OpenAI and Broadcom have unveiled Jalapeño, their first jointly developed custom AI chip — a large-scale inference accelerator designed from the ground up for large language models. The two companies revealed the chip in late June, describing a development timeline of just nine months from initial design to manufacturing tape-out, which the partners claim is among the fastest ASIC development cycles ever achieved in high-performance computing.
What Happened
Jalapeño is a reticle-sized application-specific integrated circuit built entirely around the demands of modern LLM inference. Unlike general-purpose AI accelerators adapted from earlier workloads, the chip reflects the specific requirements of running ChatGPT, Codex, and OpenAI’s expanding portfolio of agentic products at scale. Notably, OpenAI used its own AI models to accelerate portions of the chip’s design and optimization process — applying its own technology to build the hardware that will run it.
Early benchmark results shared by OpenAI indicate that Jalapeño will deliver significantly better performance per watt compared to current state-of-the-art alternatives, including the Nvidia GPUs that presently power the majority of OpenAI’s infrastructure. The companies have not publicly disclosed the exact process node, but manufacturing is confirmed to be handled by TSMC.
Why It Matters
The unveiling marks OpenAI’s first serious step toward reducing its dependence on Nvidia, which currently supplies the bulk of AI training and inference hardware across the industry. For a company operating at the scale of OpenAI — serving hundreds of millions of users across ChatGPT and the API — even modest gains in performance per watt can translate into hundreds of millions of dollars in annual infrastructure savings. The move also gives OpenAI more control over its own supply chain and long-term compute costs.
OpenAI is not alone in pursuing custom silicon. Anthropic is in early discussions with Samsung to manufacture its own custom chip targeting the 2nm process node — a clear signal that the largest AI labs are all moving to secure independent hardware roadmaps rather than remaining dependent on third-party GPU suppliers.
Background and Context
The AI chip race has intensified significantly in 2026. Google built its Tensor Processing Units years ago, Amazon has Trainium and Inferentia, and Microsoft operates Azure Maia across its data centers. OpenAI and Anthropic entering the custom silicon space marks a new phase in which the AI software leaders are becoming hardware companies by necessity, not just choice.
The scale of the planned Jalapeño deployment is substantial: OpenAI and Broadcom have announced ambitions to eventually deploy systems drawing up to 10 gigawatts of power across data centers. That energy footprint raises important questions about the AI sector’s overall electricity demand — a concern that has become harder to ignore as Google’s AI infrastructure recently drove a record 37 percent surge in its power consumption, prompting scrutiny from utility regulators and climate researchers alike.
What Comes Next
Initial deployment of Jalapeño chips is targeted for before the end of 2026, with broader rollout planned across subsequent years. Broadcom and OpenAI have described their collaboration as a multi-generation platform, implying future chip revisions are already in the planning stages. OpenAI’s commitment to custom silicon at this scale is also a broader signal about the company’s long-term ambitions: a research lab that launched barely a decade ago is now designing its own data center hardware and building toward 10-gigawatt compute clusters. The industry is watching to see whether Jalapeño delivers on its performance-per-watt promises once deployed in production environments — and whether it gives OpenAI a meaningful infrastructure edge over competitors who remain more reliant on Nvidia.
