The AI Infrastructure 'Wide Moat' Is Three Moats, Paid for Three Different Ways

A moat, used the way Warren Buffett means it, is a profit that endures without being re-dug every year. Slap that label on "AI infrastructure" and it stops meaning one thing. In the same breath that includes Nvidia — a $5.3 trillion company that keeps about 74 cents of every sales dollar while spending almost nothing on factories — it includes Microsoft, which pours triple-digit billions into data centers and, in the trailing year, collected less free cash than it did a year earlier. One label, two opposite economics.

That confusion is exactly what the phrase is there to pump past you. "Wide moat" is doing rhetorical work that the accounting doesn't support: in this stack the size of the margin and the amount of capital needed to keep it do not travel together. Strip the stack apart and what you find is not one moat but three, each built and paid for differently — and only one of them looks like the durable franchise the label promises.

The design layer: a real margin, priced at its erosion

Nvidia is the name-brand case, and its numbers are the least ambiguous. Revenue grew 83% year over year, gross margin sat near 74%, and return on invested capital near 87%. Because Nvidia is fabless — it designs, TSMC fabricates — it barely has to spend to stay in the race: roughly $7 billion of capital spending against $127 billion of free cash flow in the trailing twelve months. On paper these are the most Buffett-shaped economics in the group: a high, hard-won margin with almost no reinvestment required.

The margin is real. Whether it is a moat or a treadmill is the live question. Nvidia's 5–6x gap between what a Blackwell GPU costs to build — around $6,400, of which HBM memory is roughly 45% — and the $30,000–$40,000 it sells for is captured entirely at the design layer, and the squeeze it represents is why the customers at the bottom of the stack, the Dell and Supermicro integrators and the CoreWeave renters, live on single-digit margins. What protects that spread is not a patent but the CUDA software ecosystem and a yearly product cadence that force customers to re-buy before they can switch. And the workload most vulnerable to that lock-in is inference — the part of AI that actually runs the models, now something like two-thirds of compute spend — where custom silicon engineered for a single model beats a general-purpose GPU on price per token. That is why Google, Amazon, Microsoft, and Meta are each building their own accelerators, and why headlines project Nvidia's hyperscaler-internal inference share collapsing from around 90% toward 20–30% by 2028.

This is not a challenger beating a collapsing incumbent — Nvidia is not collapsing. It is the richest customers in the world paying to build their own exit door, and the market is pricing the erosion, not the permanence.

The shovel seller: Broadcom profits from the escape

The exit door is being built by Broadcom, which designs the custom accelerators for exactly the customers trying to leave. It reported $8.4 billion of AI semiconductor revenue in its first fiscal quarter of 2026, up 106% year over year, guided the next quarter to $10.7 billion, and has announced a $100 billion fiscal-2027 AI revenue target backed by a $73 billion committed order backlog. Like Nvidia, Broadcom is fabless and rides the same foundry, and like Nvidia it collects most of the margin without owning a factory. The difference is that its moat runs in the opposite direction: it monetizes the erosion of a rival's share.

Broadcom's own durability does not inherit Nvidia's. The custom-silicon engine serves a handful of giant named customers, so each deal — OpenAI, or Google's TPU supply agreement running to 2031 — is a repricing event rather than a naturally widening franchise. What protects Broadcom is not an ecosystem but being the only design partner at this scale, which is a narrower and more cancellable kind of moat than the label implies, even if the market now prices it at roughly 45 times trailing earnings. Its real lesson is the deeper one: the "challenger" here looks strong mainly because very rich, very solvent incumbents are funding their own escape — which is a statement about the customers' incentives, not about any single product's genius.

The application layer: a real moat, spent into silicon

Pivot to Microsoft, which owns the one franchise in this group that most closely resembles what Buffett actually buys — Windows, Office, Azure, a durable pricing power you do not have to rebuild each year. The AI bet is what threatens that. Staying in the race demands real estate and GPUs at staggering scale: Microsoft has guided calendar-2026 capital spending to roughly $175 billion, and its trailing-year capital expenditure is already about $116 billion — a build so large that its trailing free cash flow fell roughly 6% from a year earlier. Around two-thirds of that build goes into short-lived assets — the GPUs and CPUs themselves — which must be replaced and re-bought on a yearly cycle as the silicon turns over.

This is the inversion of the Buffett template. The durable franchise is being dug into hardware, not defended by it, because the premise of generative AI is that the software making Microsoft sticky gets commoditized and re-won at the silicon layer. Here the moat funds the dig; the dig does not fund a moat. That is the opposite of an enduring business that prints cash without shovels.

The name nobody puts on the moat

And then there is TSMC. The one layer that is hardest to contest — the leading-edge foundry every GPU and every custom accelerator depends on — is the one the "wide moat" listicles almost never lead with, because it is boring, capital-hungry, and its margin is merely very good rather than obscene, rated at only about 33 times earnings. But TSMC's scale economics, its sunk process R&D, and its near-monopoly on the leading edge — a genuine structural scarcity rather than a marketing gap — make it the most compounding, least cancellable franchise in the entire stack. The moat that must be re-dug constantly is precisely the one the label overlooks, because its price tells no dramatic story.

The label, then, is doing work the numbers don't. Nvidia's margin is real but funded by switching costs that are being priced at their erosion, not at permanence. Microsoft's moat is genuine but is being spent down into silicon. Broadcom profits from arbitraging the first against the second. And the moat that most closely fits Buffett's actual definition — TSMC's — is the one the narrative forgets, because durability without drama doesn't sell. The correction is not "AI has no moat." It is that there are several, they are different things, they are paid for in different ways, and two of them are being sold as one.