OpenAI and Broadcom to develop 10 GW of custom AI chips

OpenAI wants a substantial share of its future infrastructure to run on chips designed in-house. The company has agreed with Broadcom to develop and deploy 10 gigawatts of AI accelerators, with the first installations scheduled for the second half of 2026 and the project due to be completed by the end of 2029.

The deal expands OpenAI’s strategy beyond buying commercial processors. It also further increases the industrial and financial demands on a company that, in just three weeks, has announced large-scale computing plans with Nvidia, AMD and now Broadcom.

OpenAI will design the chip and Broadcom will build the system

OpenAI will handle the design of the accelerators and the systems that integrate them. Broadcom will help develop them and deploy the racks, while also providing the technologies needed to connect thousands of processors inside data centers.

An AI accelerator is a chip specialized in the mathematical operations used to train and run models. Compared with a conventional processing unit, it sacrifices versatility to gain speed and efficiency on those specific tasks.

The companies already have agreements covering co-development and supply, but today’s announcement gives the project a concrete scale: 10 GW of accelerators. They have also signed a term sheet for the rack deployment, a document that sets out the basis of the transaction before the definitive contracts are signed.

The systems will be installed both in OpenAI facilities and in its partners’ data centers. Broadcom will supply Ethernet, PCIe and optical connectivity to link the chips within each rack and between server clusters.

Ten gigawatts describe an industrial scale, not a chip count

The announced capacity does not indicate how many processors will be manufactured. The 10 GW refers to the electrical power associated with the accelerators OpenAI plans to deploy. The data centers’ total demand will be higher once networking, storage, cooling and other equipment are included.

As a measure of scale, 10 GW is equivalent to the nameplate capacity of ten one-gigawatt nuclear reactors operating simultaneously. That does not mean the project will consume that much power from day one: deployment will be gradual between 2026 and 2029.

Energy has therefore become a constraint as important as chip availability. Meeting the schedule will require semiconductor factories, data centers, grid connections and financing to be coordinated over several years.

OpenAI’s third major deal in three weeks

The Broadcom project follows two extraordinary announcements in terms of scale. On September 22, Nvidia and OpenAI announced a partnership to deploy at least 10 GW of Nvidia systems. The chipmaker said it intends to invest progressively up to $100 billion as that capacity is installed.

On October 6, OpenAI reached an agreement with AMD to deploy 6 GW of Instinct processors over several years, starting with 1 GW of MI450 chips in the second half of 2026. AMD also granted OpenAI rights to acquire up to 160 million shares, tied to deployment and share-price targets.

On paper, the three announcements represent 26 GW. They do not amount to a single finalized bill, nor do they guarantee that all the capacity will be installed: they include multiyear schedules, conditions and preliminary documents. Even so, they show the scale of the bet.

Sam Altman has cited an approximate cost of $50 billion for each gigawatt of AI infrastructure. Applying that figure purely as a rough guide to 26 GW would produce $1.3 trillion. That is not the contractual price of these agreements, but it illustrates why their financing and execution are raising questions even within the technology industry.

Custom chips to reduce dependence and tailor the hardware

OpenAI still needs Nvidia and AMD, but the Broadcom agreement gives it a third option. A custom chip can be tailored to the operations its models use most, reduce the cost per query and avoid relying entirely on one supplier’s products and timelines.

It is a familiar strategy among major cloud operators. Google develops its TPUs; Amazon has the Trainium and Inferentia chips; Microsoft is working on Maia; and Meta has its MTIA accelerators. The difference is that OpenAI does not own a data center network comparable to those companies’, so it must coordinate its silicon with cloud providers, financial partners and energy operators.

Broadcom is an important player in that market, even though it has less public visibility than Nvidia. It designs custom chips for major customers and controls a large share of the networking technologies used to connect servers. Its Ethernet strategy also competes with InfiniBand, the interconnect technology closely associated with Nvidia’s large-scale systems.

The first test will come in 2026

OpenAI says it has surpassed 800 million weekly active users, a scale that requires expanding both training capacity and the capacity used to answer queries. Designing its own chips could improve the economics of each interaction, but only if the systems deliver the expected performance and can be manufactured at volume.

The next verifiable milestone will be the start of deployment in the second half of 2026. Until then, the decisive questions will be whether the accelerators can be manufactured, whether electricity will be available and whether the preliminary agreements can be turned into financed orders. The announcement sets the size of the ambition; the first racks will show how much of it can become operational infrastructure.