OpenAI unveils first custom AI chip “Jalapeño”, targets 50% cost reduction in Nvidia alternative push

OpenAI has moved to reduce its dependence on Nvidia by unveiling a custom artificial intelligence chip developed in partnership with Broadcom. The chip, named Jalapeño, is designed to cut computing costs by around 50% compared with conventional AI graphics processing units.

On Wednesday, OpenAI and Broadcom presented Jalapeño, described as OpenAI’s first Intelligence Processor. It has been built as a custom chip aligned with the company’s long-term vision for large language model workloads.

The processor was developed from design to manufacturing in just nine months, a timeline widely considered among the fastest in high-performance semiconductor development. OpenAI attributed the speed to its use of its own AI models to accelerate parts of the design and optimisation process, effectively creating a feedback loop where its models help improve the infrastructure used to train future systems.

Early testing indicates Jalapeño delivers significantly improved performance per watt compared with leading current-generation hardware. Engineering samples are already running machine learning workloads in laboratory conditions at target frequency and power levels, including the GPT-5.3-Codex-Spark model. A full technical report is expected in the coming months.

The move marks a strategic shift for OpenAI, which has been one of Nvidia’s largest customers since 2022, relying heavily on its GPUs to train and run advanced AI systems at scale.

The introduction of Jalapeño signals an ambition to make AI computing more efficient and affordable. OpenAI President Greg Brockman said the chip forms part of a broader strategy to deliver “faster, more reliable, and more affordable AI for people and businesses.”

Deployment is expected to begin by the end of 2026, with Microsoft set to acquire around 40% of the chips in the initial rollout. Broadcom chief executive Hock Tan confirmed a longer-term roadmap that includes a follow-up version in 2028, with annual iterations planned thereafter.

Broadcom will also support production through its networking technologies, including Tomahawk chips, while Celestica will handle board manufacturing, rack systems, and large-scale integration.

Holger Mueller, an analyst at Constellation Research, said the development could pose a challenge to Nvidia, noting: “Companies that create large language models (LLMs) need to reduce costs for AI. The best partner for this is the company that has helped Google, a leader in AI cost-cutting. This is a warning for Nvidia.”

The strategy mirrors an approach pioneered by Google more than a decade ago with its Tensor Processing Units, where custom silicon was used to optimise AI workloads. However, the scale and urgency at OpenAI are significantly greater.

With hundreds of millions of users relying on its systems daily, even marginal efficiency gains could translate into substantial cost savings. OpenAI’s plan to deploy Jalapeño at scale by the end of 2026 underscores its shift from experimentation to deep infrastructure reliance on custom-built chips.