OpenAI on Wednesday unveiled Jalapeño, its first custom-designed inference processor, built in partnership with Broadcom and aimed squarely at the cost of running OpenAI models in production. The chip was co-developed using OpenAI's own AI models, and the company said early testing shows significantly better performance-per-watt than current state-of-the-art alternatives. The partnership was first announced in October, and Jalapeño is the first silicon to emerge from it.
Jalapeño is built for inference — the workload of running pre-trained models against live user prompts — not for pre-training the next generation of frontier models. OpenAI singled out real-time coding models as a target use case, emphasizing the chip's low operating cost on that workload. More compute-intensive pre-training runs will, for now, continue to lean on Nvidia hardware.
The strategic logic is straightforward. OpenAI's spending on inference scales directly with usage of ChatGPT, the API, and agentic products like Codex, and every incremental drop in cost-per-query falls to the bottom line. Google and Amazon have already gone down this path with their own AI accelerators — custom silicon designed to do one job, machine learning, more cheaply than a general-purpose GPU. OpenAI is the next major model provider to do the same.
Key facts
- 01OpenAI unveiled Jalapeño on Wednesday, its first custom-built inference processor, designed with Broadcom.
- 02The chip was co-developed using OpenAI's own AI models and is purpose-built for inference, not pre-training.
- 03Early testing shows significantly better performance-per-watt than current state-of-the-art alternatives, according to OpenAI.
- 04The Broadcom partnership was officially announced in October, with Jalapeño the first silicon to emerge from it.
- 05Google and Amazon have built similar custom AI accelerators to reduce dependence on Nvidia GPUs.
OpenAI president Greg Brockman described the company's approach on its in-house podcast shortly after the Broadcom deal was announced, framing the chip program as a targeted bet on workloads that off-the-shelf hardware was leaving on the table.
Broadcom is the manufacturing and design partner that turns OpenAI's workload knowledge into actual silicon. The arrangement mirrors how Google built its TPUs in collaboration with Broadcom for years, and how Amazon developed Trainium and Inferentia with similar outside silicon expertise. OpenAI gets a chip tuned to its kernels and memory patterns; Broadcom gets a marquee customer at the frontier of AI compute demand.
The Nvidia question hangs over all of this. Nvidia GPUs remain the default substrate for both training and inference across the industry, and OpenAI is one of its largest customers. Jalapeño does not change that overnight — pre-training and the most demanding inference workloads will still run on Nvidia hardware for the foreseeable future. But every workload OpenAI can move onto its own silicon is a workload it no longer has to rent at Nvidia's margins.
OpenAI framed the chip launch as part of a deliberate full-stack strategy. "OpenAI is not only developing frontier models or building products on top of them; it is designing the infrastructure underneath them: chip architecture, kernels, memory systems, networking, scheduling, deployment systems, and product experience," the company wrote in its announcement. "Because OpenAI operates across the stack, each layer can be optimized around the same goal: making its models faster, more reliable, and more affordable for users."
That vertical integration story now stretches from custom chips through data centers, model training, agentic products like Codex, and the consumer ChatGPT app. It is the same playbook Google has run for a decade — own the silicon, own the model, own the surface — applied to a company that did not exist as a model lab a decade ago.
The caveats are real. Jalapeño is still being tested, and OpenAI has not released independent benchmark data, power figures, or unit economics. "Significantly better performance-per-watt" is the company's own claim, against unnamed comparisons. Custom silicon programs are also expensive and slow: Google's TPU effort took years to reach competitive parity with Nvidia on a wide range of workloads, and Amazon's Trainium adoption has been uneven outside AWS's own services. Whether Jalapeño moves the needle on OpenAI's inference bill in 2026 or in 2028 is the question that matters.
The bigger signal here is about where the AI cost curve is headed. Inference, not training, is the recurring expense that determines whether frontier AI products can ever earn their keep, and the largest model providers are now all building their own chips to bend that curve. OpenAI joining Google and Amazon in custom silicon narrows Nvidia's monopoly on the most lucrative slice of AI compute and reshapes the competitive picture for every model lab that still rents its GPUs by the hour.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
