Broadcom’s AI chips are custom accelerators, which Broadcom calls XPUs, built with and for a specific customer such as Google, Meta or OpenAI, plus the Ethernet networking that ties them together. How the design work is split depends on the customer: OpenAI says it designs its own accelerator, while Broadcom says it develops and supplies Google’s TPUs. Broadcom brings design IP, packaging, networking and supply, and its advanced packaged designs are built on TSMC process nodes. As of October 2026 this business is large enough that Broadcom reports “AI semiconductor revenue” separately, at $16.7 billion in its latest quarter.

On this page

What an XPU is

“XPU” is Broadcom’s umbrella term for a custom AI accelerator, as opposed to a merchant GPU that any buyer can order from a catalog. When Broadcom announced its custom-accelerator push in March 2024, it described “foundational technologies and advanced packaging capabilities aimed at building the highest performance, lowest power custom AI accelerators,” naming its SerDes and DSP technology as the base.

Packaging is a large part of that offer. In December 2024 Broadcom introduced its 3.5D XDSiP platform, which combines 3D silicon stacking with 2.5D packaging and can integrate more than 6,000 mm² of silicon and up to 12 HBM stacks in one package. Its face-to-face (F2F) method connects the top metal layers of stacked dies directly. Broadcom said the lead 3.5D XPU uses TSMC process nodes and TSMC’s CoWoS packaging, that more than five 3.5D products were in development, and that production shipments would start in February 2026. Fujitsu named its FUJITSU-MONAKA processor, on a 2-nanometer process, as one design using it.

Who designs what

The split of work differs by customer, and the companies describe it in their own words.

  • OpenAI said on October 13, 2025 that it designs the accelerators and systems, while Broadcom develops and deploys them. The plan covers 10 gigawatts of OpenAI-designed accelerators in racks that use Broadcom’s Ethernet, PCIe and optical connectivity, with deployment starting in the second half of 2026 and finishing by the end of 2029.
  • Google: in a Form 8-K dated April 6, 2026, Broadcom said it will “develop and supply custom Tensor Processing Units (‘TPUs’) for Google’s future generations of TPUs,” along with networking and other components, through 2031.
  • Meta announced an extended partnership on April 14, 2026 covering future generations of its MTIA (Meta Training and Inference Accelerator) chips through 2029, including chip design, packaging and Ethernet networking. The companies called one of the chips “the industry’s first 2nm AI compute accelerator” and said the initial commitment exceeds 1 GW, with multiple gigawatts planned over time.
  • Anthropic’s arrangement is for TPU capacity. On April 6, 2026 Anthropic announced “multiple gigawatts of next-generation TPU capacity” with Google and Broadcom, expected to come online from 2027. Broadcom’s 8-K puts Anthropic’s access through Broadcom at about 3.5 GW of TPU-based compute beginning in 2027, and adds that consumption “is dependent on Anthropic’s continued commercial success.”
Customer Chip family What the companies announced Date
Google TPU Broadcom to develop and supply future TPU generations through 2031 Apr 6, 2026
Anthropic TPU capacity About 3.5 GW of TPU-based compute via Broadcom from 2027 Apr 6, 2026
Meta MTIA Multi-generation partnership through 2029; initial commitment over 1 GW Apr 14, 2026
OpenAI OpenAI-designed accelerator 10 GW deployed from H2 2026 to end of 2029 Oct 13, 2025
Fujitsu FUJITSU-MONAKA Uses Broadcom’s 3.5D XDSiP packaging on a 2 nm process Dec 5, 2024

The AI chips catalog lists the TPU, MTIA and other accelerators with their vendor-published specifications.

How big the business is

Broadcom reports AI revenue as a company fact in its quarterly release. For fiscal Q3 2026, which ended August 2, 2026 and was reported on September 2, 2026:

Metric (fiscal Q3 2026) Figure
Total revenue $29.6 billion
Semiconductor solutions revenue $20.839 billion
AI semiconductor revenue $16.7 billion
AI semiconductor growth, year over year 221%
AI semiconductor growth, quarter over quarter 54%
Q4 FY2026 AI semiconductor revenue guidance $21.7 billion

CEO Hock Tan said in the release: “Demand for our custom AI accelerators and networking continues to be very strong.”

Financing the build-out

Broadcom has also set up a vehicle to finance the capacity its chips go into. On June 9, 2026, Broadcom, Apollo and Blackstone announced the “AI XPV Platform,” intended to support more than 20 GW of AI compute through 2028 using “Broadcom’s XPUs and networking solutions customized for leading frontier AI labs,” naming Anthropic and OpenAI. An initial $35 billion tranche, led by Apollo with Blackstone, supports Anthropic’s previously announced expansion of more than 1 GW, deploying from mid-2026 at Fluidstack-based sites. Projects of this kind are tracked in the AI data centers tracker.

Marvell, the other custom-silicon house

Marvell also designs custom AI silicon for cloud providers. On December 2, 2024 it announced a five-year, multi-generational agreement with AWS covering “custom AI products” alongside optical DSPs, AEC DSPs, PCIe retimers, data-center interconnect modules and Ethernet switching silicon. In May 2025 Marvell and Nvidia said Marvell’s custom silicon would be offered with Nvidia’s NVLink Fusion, which Nvidia describes as a way of “integrating custom XPU silicon with NVIDIA NVLink connectivity.”

For its fiscal Q2 2027, reported August 27, 2026, Marvell gave net revenue of $2.739 billion and data center revenue of $2.171 billion. Marvell’s CEO pointed to “a significant acceleration in our Custom business beginning in the second half of fiscal 2027.” Company profiles for Broadcom, Marvell and their customers are in the AI companies directory.

Why labs commission custom chips

The companies’ own statements give the stated reasons. Mark Zuckerberg said the Broadcom partnership “will give us greater performance and efficiency” as Meta rolls out more than 1 GW of its custom silicon. Sam Altman said that “developing our own accelerators adds to the broader ecosystem of partners all building the capacity required to push the frontier of AI.” Anthropic says it trains and runs Claude on AWS Trainium, Google TPUs and Nvidia GPUs, and describes this as a diversified hardware strategy.

What custom silicon does not remove is the shared supply chain: Broadcom’s 3.5D designs rely on TSMC process nodes and CoWoS packaging, and those packages carry HBM stacks made by the memory makers.

Sources