Broadcom (AVGO): what it does, how it makes money, and why it owns custom AI silicon

The standard $AVGO story is that Broadcom is a sprawling acquisition machine: a bit of chips, a bit of VMware, a fat dividend, and a CEO who buys software companies and raises prices. That story is half-right, and it misses the half that now drives the stock.

The more accurate frame: Broadcom is the second-largest AI-chip company on the market, and the largest by a different metric than the one everyone watches. It does not sell the GPU. It designs the custom silicon that $GOOGL, $META, and a short list of other hyperscalers build so they can buy fewer GPUs. In its fiscal Q2 2026 (reported June 3, 2026), Broadcom's AI semiconductor revenue hit $10.8B, up 143% year on year, with AI bookings above $30B. This piece walks through what the company actually does, how its segments make money, where it sits in the AI-compute stack, and the concentration risk underneath the growth.

Why it matters now

On June 3, 2026, Broadcom reported fiscal Q2 revenue of $22.2B, up 48% year on year, and guided fiscal Q3 to roughly $29.4B, up 84%, with AI semiconductor revenue expected near $16B (up more than 200%). The print confirmed the part of the thesis that is hardest to fake: the custom-accelerator backlog is converting into shipped silicon, and the next quarter's guide accelerates rather than decelerates. The stock still sold off on the day, a positioning reaction after a run into the print near its highs, not a break in the numbers. That gap between a strengthening business and a jumpy tape is exactly why the evergreen question, what does this company actually own, is worth answering carefully.

What does Broadcom do?

Broadcom is two companies bolted together. The larger and faster-growing one is Semiconductor Solutions; the steadier one is Infrastructure Software.

On the semiconductor side, the AI story has two legs. The first is custom accelerators, marketed as XPUs. A hyperscaler that wants its own chip (to run its own models cheaply, at its own scale) does not start from a blank wafer. It brings the architecture and hands Broadcom the hard parts: the IP, the SerDes, the packaging, the networking, and the manufacturing relationship with $TSM. Broadcom co-designs the result. The canonical example is Google's TPU, now seven generations deep and co-designed with Broadcom since 2014. $META's MTIA inference and training silicon is the second large program. The second leg is networking: the Tomahawk and Jericho ethernet switch chips that move data between thousands of accelerators inside a training cluster. Tomahawk 5 (51.2 terabits per second) is shipping; Tomahawk 6 (102.4 Tbps) is sampling.

The rest of the semiconductor segment is the legacy Broadcom: wireless RF for $AAPL handsets, broadband, storage connectivity, and enterprise networking. These are cyclical, cash-generative, and not the reason the multiple is what it is.

The software side is VMware (acquired for $69B in November 2023), plus the earlier Symantec and CA assets. CEO Hock Tan runs the same playbook each time: convert perpetual licenses to subscriptions, trim the long tail of small customers, and harvest the margin. VMware now runs at roughly $20B in annual recurring revenue with software gross margins near 90%.

How Broadcom makes money

Revenue splits across the two segments, monetized very differently: per-chip pricing on the silicon side, multi-year enterprise contracts on the software side. Trailing twelve-month revenue is $68.3B (as of 2026-06-03). Gross margin sits near 77%, operating margin near 45%, because the mix is shifting toward high-value custom silicon and high-margin software at the same time.

The contracts that define the AI leg:

• Google TPU (v5/v6/v7 co-design, multi-generation) and Meta MTIA are the two disclosed large XPU programs.
• A third large hyperscaler XPU customer was disclosed in development in September 2024, with additional customers confirmed since; management has guided the AI serviceable market to the tens of billions across multiple customers.
• A multi-year Apple wireless component extension signed in 2026 de-risks the RF baseline through fiscal 2028 and beyond.
• VMware, closed November 2023, now around $20B ARR after the subscription conversion.

The headline risk is concentration. Broadcom's top hyperscaler accelerator customers represent a large share of AI semiconductor revenue, and the wireless segment leans heavily on a single customer in Apple. When a handful of buyers drive the growth, any one of them pausing a chip program moves the guide. That is the number to hold onto through the rest of this piece.

Where it sits in the AI compute bubble

Broadcom is the primary-tier name in QA's AI Compute Accelerators bubble, the cluster of pure-play silicon that the AI capex cycle runs through. It also carries a secondary weight in the Networking / Optical bubble, which is rare: most names sit in one bucket, Broadcom genuinely spans compute and the fabric that connects it. On the AI Inference theme, the custom-XPU story is fundamentally an inference-economics bet: at hyperscaler scale, a purpose-built accelerator can deliver materially lower total cost of ownership than a general-purpose GPU for a fixed model architecture.

