Broadcom's custom AI silicon generated $16.7 billion in revenue during its fiscal third quarter, and CEO Hock Tan told investors the growth is not slowing down.

Semiconductor solutions climbed 127% to $20.8 billion

Broadcom reported revenue of $29.6 billion for the quarter ended August 2, an increase of 86% from a year earlier, according to the company's earnings release.

The semiconductor solutions segment grew 127% to $20.8 billion, while infrastructure software rose 29% to $8.75 billion.

AI semiconductor sales reached $16.7 billion, up 221% from the same period in 2025 and up 54% from the prior quarter, Tan said. He described demand for the company's custom accelerators and networking gear as "very strong."

Broadcom generated $13.7 billion in free cash flow, roughly 46% of revenue, with a non-GAAP operating margin of 66%.

In the earnings release, Tan said he expects AI semiconductor revenue to climb to $21.7 billion in the fourth quarter — a 236% rise from the same period a year ago.

CFO Amie Thuener said the company's fourth-quarter revenue guidance is approximately $34.8 billion, up 93% year over year. Broadcom expects its non-GAAP operating margin to hold steady at 66%.

Google's custom silicon deal runs through 2031

The board declared a quarterly dividend of $0.65 per share. Broadcom's fiscal fourth quarter ends November 1, and the company cautioned that actual results may differ materially from its outlook.

The demand comes from a short list of named clients that are building their own chips to reduce dependence on Nvidia GPUs. For hyperscalers spending tens of billions of dollars on AI infrastructure, designing their own accelerators has become a way to control both cost and supply, since custom silicon is tuned to their specific models rather than to general-purpose workloads.

In April, Broadcom signed a long-term agreement with Alphabet's Google to supply custom AI chips through 2031, extending a TPU partnership that has been in place for roughly 10 years, as Cryptopolitan reported.

The same period brought a deal to supply Anthropic with 3.5 gigawatts of Google-powered computing capacity starting in 2027, a commitment that could approach 5 gigawatts over time.

In June, ChatGPT's maker unveiled "Jalapeño," an inference chip designed and built with Broadcom. It is the first component of a larger hardware platform, according to both companies.

OpenAI president Greg Brockman framed the move as a way to "serve more intelligence with greater efficiency." The two companies set a target of eventually powering roughly 10 gigawatts of capacity.

Broadcom's strategy centers on the ASIC, an application-specific chip that is less flexible than a general-purpose GPU but cheaper and tuned to specific AI workloads. That tradeoff has found a market in inference — the work of running trained models for end users — which is growing as AI products move from development into widespread daily use.

Nvidia's lead in inference is not as dominant as it is in training, and none of Broadcom's customers is abandoning Nvidia. OpenAI's Sam Altman has publicly said Nvidia makes "the best AI chips in the world," and Anthropic runs Claude on AWS Trainium, Google TPUs, and Nvidia GPUs to distribute its risk. How quickly the multi-gigawatt commitments with Google, Anthropic, and OpenAI translate into shipped silicon will become clearer as Broadcom reports its fiscal fourth quarter, which ends November 1.