Half of understanding the AI supply chain is decoding the vocabulary, and the most precise place to find a company's own definitions is often the proxy statement, where the business gets described for shareholders. Broadcom (AVGO) lays out its AI-silicon terms there.

In its 2026 proxy statement on Schedule 14A filed March 2, 2026, Broadcom describes its offerings as including, but not limited to, custom AI accelerators, XPUs, ASICs, networking and connectivity solutions, and AI racks or systems based on its XPUs. The filing was surfaced through SEC filings, the SEC filing data API and evidence index, and the enumeration is quoted from the document.

Walk the anatomy. An ASIC - application-specific integrated circuit - is silicon hardwired for one job rather than general-purpose compute. An XPU, in Broadcom's usage, is a custom AI accelerator of exactly that kind: a chip co-designed with a hyperscaler to run that customer's models efficiently, as opposed to a merchant GPU sold to everyone. The proxy treats 'XPU' and 'custom AI accelerator' as the same category - it pairs the two terms directly, describing the ability “to provide custom AI accelerators or XPUs, connectivity solutions, and racks or systems based on our XPUs.” That “or” is the glossary entry: the company is telling shareholders the two words name one thing.

Where the enumeration appears matters as much as its wording. It is not in a marketing section; it sits in the compensation discussion, defining a performance metric. The proxy specifies the “technologies and products used or expected to be used by our customers for AI enablement, including but not limited to our custom AI accelerators, XPUs, ASICs, networking and connectivity solutions, and AI racks or systems based on our XPUs” - and then labels that whole bundle “AI Revenue,” the metric chosen for the 2025 Tan PSU (performance stock unit) award. In other words, Broadcom's own definition of what counts as its AI business is fixed in a signed filing because executive pay is tied to it. That gives the list unusual precision: it is the boundary the board drew around “AI” for the purpose of paying its CEO.

The phrase 'AI racks or systems based on its XPUs' extends the picture upward. Broadcom is describing not just selling a chip but the networking and the rack-scale system that wires many XPUs together - the connectivity layer being as essential to a cluster as the accelerators themselves. That is why networking and connectivity sit in the same list as the silicon. The proxy frames the rationale plainly: AI Revenue was selected as the metric because, in the filing's words, the company “must successfully compete in a highly competitive technology industry” that is rapidly evolving - the board betting compensation on the whole stack from custom silicon to system.

The same filing quantifies why the category earned its own metric. The proxy reports that AI semiconductor revenue grew 65% from $12.2 billion in fiscal 2024 to $20.2 billion in fiscal 2025, and that this AI semiconductor revenue represented 55% of fiscal 2025 semiconductor revenue of $36.9 billion - attributing the cash-flow and revenue gains to “the increase in our custom accelerator business.” Set against the vocabulary list, those figures show the XPU/ASIC line is not a side project: more than half of Broadcom's semiconductor revenue now flows from the custom-accelerator category the glossary defines. The definitions and the dollars are in the same document, reinforcing each other.

There is a second reason the proxy's vocabulary is unusually trustworthy: the same document records the executives whose work it describes in these terms. In the named-executive summaries, the filing credits leadership with implementing “expansion of existing and new business models and strategies to maximize our ability to provide custom AI accelerators or XPUs, connectivity solutions, and racks or systems based on our XPUs to meet the unprecedented demand for compute and AI infrastructure,” and with leading “key strategic customer engagements” around AI semiconductor solutions. The phrasing is consistent from the compensation rationale to the individual performance write-ups - the company uses the same words for the same things throughout, which is exactly what makes a proxy a better glossary than a press release: internal consistency under the legal weight of a signed SEC filing.

For the careful reader, the distinction the list draws is the useful part. “Custom AI accelerators or XPUs” and “ASICs” describe the silicon; “networking and connectivity solutions” describe the fabric between chips; and “AI racks or systems based on our XPUs” describe the assembled cluster. Broadcom is positioning itself across all three layers, and the proxy's choice to define AI Revenue as the union of them - rather than chips alone - is itself a statement that the company sees its AI franchise as the full stack from accelerator to system. The 65%-to-$20.2-billion jump, at 55% of semiconductor revenue, is the financial weight behind that framing.

For readers building a mental model of who makes what in AI infrastructure, the proxy's glossary is a clean primary source: the terms are Broadcom's own, in a signed filing, pinned to a pay metric, a revenue figure, and the executives accountable for them. Read the offerings list and the AI Revenue definition in the DEF 14A directly rather than inferring the vocabulary from secondary coverage - the document not only names the parts (custom accelerators, XPUs, ASICs, connectivity, racks) but tells you, in dollars, how much of the company they now are.