Quick Facts
- OpenAI and Broadcom co-developed Jalapeño, a purpose-built AI inference chip, from design to manufacturing tape-out in nine months using TSMC’s 3-nanometer process.
- Early testing shows Jalapeño delivers performance per watt substantially better than current state-of-the-art chips, though final benchmarks have not been published.
- OpenAI aims to power 10 gigawatts of compute with custom chips by 2029, with Microsoft expected to purchase 40 percent of initial production.
OpenAI and Broadcom on June 24 unveiled Jalapeño, OpenAI’s first custom silicon product. The chip is an application-specific integrated circuit built from scratch for large language model inference and marks the first product in a multi-generation compute platform the two companies are developing together.
Broadcom CEO Hock Tan and President Charlie Kawwas formally delivered the chip to OpenAI CEO Sam Altman and President Greg Brockman. Celestica is handling board, rack, and system integration as the third partner in the program.
A Nine-Month Design Cycle
OpenAI and Broadcom completed the chip’s design-to-tape-out cycle in nine months, a timeline the companies describe as potentially the fastest ASIC development ever achieved in high-performance advanced semiconductors. OpenAI’s own AI models accelerated parts of the design and optimization process, creating a closed loop where models serving end users helped build the hardware that will run future models.
Richard Ho, head of OpenAI’s hardware program, said the team “optimized the architecture around the kernels, memory movement, networking, and serving patterns that matter most for frontier AI models.” Engineering samples are already running machine learning workloads in laboratory testing at production target frequency and power levels, including GPT-5.3-Codex-Spark.
Why OpenAI Built Its Own Chip
Jalapeño is an ASIC, which is less flexible than a general-purpose GPU but less expensive and optimized for specific tasks. The architecture reduces data movement and balances compute, memory, and networking to push realized utilization closer to theoretical peak performance.
Tan told CNBC that companies serious about leading in AI should not rely on third-party GPUs: “It’s such a key part.” He added that compute demand from Broadcom’s six customers is “simply insatiable” and that elevated demand is expected through 2028.
Greg Brockman framed the chip as part of a broader infrastructure play. “The world is moving to a compute-powered economy,” he said. “Jalapeño is part of our long-term full-stack infrastructure strategy to make compute more abundant, resulting in AI which is faster, more reliable, more affordable for people and businesses.” Brockman also told CNBC that OpenAI “cannot get compute fast enough.”
Deployment Timeline
Broadcom’s Tan said the companies will begin small prototype deployments in late 2026 before scaling. Microsoft is expected to purchase 40 percent of early chip production. OpenAI says meaningful volume will arrive in 2027.
The two companies have also established a multi-generation roadmap. The next chip generation is planned for 2028, with annual iterations after that. OpenAI’s stated goal is to have custom silicon powering 10 gigawatts of compute by 2029.
For software and technology executives, the announcement signals that the largest AI labs are moving fast to control their own compute supply chains. If AI-assisted chip design continues to compress development timelines, the cost of running AI workloads at scale could fall significantly over the next several years.
This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.