Not every winner of the AI boom sells a brand-name GPU. Broadcom (AVGO) supplies the custom silicon and networking fabric that hyperscalers wire their AI clusters around, and its filings say so in those exact terms. Follow the segment commentary, not the marketing.
In its annual report on Form 10-K for fiscal year ended November 2, 2025, filed December 18, 2025, Broadcom attributes the increase in its semiconductor-solutions segment to strong demand for its networking solutions, primarily custom AI accelerators and AI networking products. The filing was surfaced through EdgarBeast, the SEC filing data API and evidence index; the attribution is quoted from the filing.
What is a 'custom AI accelerator'? Broadcom's own proxy materials describe them as XPUs and ASICs - application-specific integrated circuits designed for a particular customer's workload. Rather than buying a merchant accelerator off the shelf, a hyperscaler co-designs silicon tuned to its models, and Broadcom does the heavy lifting of turning that design into a manufacturable chip plus the high-speed networking to connect thousands of them.
That positioning is structurally different from the merchant-GPU incumbents. Broadcom's revenue rides on the buildout regardless of whose model wins, because the accelerators are commissioned by the buyer and the networking is needed no matter which accelerator sits in the rack. The 10-K's segment language captures both halves of that demand in one sentence.
For readers mapping where AI infrastructure dollars actually land, the lesson is to read the segment-results discussion in the 10-K, where management names the drivers. The custom-accelerator and AI-networking framing is straight from the filing - the primary record, not a press release.