Broadcom AI Chip Revenue Triples as Custom Accelerator Demand Surges

Week of August 31, 2026

Broadcom Sees AI Revenue Reaching About $230 Billion by Fiscal 2028

Custom accelerator demand from Google, Anthropic, OpenAI and Meta underpins Broadcom's AI revenue path, which depends on data centers being ready on time.

View the data

	AI revenue (projected)
FY2026	$58B
FY2027	$115B
FY2028	$230B

More charts in By the Numbers

Broadcom reported record results on September 2 for its fiscal third quarter, with AI chip revenue more than tripling from a year earlier. The company said demand for the custom accelerators it designs for Google, Anthropic, OpenAI and Meta keeps climbing. For enterprise technology leaders, the results say two things: more of the AI capacity they rent will run on purpose-built chips rather than general-purpose GPUs, and the bottleneck on that capacity is now power and buildings as much as silicon.

What did Broadcom report for Q3 fiscal 2026?

Revenue for the quarter ended August 2 was $29.6 billion, up 86% from a year earlier, according to Broadcom's earnings release filed with the SEC. AI semiconductor revenue reached $16.7 billion, up 221% year over year and 54% from the prior quarter, the release said, which means AI chips now account for more than half of the company's total sales.

CEO Hock Tan said in the release that demand for "custom AI accelerators and networking continues to be very strong." The company guided to about $34.8 billion in fourth-quarter revenue, up 93%, with roughly $21.7 billion of that from AI semiconductors.

Broadcom's infrastructure software segment, which is mostly VMware, grew 29% to about $8.8 billion, according to the release.

Who is buying Broadcom's custom AI chips?

On the earnings call, management described a customer list that reads like a directory of the largest AI model builders, according to a Benzinga transcript of the call and Nasdaq's summary:

• Google: Broadcom shipped Ironwood TPU v7 chips in high volume during the quarter and began production shipments of the next-generation TPU v8i.
• Anthropic: Broadcom said Anthropic is deploying about 1 gigawatt of Ironwood capacity in 2026 and expects it to add 5 gigawatts of TPU v8i in 2027, with what Tan called line of sight to another 10 gigawatts in 2028. Tan said Anthropic is on track to become Broadcom's largest custom-chip customer in 2027.
• OpenAI: Broadcom has begun shipping OpenAI's first-generation custom accelerator, with about 1.3 gigawatts of deployment planned for 2027.
• Meta: Broadcom expects production shipments of Meta's MTIA accelerator to begin in the fourth quarter.

Broadcom also laid out an AI revenue path of roughly $58 billion for fiscal 2026, about $115 billion for fiscal 2027 and about $230 billion for fiscal 2028. These are company projections that depend on customers' data centers being ready on time. Tan said demand already exceeds the 2027 outlook, and management named land, power, data center shells, leading-edge silicon, substrates and high-bandwidth memory as the constraints, ChannelWeb reported.

Why does this matter for enterprise AI buyers?

Most large companies will never buy a Broadcom accelerator directly. They will meet these chips through the cloud platforms and model APIs they already pay for. That makes the shift toward custom silicon an indirect but real factor in enterprise AI costs.

Broadcom's argument, as Tan put it on the call, is that custom accelerators can deliver better performance, cost and power efficiency for specific large language model workloads than general-purpose chips. If that holds at the scale its customers are planning, the providers behind the models enterprises use most could gain more control over their own unit costs. Whether those savings reach enterprise customers as lower prices is a separate question. At The AI Enterprise Conference 2026, Verisk Global Chief Data Officer Louis Dimodugno warned that token prices could rise once frontier model providers go public, a point covered in our story on how Verisk lets AI pay for AI.

The supply list matters too. When the limits are power, land and memory, capacity gets allocated, and large model builders with multi-gigawatt commitments sit at the front of the line. Enterprises that depend on a single provider for critical AI workloads should ask how that provider's capacity plans hold up if build-outs slip.

What about VMware and private AI?

The quieter part of the report concerns enterprises more directly. Broadcom introduced VMware Private AI Cloud, aimed at companies that want to build and run AI next to their existing applications, and said VMware Cloud Foundation is helping customers move workloads from public to private clouds, according to the call transcript.

That matches what practitioners have described at AI Connect events. Fulcrum Digital founder Rajesh Sinha told The AI Enterprise Conference 2026 audience that for many regulated clients "everything we are talking today is going to be inside your firewall," a theme we explored in seven layers of sovereign AI architecture. Everpure VP of AI Kaycee Lai made a similar case for bringing AI to the data and argued the eventual winners are not defined by hardware spend. "They're not the ones who buy the most GPUs," he said.

The VMware angle cuts both ways for buyers. Running private AI on the same platform that already hosts core applications can simplify operations and keep sensitive data in house, but it also concentrates more of the stack with one vendor. Buyers should weigh that dependency as carefully as the convenience.

What to watch next

The next signals are whether Broadcom's customers hit their 2027 deployment targets and whether hyperscalers pass custom-chip savings through to enterprise pricing. For CIOs, the practical step is to treat AI capacity as a supply-chain risk: know which chips and data centers sit behind your critical models, and keep at least one workload portable across providers or on private infrastructure.

Frequently asked questions

How much AI revenue did Broadcom report in Q3 fiscal 2026?

Broadcom reported $16.7 billion in AI semiconductor revenue for the quarter ended August 2, 2026, up 221% year over year, on total revenue of $29.6 billion.

Which companies use Broadcom custom AI chips?

According to Broadcom's earnings call, Google, Anthropic, OpenAI and Meta are among the customers deploying or preparing to deploy Broadcom-designed accelerators.

Why do custom AI chips matter to enterprises?

Most enterprises use them indirectly through cloud and model providers, so the chips affect providers' unit costs, capacity and, potentially, the prices enterprises pay for AI.

Filed under: Data & AI Infrastructure, AI Strategy & ROI