenAI reports 1.5× to 1.9× higher peak performance per watt and 1.7× to 3.6× lower end-to-end latency.
The per-user throughput figures are also striking. Jalapeño reached 1,459 tokens per second per user with GPT-OSS 120B, compared with 535 for the comparison system. With DeepSeek R1 670B, the figures were 700 versus 169. With Kimi K2.5 1T, they were 694 versus 182.
For power-normalized peak throughput, the published numbers were 85,448 mixed tokens per second per kilowatt for Jalapeño versus 44,960 for Nvidia GB200 on GPT-OSS 120B; 19,641 versus 11,781 for DeepSeek R1 670B against Nvidia GB300; and 18,195 versus 11,862 for Kimi K2.5 1T against Nvidia GB300.
Those numbers are useful, but they need to stay attached to their conditions. Performance per watt is not the same as total system performance, and tokens per user is not the same as every possible measure of responsiveness. The largest reported multipliers came from operating points chosen around latency. One comparison reached 104.3× more throughput at Nvidia’s minimum-latency operating point; that does not mean Jalapeño is 104.3× faster overall.
The numerical results were supplied by OpenAI, and only selected runs were observed in OpenAI’s lab. The complete InferenceX and AgentX suites were not run. That makes the published data meaningful evidence about the tested configurations, but not a universal verdict on every production workload.
The comparison also did not include Nvidia’s Vera Rubin generation. That omission does not invalidate the GB200 and GB300 results, but it limits what the results can say about the newest Nvidia platform.
A chip designed around OpenAI’s software and models
The main strategic advantage is co-design. OpenAI can optimize data placement, kernels, scheduling, memory access, and network behavior around the models it expects to serve. The company also used its own models and Codex to assist with hardware optimization, kernel development, placement, and scheduling.
That description should not be mistaken for an autonomous chip-design process. AI assistance can accelerate exploration and implementation work; it does not turn a production semiconductor into a software-generated object with no engineering, verification, or validation around it.
The bargain is straightforward: specialization can improve efficiency when the workload is predictable, while generality makes a platform easier to reuse. Jalapeño’s results are most relevant to organizations serving large language models at scale, where a small improvement in latency or energy use can multiply across a large fleet.
For developers or companies looking for a broadly compatible accelerator, the story is different. Jalapeño is not presented as a retail component or a general-purpose replacement for the software portability of Nvidia’s platform.
Nine months from RTL to tape-out
OpenAI’s reported timeline places the start of initial register-transfer-level (RTL) work in February 2025 and the tape-out—the point at which the final chip design is sent for manufacturing—in November 2025. First silicon arrived in May 2026, according to the reported development sequence.
That nine-month interval refers specifically to the path from initial RTL work to tape-out. It is not the entire history of the hardware effort. OpenAI had identified an architecture concept in October 2024, and the company publicly unveiled Jalapeño with Broadcom on June 24, 2026. OpenAI published its first measured performance results on August 25, 2026.
The timeline is notable because custom silicon normally demands a long chain of architecture, design, verification, manufacturing, packaging, and system integration work. OpenAI’s use of AI-assisted tools is part of that engineering story, not a shortcut around it.
A strategic alternative, not an Nvidia replacement
Jalapeño matters because it gives OpenAI more control over a critical cost center: serving models. A specialized chip could help the company manage power, latency, and the cost of generating tokens as demand grows. It also reduces reliance on a single hardware strategy, even while Nvidia remains an important supplier.
OpenAI plans to begin deploying Jalapeño inside its own compute infrastructure by the end of 2026, with production expected to ramp in 2027. Those are forward-looking plans, not a completed deployment or a statement that the chip is broadly available to customers.
The public evidence also does not establish how Jalapeño performs on long-context, multi-turn agent workloads. InferenceX results show selected serving conditions; they do not answer every question about complex agents, where memory traffic, orchestration, and repeated model calls can change the balance between throughput and latency.
That is the cleanest way to read the announcement. Jalapeño is a serious attempt by OpenAI to own more of the inference stack, and its reported results are impressive within the published scope. But the story is not “OpenAI defeated Nvidia.” It is that a model company is building purpose-made infrastructure for the part of AI computing it expects to use most—and keeping Nvidia in the fleet while it does so.
