openai and broadcom custom ai accelerator partnership

OpenAI and Broadcom announced a 10-gigawatt custom accelerator program on October 13, 2025. OpenAI will design the accelerators and systems. Broadcom will help develop and deploy the accelerator racks and Ethernet networks.¹

what the companies announced

The announcement describes a multi-year systems program, not a simple chip order. Its scale is stated in gigawatts, a measure of electrical power. That unit can describe the capacity of a fleet, but it does not reveal the number of accelerators. Chip count depends on rack design, use, cooling, networking, and the power used by each accelerator.

OpenAI said custom design lets it apply lessons from its models and products to the hardware. Broadcom described both scale-up links inside a computing system and scale-out links between systems.¹ The announcement did not limit the accelerator to inference, and it did not publish benchmark results.

deployment schedule

The public schedule has two bounds:

• deployment starts in the second half of 2026;
• deployment is planned to finish by the end of 2029.

The companies did not publish quarterly tapeout, prototype, yield, or unit-volume milestones. Earlier reports supplied some of those details, but they were not part of the joint announcement. This page therefore does not present them as settled facts.

what remains unknown

As of September 8, 2026, the announcement does not identify:

• the semiconductor foundry or process node;
• the package design or high-bandwidth memory supplier;
• the accelerator’s name, arithmetic formats, memory capacity, or thermal design power;
• the number or price of chips and racks;
• training or inference throughput;
• the program’s total contract value.

Broadcom has separately published its 3.5D XDSiP packaging platform, which can combine more than 6,000 square millimeters of silicon and as many as 12 high-bandwidth memory stacks.² That product exists, but Broadcom has not said in the partnership announcement that OpenAI’s system uses it. The same distinction applies to reported TSMC and SK hynix involvement.

business context

Broadcom reported $16.7 billion of AI-semiconductor revenue for its fiscal third quarter of 2026, up 221$21.7 billion for the fourth quarter.³ Those figures cover Broadcom’s AI semiconductor business, not the OpenAI program alone.

The program also does not remove OpenAI’s need for general-purpose accelerators. A custom chip can lower cost or power for a stable workload, while a programmable GPU can support new models and kernels sooner. A useful comparison will require measured throughput, energy, reliability, and total system cost after production hardware is available.

evidence note

The September 2025 version of this page described a $10 billion inference-chip order and supplied specific foundry, memory, package, cost, and unit-volume estimates. The October 2025 announcement did not confirm those terms. This revision removes them instead of carrying them forward as a secondhand specification.

references

[1] OpenAI and Broadcom. (2025, October 13). OpenAI and Broadcom announce strategic collaboration to deploy 10 gigawatts of OpenAI-designed AI accelerators.

[2] Broadcom. (2024, December 5). Broadcom delivers industry’s first 3.5D face-to-face technology for AI XPUs.

[3] Broadcom. (2026, September 2). Third-quarter fiscal 2026 results.