Broadcom is the indispensable backbone of AI infrastructure — that’s what Wall Street keeps telling itself. The stock slid from $343.50 in January to $314.40 by March 2026, and the people defending it keep calling it a buying opportunity. Or the slide is the market slowly working out something it hasn’t been willing to say directly.
Broadcom’s entire AI thesis rests on a handful of hyperscalers continuing to pay a premium for custom silicon they are increasingly capable of designing themselves. Google, Broadcom’s anchor AI customer, has been building its own Tensor Processing Unit architecture for years. The question isn’t whether hyperscaler insourcing is happening. It’s how fast, and whether Broadcom’s current valuation has absorbed any of that velocity.
Fiscal 2025 revenue reached $63.9 billion, per Broadcom’s annual report, up from $51.6 billion the year before. Operating income nearly doubled, from $13.5 billion to $25.5 billion. Those are the numbers of a company at peak leverage — exactly when the structural questions deserve the most scrutiny.
What the Operating Income Number Actually Says
The $25.5 billion operating income figure implies a margin of roughly 40%, up from about 26% in fiscal 2024. That jump demands explanation beyond top-line growth. VMware, acquired in late 2023, finally cycled through the income statement with its full software margin contribution. Operating leverage on AI chip demand did the rest. But a 14-percentage-point margin expansion in a single year, in a capital-intensive industry, either reflects a genuine competitive moat or a temporary window of pricing power. If this margin compresses even 500 basis points as hyperscaler bargaining power increases — and it will increase, as customers accumulate internal design capability — roughly $3 billion in annual operating income evaporates without a single dollar of revenue loss.
R&D spending came in at $11 billion for fiscal 2025, approximately 17.2% of revenue, down slightly as a share from 18% the prior year. The absolute number is large. The direction is ambiguous. Broadcom is spending more in dollar terms to defend a design business that depends entirely on customers choosing not to compete with it.
Capex remains almost comically light — $0.6 billion against $63.9 billion in revenue, a ratio of roughly 1%. That’s the fabless model: someone else builds the fabs. It also means Broadcom’s control over its own supply chain is narrow. When TSMC capacity gets contested, Broadcom waits in line like everyone else.
The Customer That Designs Its Own Replacement
Google’s TPU program is not a rumor or a distant threat. It is a multi-generational, multi-billion-dollar internal semiconductor effort that has been compounding for nearly a decade. Every generation of TPU that performs competitively feeds an internal ROI calculation that eventually asks: why are we still paying Broadcom? The relationship works because Broadcom’s custom ASIC design expertise is genuinely scarce. That scarcity is time-limited. Hyperscaler engineering teams are accumulating the institutional knowledge required to design at Broadcom’s level, slowly and deliberately, inside the walls of its own customers.
Concentration risk bites differently than it looks on the surface. Not a cliff — one customer leaves, revenue craters. A gradient. Google keeps Broadcom for next-generation designs while running more workloads on internal silicon. Revenue stays. Margin negotiating leverage transfers. Broadcom’s pricing power, the actual source of that 40% operating margin, erodes in increments that don’t surface in quarterly filings until they’re structural.
A version of this story exists where none of that happens. Hyperscaler AI compute demand grows fast enough that even with aggressive insourcing, the total addressable market expands faster than any one customer can absorb internally. Broadcom adds new hyperscaler relationships — Meta, Amazon, Microsoft — and the concentration risk dilutes organically. CEO Hock Tan’s $100 billion AI revenue target might actually be conservative if inference compute scales the way training compute did. The weakest assumption in the bear case is that Google’s internal design capability reaches production-grade maturity on the timeline consensus fears — a bet against the stickiness of Broadcom’s accumulated design IP. The thesis breaks if AI infrastructure spending plateaus before Broadcom has diversified its customer base, if Google reaches that maturity faster than expected, or if a second-generation custom chip competitor captures the next wave of hyperscaler procurement. Any one of those conditions is plausible. All three arriving in the same 18-month window would reprice AVGO aggressively.
Nvidia sits at $118.9 billion in fiscal 2025 revenue, per its earnings filings, and owns the ecosystem — CUDA, the developer toolchain, the inference stack. AMD reported $24.8 billion and competes on merchant silicon. Broadcom sells bespoke. No standard product, no public roadmap, no ecosystem moat in the traditional sense. The moat is relationship depth and design expertise. Against Nvidia’s ecosystem lock-in, that’s a narrower competitive position than the revenue numbers suggest. The comparison that matters isn’t Broadcom versus Nvidia on revenue scale. It’s Broadcom versus Google’s internal TPU team on design quality — a competition with no public scoreboard.
Co-founder Henry Samueli’s selling at these levels doesn’t prove a thesis. Executives sell for any number of reasons. But the data point lands next to a stock that has given back roughly 8.5% since January, in a period when AI infrastructure optimism has generally remained elevated. The non-alignment between external narrative and internal action is a texture worth registering.
Broadcom’s fiscal 2025 performance is, on its face, exceptional. The contrarian case isn’t that the company is broken. It’s that the valuation prices in a future where the customer relationship stays exactly as it is — where Google keeps outsourcing, keeps paying current margins, keeps choosing Broadcom over its own engineers. That future is possible. It is not the most likely version of how concentrated, technically sophisticated customers behave over a decade-long time horizon. Cyclical growth is already priced in; structural margin erosion is not.
The most uncomfortable thing about Broadcom’s position isn’t the risk — it’s that the risk comes dressed as the strength.