The useful way to place Broadcom on the stack is by what it is not. It is not the GPU ($NVDA). It is not the foundry (TSM). It is the design partner and the interconnect, the layer that lets a hyperscaler turn its own architecture into shipped silicon and then wire thousands of those chips together. Its closest peer on the custom-ASIC side is $MRVL; on the optical-interconnect side, names like $CRDO and Fabrinet ($FN) move with it. Empirically, AVGO's tightest historical correlations in the QA universe run to Celestica ($CLS, 0.66), TSM (0.60), CRDO (0.57), and NVDA (0.52), the supply chain that ships when AI capex ships.

For the deeper GPU-versus-ASIC framing, see Nvidia vs custom ASICs: TPU, Trainium, MTIA.

The numbers

The trajectory is the story. A $68B-revenue business does not usually post 48% top-line growth, and the AI segment inside it is compounding far faster than the whole. The trailing P/E (above 90) looks extreme until you account for the AI ramp and the VMware integration costs rolling off; on forward earnings the multiple compresses sharply, which is why the bull and bear arguments are really an argument about whether the AI-silicon ramp holds.

The bull case

• AI semiconductor revenue went from $10.8B in Q2 to a ~$16B guide for Q3 (+200% YoY), and AI bookings above $30B give the backlog credibility rather than a hope.
• Broadcom co-designs the majority of hyperscaler custom accelerators, a position built on multi-generation relationships (Google since 2014) that are expensive and slow for a customer to unwind.
• The networking leg compounds the silicon leg: Tomahawk and Jericho ethernet win share against InfiniBand inside scale-out fabrics, and the customer buying the XPU often buys the switch.
• Software is a margin flywheel: VMware at ~90% gross margin and improving free-cash-flow conversion funds the dividend and buyback that put a floor under the stock.
• The Apple wireless extension de-risks the cyclical RF baseline through fiscal 2028, removing one of the older bear arguments.

The bear case

• Concentration. A handful of hyperscalers drive the AI semiconductor line; any single TPU or MTIA pause hits the guide hard, and the wireless segment is a single-customer story in a cyclical handset market.
• The custom-ASIC moat is narrower than Nvidia's software moat. $MRVL competes for the same hyperscaler sockets, and a hyperscaler can run a dual-sourced design over time.
• VMware revenue is still declining in some segments as the long tail of customers is trimmed; the margin harvest and the revenue line can pull in opposite directions.
• Valuation prices a clean ramp. A price-to-sales ratio above 30 and a trailing P/E above 90 leave little room for a missed quarter or a delayed customer chip.
• Hyperscaler capex is the whole cycle. If AI infrastructure spend slows, the custom-silicon order book is one of the first places it shows up.

How to access

Broadcom trades on the Nasdaq as AVGO, a clean US listing, so direct ownership needs nothing exotic. To trade it from a US-retail account alongside the rest of the AI-supply-chain names, see /stack/ibkr.

Most diversified investors already own Broadcom without buying the single stock. It is a top holding in the VanEck Semiconductor ETF (SMH, roughly 8%), and a meaningful weight in broad tech and market funds: XLK (~6%), IWF (~6%), VGT (~5%), and even SPY (~3%). If you hold a US large-cap index fund, Broadcom is already one of your largest AI-chip exposures, you just may not have known its weight.

Bubble shifts and rule-based alerts on $AVGO, the kind that fire when a name crosses a curated level or its bubble correlation breaks, are part of /pro.

What to watch

• Next earnings: fiscal Q3 2026, expected around 2026-09-03. The AI semiconductor line versus the ~$16B guide is the single cleanest read.
• Whether AI bookings keep building above the $30B mark, or flatten, the leading indicator for the FY2027 revenue path.
• Identity and ramp of the additional XPU customers beyond Google and Meta; each confirmed program widens the base and reduces single-customer risk.
• VMware: whether ARR growth and the revenue line converge or keep diverging as the customer trim runs its course.
• A bubble-level shift: if the AI Compute Accelerators bubble breaks its correlation with NVDA and the broader capex tape, the structural read on $AVGO changes from "rides the cycle" to "stands on its own backlog."

Live data on this ticker: /stocks/avgo. Price, ETF holdings, bubble correlation, bot positions.

Bubble context: /bubbles/semiconductors. The cluster this name belongs to and how it's moving.

QuantAbundance is educational research. Nothing here is investment advice. See /disclosures.