Last reviewed
Jun 2, 2026
Sources
13 citations
Review status
Source-backed
Revision
v1 ยท 2,036 words
Add missing citations, update stale details, or suggest a clearer explanation.
The OpenAI-Broadcom collaboration is a strategic partnership, announced on October 13, 2025, under which OpenAI designs its own custom AI inference accelerators and Broadcom co-develops, manufactures, and deploys them together with Broadcom's Ethernet-based networking. The two companies committed to deploy 10 gigawatts (GW) of OpenAI-designed accelerators and associated network systems, with rack deployments targeted to begin in the second half of 2026 and complete by the end of 2029.[1][2] The arrangement is one of several large-scale compute deals OpenAI struck in 2025 to secure hardware for training and serving its frontier models, and it marks OpenAI's move into designing its own silicon rather than relying solely on merchant GPUs.[3][4]
OpenAI and Broadcom described the collaboration as a plan to "develop and deploy 10 gigawatts of custom AI accelerators."[1] Under the terms made public, OpenAI is responsible for designing the accelerators and the systems that house them, while Broadcom develops and deploys those systems and supplies the networking that ties them together.[1][2] The deployments are intended to run across OpenAI's own facilities and partner data centers, and the racks are scaled "entirely with Ethernet and other connectivity solutions from Broadcom."[1]
Ten gigawatts is a measure of electrical power capacity rather than a chip count, but the scale implies a very large number of devices. Industry coverage at the time of Broadcom's fiscal 2025 disclosures suggested a custom-silicon order of this magnitude could amount to several million processors.[5] The chip at the center of the effort is oriented toward inference, the stage in which a trained model produces outputs, rather than toward model training, where OpenAI continues to lean heavily on GPUs from other suppliers.[6][7]
Unlike OpenAI's contemporaneous agreements with chip vendors, the Broadcom collaboration was announced without a disclosed dollar value and without equity warrants or cross-investment between the two firms.[4] The companies framed it as a multi-year engineering and deployment partnership rather than a financing arrangement.[1][4]
By 2025 OpenAI's demand for computing capacity had grown faster than any single supplier could comfortably meet, and the company began pursuing hardware on multiple fronts at once. Designing a custom accelerator was part of that strategy. Reporting from late 2024 indicated OpenAI was working with Broadcom and the foundry TSMC on its first in-house chip, an inference-focused design, and that OpenAI had abandoned earlier ambitions to build its own chip fabrication plants on cost and timing grounds.[6][7] The internal silicon team was led by Richard Ho, an engineer who had previously worked on Google's tensor processing unit program, and it grew to roughly 40 people as the design progressed toward a tape-out.[6]
The rationale OpenAI gave for building its own chip is that doing so lets it fold lessons from training and operating frontier models directly into the hardware. As OpenAI president Greg Brockman put it, "by building our own chip, we can embed what we've learned from creating frontier models directly into hardware."[1][8] Custom silicon also offers a route to reduce dependence on Nvidia, whose GPUs dominated AI data centers and commanded high margins; several large AI operators, including Google, Meta, and ByteDance, had already commissioned custom accelerators through Broadcom before OpenAI did.[5][9]
Broadcom's part of this background is its established custom-silicon, or XPU, business. The company designs application-specific accelerators for hyperscale customers and pairs them with its own networking products, an end-to-end approach that spans the accelerator, the Ethernet switches, the network interface cards, and the optical interconnects.[9][4] In its fiscal third quarter of 2025 (reported in September 2025), Broadcom said it had secured more than 10 billion dollars of orders for AI racks from a new customer, which analysts widely assumed to be OpenAI. Broadcom later clarified that the specific 10-billion-dollar order was not OpenAI's, and in December 2025 identified that customer as Anthropic.[10][11]
The division of labor is the defining feature of the collaboration. OpenAI owns the chip and system design; Broadcom turns that design into manufacturable, deployable hardware and supplies the network fabric. The joint announcement states plainly that "OpenAI will design the accelerators and systems, which will be developed and deployed in partnership with Broadcom."[1]
| Function | Responsible party |
|---|---|
| Accelerator (chip) architecture and system design | OpenAI[1] |
| Co-development of accelerators and network systems | Broadcom[1][2] |
| Manufacturing (foundry) | TSMC (per prior reporting on OpenAI's custom chip)[6] |
| Networking fabric (Ethernet, PCIe, optical) | Broadcom[4][8] |
| Rack assembly and deployment | Broadcom[1][2] |
| Hosting | OpenAI facilities and partner data centers[1] |
Broadcom's networking contribution is built on standards-based Ethernet rather than a proprietary interconnect. Charlie Kawwas, president of Broadcom's Semiconductor Solutions Group, said that "custom accelerators combine well with standards-based Ethernet networking to provide cost and performance optimized AI infrastructure."[1] The wider portfolio Broadcom brings to the racks includes Ethernet switching, PCIe connectivity, and optical interconnects.[4][8] Broadcom's CEO, Hock Tan, said the company was "thrilled to co-develop 10 gigawatts of accelerators and network systems with OpenAI," and OpenAI CEO Sam Altman called the partnership "critical to building infrastructure needed to unlock AI's potential."[1]
Financial terms were not part of the announcement. Whereas OpenAI's deals with other suppliers came bundled with multibillion-dollar investments or equity, the Broadcom collaboration was presented purely as a hardware engineering and deployment program, with no investment commitment or warrant disclosed.[4]
The headline figure is 10 gigawatts of accelerator capacity. To put that in context, a single large AI data center campus is often measured in hundreds of megawatts, so 10 GW represents capacity on the order of many such campuses dedicated to OpenAI's custom hardware.[3][4] The accelerators are intended chiefly for inference, the high-volume, cost-sensitive workload of running models in production, which is also the segment where reducing reliance on expensive merchant GPUs has the largest financial effect.[6][5]
A distinctive technical choice is that the racks are scaled "entirely with Ethernet."[1] Many high-end AI clusters historically used proprietary or specialized scale-up interconnects, but Broadcom has pushed Ethernet as the fabric for AI at both the scale-up and scale-out levels, supported by its switch silicon. The aiwiki article on the Broadcom Tomahawk 6 switch describes the class of high-radix Ethernet switching Broadcom markets for exactly this kind of large AI deployment. By standardizing on Ethernet, OpenAI and Broadcom aim for a networking stack that is openly specified and, in their framing, optimized for cost and performance at scale.[1][4]
The collaboration follows a multi-year deployment schedule.
