OpenAI is moving deeper into the infrastructure behind artificial intelligence. With Broadcom, it has unveiled Jalapeño, a custom chip designed specifically for large language model inference.
The announcement signals a shift from relying only on general-purpose or third-party AI hardware toward a more controlled stack. OpenAI says the goal is to run models faster, more reliably, and at lower cost.
A chip built for inference, not general computing
Jalapeño is described as OpenAI’s first so-called Intelligence Processor. It is a custom accelerator built for the inference stage of large language models, the phase where a trained model responds to user requests.
According to the source, OpenAI says Jalapeño is not a modified general-purpose chip. It was designed from scratch for modern LLM inference, making it part of a broader push to tailor hardware more closely to the workloads that power AI products.
The chip is also the first entry in a multi-generation platform that OpenAI and Broadcom are building together. That matters because a platform approach suggests the companies are not treating Jalapeño as a one-off experiment, but as the beginning of a longer hardware roadmap.
How the partnership is divided
The project brings together several companies with different roles in the hardware chain. OpenAI handles the chip design. Broadcom contributes silicon manufacturing and networking technology, including its Tomahawk networking chips. Celestica is responsible for boards, racks, and system integration.
The unveiling included a handoff of the first wafer. Broadcom CEO Hock Tan and President Charlie Kawwas presented it to OpenAI CEO Sam Altman and President Greg Brockman.
For OpenAI, the move marks its first step into custom hardware after years focused on models and products. For Broadcom, the partnership places its manufacturing and networking expertise inside a high-profile AI infrastructure effort.
The performance claim is still unproven
OpenAI says early tests showed performance per watt that is "substantially better" than current state-of-the-art hardware. That is the central technical claim around Jalapeño, because performance per watt affects both operating cost and the amount of infrastructure needed to serve AI models at scale.
But the source makes clear that these numbers are self-reported and not finalized. A technical report is expected to follow, and the available information does not yet say which chips Jalapeño was tested against, what tasks were used, or what conditions applied.
That uncertainty is important. Without those details, it is not possible to fairly compare Jalapeño with competing inference hardware. The claim may prove meaningful, but it still needs technical backing.
The architecture reportedly reduces data movement and pushes utilization closer to its theoretical maximum. Engineering samples are already running ML workloads in the lab, including the GPT-5.3-Codex-Spark model. The same model currently runs on Cerebras hardware, which also specializes in inference.
Development moved quickly
OpenAI says the path from design to tape-out took just nine months. The company describes that as the fastest ASIC development cycle for high-performance semiconductors it is aware of.
OpenAI’s own models helped speed up parts of the design process. That detail is notable because it shows AI being used not only as the workload for new hardware, but also as a tool in building that hardware.
At the same time, the source notes that rumors about OpenAI’s chip plans have been circulating since 2023. So while the tape-out timeline is presented as unusually fast, the broader hardware ambition has been visible for longer.
Why OpenAI wants more control of the stack
The strategic argument behind Jalapeño is straightforward: if OpenAI can control more of the path from chip to product, it may be able to tune each layer for its own models and services.
That could affect several practical areas:
- Cost: custom inference hardware is intended to make running AI models cheaper.
- Reliability: OpenAI says custom hardware can make deployment more dependable.
- Performance: the company is emphasizing better performance per watt, though the claim still awaits fuller technical evidence.
- Scale: large-scale deployment is planned for late 2026.
Broadcom CEO Tan says the first deployment is planned for late 2026 at gigawatt scale, together with Microsoft and other partners. Broadcom has reportedly demanded that Microsoft guarantee it will buy 40 percent of the chips to secure the first phase.
That planned deployment shows the ambition behind Jalapeño. This is not just a lab project or a small hardware test. OpenAI and Broadcom are positioning it as infrastructure for large-scale AI operations.
Still, the most important open question remains technical. Jalapeño has been introduced as a purpose-built LLM inference chip with strong efficiency claims, but the industry will need the promised technical report before it can judge how large the advantage really is.