The headline, of course, is that OpenAI has beaten Nvidia. According to the published results, Jalapeño delivers 1.5 to 1.9 times more AI work per watt at maximum throughput than the compared systems. At the same time, end-to-end latency was 1.7 to 3.6 times lower.

GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T were tested. This means the architecture does not just work with OpenAI's own models. The chip is designed for a variety of modern language models, but remains a specialized inference chip. It runs models; it does not train them.

This distinction is crucial. Jalapeño is not a drop-in GPU replacement for every conceivable workload. The chip was designed from the ground up to execute language models quickly and efficiently. Less data movement, tailored memory and interconnect architecture, and a clear focus on real-world inference workloads are intended to extract more usable work from every watt.

Yet good hardware alone is not enough. An AI chip requires highly optimized kernels. These are small programs that dictate precisely how the chip executes specific mathematical operations. A poor kernel can throttle outstanding hardware. And writing good kernels is painstaking, specialized work.

This is precisely where Nvidia's historical advantage lies. For years, CUDA has provided programming languages, libraries, tools, documentation, and a vast talent pool of experienced developers. Anyone building a new chip architecture must not only deliver better silicon; they must also overcome this software barrier.

OpenAI is attempting to solve this problem differently. Instead of first building a human-friendly ecosystem, Jalapeño is programmed using the low-level language Gluon. According to the report by SemiAnalysis, kernels can span around 3,000 lines and are secured through correctness proofs and a dedicated verification tool.

Codex is increasingly taking over this work. When an MLA kernel for DeepSeek R1 was missing, the system was able to quickly generate a working, efficient implementation, according to SemiAnalysis. OpenAI also reports that three open models not originally planned for Jalapeño were brought up to high performance within two months.

On selected components of GPT-OSS, AI-generated implementations ran 1.5 to 1.8 times faster than existing versions written by human experts, according to OpenAI. This applies specifically to individual blocks and not the complete model. Nevertheless, it illustrates the direction of travel.

OpenAI also leveraged AI in the chip design itself. Models helped explore variants, optimize circuitry, and shorten the loops between design, measurement, and verification. From initial design to tape-out took nine months, according to OpenAI and Broadcom.

This creates a feedback loop: existing models help design better hardware. That hardware runs subsequent models more efficiently. New models, in turn, write better software for the next hardware generation. This is not fully autonomous self-improvement. Highly skilled engineers remain central. But their leverage is growing massively.

The results nonetheless warrant a sober assessment. SemiAnalysis personally verified the runs at the OpenAI lab, but did not execute the full InferenceX suite. The underlying numbers were provided by OpenAI. More realistic AgentX tests involving long contexts and multi-turn interactions are still missing.

The comparison with Nvidia is also less definitive than some headlines suggest. Jalapeño currently exists as early silicon and must still be qualified for widespread deployment. Nvidia's Rubin platform is likewise still in an early optimization phase. Both software stacks will become faster.

OpenAI plans the initial deployment in its own infrastructure by the end of 2026. The second generation is already well into development, and the third is in preparation. Whether Jalapeño maintains the same advantages under real-world production workloads remains to be seen.