| Date | Milestone |
|---|---|
| Late 2024 | Reports that OpenAI is designing a custom inference chip with Broadcom and TSMC, and has dropped foundry-building plans[6] |
| September 2025 | Broadcom discloses a new 10-billion-dollar custom-silicon customer on its fiscal Q3 call; OpenAI widely assumed (later clarified)[5][10] |
| October 13, 2025 | OpenAI and Broadcom announce the 10 GW custom-accelerator collaboration[1][2] |
| Second half of 2026 | Deployment of accelerator and network racks targeted to begin; OpenAI's first custom chip expected to debut[1][7] |
| End of 2029 | Full 10 GW deployment targeted for completion[1][2] |
OpenAI's first custom chip had been expected to reach manufacturing around 2026, consistent with the announced start of rack deployments in the second half of that year.[6][7] The companies described the end-2029 completion target as the point by which the full 10 GW would be in place.[1][2]
The Broadcom collaboration was one of a cluster of infrastructure agreements OpenAI announced across 2025, which together pointed to roughly 26 GW of contracted AI compute from chip and system partners, on top of the data-center buildout under the Stargate Project.[3][12] The Broadcom deal is distinguished from the others by being a custom-silicon program with no disclosed investment, in contrast to the GPU-supply deals that carried large financial components.
| Partner | Capacity | Hardware | Financial structure | First deployment |
|---|---|---|---|---|
| Broadcom | ~10 GW | OpenAI-designed custom accelerators + Ethernet | No disclosed dollar terms or equity[1][4] | H2 2026[1] |
| Nvidia | At least 10 GW | Nvidia systems (Vera Rubin platform) | Nvidia to invest up to $100B as gigawatts deploy[13] | H2 2026[13] |
| AMD | 6 GW | AMD Instinct GPUs | Warrant for OpenAI to buy up to 160M AMD shares[3] | 2026 (1 GW initial)[3] |
| Oracle | ~5 GW (Stargate) | Data-center capacity / hosting | Reported ~$300B over five years[12] | Phased from 2025[12] |
The Nvidia partnership, announced September 22, 2025, covers at least 10 GW of Nvidia systems with the first gigawatt slated for the second half of 2026 on the Vera Rubin platform, and Nvidia intending to invest up to 100 billion dollars in OpenAI as each gigawatt is deployed.[13] The AMD agreement covers 6 GW of AMD Instinct GPUs and includes a warrant letting OpenAI acquire up to 160 million AMD shares.[3] Oracle supplies data-center capacity as part of the Stargate buildout, reported in the range of hundreds of billions of dollars over five years.[12] Against that backdrop, the Broadcom collaboration supplies the path for OpenAI to run inference on hardware of its own design rather than on purchased GPUs.[4][5]
The collaboration is significant on two levels. For OpenAI, it represents vertical integration into silicon: the company is no longer only a buyer of accelerators but a designer of them, with the stated aim of tailoring hardware to its own models.[1][8] For the wider market, it added OpenAI to the roster of large AI operators commissioning custom chips through Broadcom, reinforcing a shift toward application-specific accelerators and away from exclusive reliance on Nvidia GPUs, particularly for inference workloads that dominate the cost of operating large models.[5][9] Commentators noted that if OpenAI's custom accelerators handle inference at scale, they could divert some of the high-volume workloads that underpin merchant-GPU revenue.[4][9]
The choice to standardize on Ethernet rather than a proprietary fabric was also read as a vote for open networking standards in AI clusters, an area where Broadcom has positioned its switch products against vendor-specific interconnects.[1][9]
The collaboration sits inside a broader 2025 race among AI developers to secure both chips and the electrical capacity to run them. Designing custom accelerators is capital- and talent-intensive and carries execution risk: a chip program can slip, and 10 GW of capacity depends on power, manufacturing, and data-center construction proceeding on schedule. The companies presented the 2026-to-2029 schedule as a target rather than a guarantee, and several of OpenAI's compute commitments, including this one, were announced as long-term plans whose financing and buildout would unfold over years.[3][4] OpenAI's custom chip is built with Broadcom for design and integration and TSMC for fabrication, mirroring the standard fabless model used by other large custom-silicon customers.[6][9]