OpenAI and Broadcom to deploy 10 gigawatts of custom AI accelerators

OpenAI and Broadcom announced on October 13, 2025 a collaboration to deploy 10 gigawatts of AI accelerators designed by OpenAI itself, with racks targeted to start rolling out in the second half of 2026 and complete by the end of 2029.

The scale is the headline. Ten gigawatts of custom silicon represents one of the largest committed compute buildouts disclosed by any AI developer, and it puts OpenAI in the small group of companies — alongside hyperscalers like Google, Amazon and Meta — that design their own training and inference chips rather than relying solely on merchant suppliers. For Broadcom, which has built a sizable business co-developing custom accelerators for large cloud operators, the deal extends that model to the most valuable customer in AI.

What did the two companies actually agree to?

According to the joint announcement, made from San Francisco and Palo Alto, OpenAI will design the accelerators and systems, and Broadcom will develop and deploy them in partnership with OpenAI. The systems will include Broadcom accelerators and Ethernet solutions for both scale-up and scale-out connectivity.

The two companies have signed a term sheet to deploy racks incorporating the AI accelerators and Broadcom networking solutions. They say they already have long-standing agreements on the co-development and supply of the accelerators — meaning Monday's announcement formalizes and scales an existing relationship rather than starting one from scratch.

The racks will be deployed across OpenAI's own facilities and partner data centers, not a single dedicated campus. Every rack will be scaled with Ethernet and other connectivity solutions from Broadcom, including what Charlie Kawwas, President of Broadcom's Semiconductor Solutions Group, described as the company's "end-to-end portfolio of Ethernet, PCIe and optical connectivity solutions."

Why does OpenAI want to design its own chips?

OpenAI's stated rationale is that owning the chip design lets it compress what it learns from frontier-model development directly into hardware. The company has grown to over 800 million weekly active users, with adoption across global enterprises, small businesses and developers — demand that generic infrastructure serves less efficiently than silicon tuned to specific workloads.

"By building our own chip, we can embed what we've learned from creating frontier models and products directly into the hardware, unlocking new levels of capability and intelligence," said Greg Brockman, OpenAI co-founder and President.

Sam Altman framed the move as infrastructure-building rather than chip-making for its own sake. "Partnering with Broadcom is a critical step in building the infrastructure needed to unlock AI's potential and deliver real benefits for people and businesses," said Altman, co-founder and CEO of OpenAI. "Developing our own accelerators adds to the broader ecosystem of partners all building the capacity required to push the frontier of AI to provide benefits to all humanity."

That phrasing — "adds to the broader ecosystem of partners" — signals that Broadcom joins, rather than replaces, OpenAI's existing infrastructure relationships. Custom accelerators become one pillar of a portfolio approach to compute capacity.

What does the deal mean for Broadcom?

For Broadcom, the collaboration is a validation of its two core strategic bets in the AI datacenter market: custom accelerators and Ethernet.

"Broadcom's collaboration with OpenAI signifies a pivotal moment in the pursuit of artificial general intelligence," said Hock Tan, President and CEO of Broadcom. "OpenAI has been in the forefront of the AI revolution since the ChatGPT moment, and we are thrilled to co-develop and deploy 10 gigawatts of next generation accelerators and network systems to pave the way for the future of AI."

The networking side of the deal may matter as much as the silicon. AI clusters have traditionally relied on proprietary interconnects for scale-up — connecting GPUs within a rack — while Ethernet handled scale-out between racks. Broadcom's position is that Ethernet can do both, at lower cost and with an open standard.

"Our partnership with OpenAI continues to set new industry benchmarks for the design and deployment of open, scalable and power-efficient AI clusters," said Kawwas. "Custom accelerators combine remarkably well with standards-based Ethernet scale-up and scale-out networking solutions to provide cost and performance optimized next generation AI infrastructure."

The company said the collaboration "reaffirms our AI infrastructure portfolio leadership" and reinforces what it sees as the decisive role of custom accelerators and Ethernet as the scale-up and scale-out technology for AI datacenters.

What is the timeline?

The companies laid out a specific schedule:

• Co-development: OpenAI-designed accelerators and systems developed in partnership with Broadcom, with Broadcom accelerators and Ethernet solutions for scale-up and scale-out.
• Term sheet signed: covering deployment of racks with the accelerators and Broadcom networking.
• Deployment start: targeted for the second half of 2026.
• Full deployment: complete by the end of 2029.
• Locations: OpenAI facilities and partner data centers.

A roughly three-and-a-half-year deployment window ending in 2029 indicates a sustained, multi-phase buildout rather than a one-time delivery — and gives Broadcom a committed revenue runway in custom silicon through the end of the decade.

Why does this matter for the wider market?

The announcement tightens the connection between the largest AI developer and one of the two dominant designers of custom AI silicon. It also escalates the architectural contest inside the datacenter: Ethernet versus proprietary interconnects for scale-up networking. If OpenAI's racks scale entirely on Ethernet and deliver the performance OpenAI needs, that strengthens the case Kawwas made about standards-based, cost-optimized clusters — and pressures vendors whose networking stacks depend on closed interconnects.

For OpenAI, the deal is part of its answer to the binding constraint of frontier AI development: compute. With over 800 million weekly active users and training runs that grow more demanding with each model generation, owning the accelerator design gives OpenAI a measure of control over its cost structure and hardware roadmap that pure purchasing relationships cannot.

OpenAI says the collaboration supports its mission "to ensure that artificial general intelligence benefits all of humanity." The company also published a conversation on the deal: episode 8 of the OpenAI Podcast, in which Sam Altman and Greg Brockman sit down with Broadcom's Hock Tan and Charlie Kawwas.

The first hardware arrives in the second half of 2026 — which means the market will be watching, over the next year, for signs of how OpenAI's first-generation accelerator designs perform against the merchant silicon it will run alongside.