OpenAI And Broadcom Name Jalapeño AI Accelerator
OpenAI and Broadcom unveiled Jalapeño, their first custom AI accelerator, with initial deployment targeted by the end of 2026 and a ramp expected in the following years.
OpenAI Moves From Chip Buyer To Chip Designer
OpenAI and Broadcom have unveiled Jalapeño, the first custom AI accelerator from their partnership and OpenAI's clearest move into silicon built for its own workloads.
The companies describe the chip as an Intelligence Processor and as the first accelerator in a platform intended to make advanced AI faster and more reliable.
The project changes OpenAI's infrastructure position.
The company has been one of the largest buyers of Nvidia GPUs since the generative AI boom began in 2022, but its demand for inference capacity keeps forcing it to seek other sources of silicon.
Jalapeño is designed for inference, the process of serving models to users in ChatGPT and other applications.
OpenAI President Greg Brockman said the company completed the design cycle in nine months while using its own AI models to speed the work.
The claim gives the announcement a product-development angle as well as an infrastructure angle: OpenAI is using AI tools to accelerate the design of hardware that will then run AI services.
Broadcom Puts Custom Silicon Behind The Stack
Broadcom will make the chips, giving OpenAI a manufacturing and custom-silicon partner rather than a fully internal chip supply chain.
The companies had already announced in October that they planned to develop and deploy racks of OpenAI-designed chips after 18 months of joint work.
The earlier plan pointed to deployment starting late this year and to a long-term buildout large enough to require 10 gigawatts of power.
The new disclosure narrows the story to a named chip and a physical sample scheduled for delivery to OpenAI on Wednesday.
Hock Tan said there would be small prototype development in late 2026 before a broader ramp.
He said the work should scale in '27 and go full tilt in the first half of '28.
Those timing markers matter because they separate a chip reveal from actual infrastructure delivery.
Compute Demand Still Sets The Limit
OpenAI has not abandoned outside accelerator suppliers.
The company has deals involving Amazon Web Services Trainium chips, Nvidia rival AMD and Cerebras, alongside its dependence on Nvidia GPUs.
Jalapeño therefore looks less like a replacement for the existing AI chip market and more like another supply path for a company that says it cannot get compute quickly enough.
Broadcom's role also reflects a wider shift among hyperscalers and frontier labs.
Custom ASICs can be less flexible than GPUs, but they can be designed around specific AI tasks and cost targets.
OpenAI said it also designed large parts of the computer system where Jalapeño will be used, extending the project beyond a single chip.
The economics remain tied to inference growth rather than a public sales forecast for Jalapeño.
OpenAI and Broadcom named the accelerator, the use case and the deployment path, but they did not publish chip volume, system pricing, power efficiency data or customer access beyond OpenAI's own workloads.
Those gaps keep the announcement at the infrastructure-buildout stage rather than proving cost relief for ChatGPT-scale demand.
The launch gives OpenAI a named accelerator, a Broadcom manufacturing partner, a late 2026 prototype target and a '27-to-'28 scaling window.
The unresolved operating burden is whether prototype silicon can become enough deployed inference capacity to ease OpenAI's compute shortage without adding another power bottleneck to its AI infrastructure plans.
