There’s a story Warren Buffett has told for decades, in one form or another: he doesn’t buy businesses; he buys moats. A moat is the thing that lets you raise your price next year and keep your customers anyway, the toll bridge nobody can build next to, the brand nobody can fake, the manufacturing process nobody else has figured out. Everything else, he says, is just a business borrowing time.

Right now, the AI industry is being priced by markets as if the whole stack has a moat. It doesn’t. And the clearest way to see the difference between businesses with real pricing power and those renting the appearance of it is an old, unglamorous analogy: a farmer, a landlord, and a tractor company that all claim to be part of the same investment.

Picture a valley where one landlord owns the only fertile soil for a hundred miles. Every farmer who wants to grow anything has to lease from him. He doesn’t need to advertise, doesn’t need to compete, doesn’t need to worry that a cheaper field will open up next door; there is no next door. So when demand for wheat rises, he doesn’t just collect more rent. He raises the rent itself every season, and the farmers pay it because the alternative is not farming at all.

In the first quarter of 2026, TSMC’s own disclosures show high-performance computing and the AI-related segment climbing to 61% of total revenue, up from 55% the quarter before and just 51% a year earlier. This isn’t a company hoping to catch a wave; it’s a company that already owns the harbor. Chips at 7-nanometer or smaller now account for 74% of revenue, with the most advanced 3-nanometer node alone making up a quarter of sales gross margin: 66.2%.

And here is the part that separates a moat from a lucky quarter: TSMC is raising prices. Advanced-node pricing is going up 3 to 10% starting this year, with reports of increases as steep as 15% on 3nm chips in the back half of 2026. This is not a company passing through costs reluctantly. It is a company that knows its customers have nowhere else to go and is pricing accordingly, with capital expenditure guidance jumping to $52–56 billion for the year to keep the fields under cultivation. One layer up, the memory makers are doing the same thing: Samsung and SK Hynix pushed HBM3E prices up roughly 20% for 2026 contracts unusual enough that industry press flagged it as a departure from normal cycles.

This is what a moat looks like in a financial statement. Disclosed, audited, repeatable, and getting wider.

Now picture the farmer. He leases the land. He leases the tractor. He buys the seed. And when harvest comes, he has to sell his wheat into a market where a dozen other farmers, some of them foreign and heavily subsidized, are growing the same wheat and racing each other to the bottom on price. If he raises his price by even a few percent, buyers simply walk to the next stall. He has no moat; wheat is wheat, so the only lever he has left is to farm more efficiently: better seed, better timing, tighter margins on his own labor.

OpenAI and Anthropic have both gotten genuinely better at the mechanics of running their models. Anthropic's reported inference-only margins for Anthropic have reportedly climbed into the 70–80% range, according to SemiAnalysis-sourced reporting, a dramatic improvement from deeply negative margins just two years earlier. That’s real, and it matters. But look at where that improvement is coming from: efficiency, not pricing power. Token prices across the industry have been falling, not rising, because the moment one lab tries to charge more, a competitor DeepSeek, an open-weight model, or whoever’s newest offers the same capability for less. The wheat is a commodity, and everyone knows it.

That gap between the labs’ rosy inference margins and their much thinner all-in economics tells the real story. Anthropic’s overall gross margin the number that includes the cost of actually training the next model, not just serving the current one, sat closer to 40% in 2025. OpenAI is projected to burn something like $14 billion in 2026, with breakeven not expected until 2029 or 2030 at the earliest. And a meaningful chunk of the customer base isn’t paying at all: roughly 95% of OpenAI’s roughly 900 million weekly users use it for free, with about a third of all inference spend serving people who generate zero revenue.

This is a business efficient enough to survive, but not powerful enough to set its own price. That is precisely the condition Buffett spent a career avoiding.

Here’s where the analogy earns its keep, because there’s a third character in this story, and he’s the strangest one.

Imagine the landlord doesn’t just rent out land. He also owns the tractor company. And when the farmer can’t afford this year’s rent, the landlord hands him a loan on the condition that the farmer spends it buying a new tractor from the landlord’s other business. The farmer’s books now show fresh capital and new equipment. The landlord’s books now show new revenue and a healthy loan asset. Money has moved. Paperwork has been generated. But look closely, and it’s the same dollar, cycling through two arms of the same owner, dressed up to look like two separate transactions proving two separate demands.

This is, almost exactly, the structure that’s drawn the label “circular financing” in the actual AI industry. Nvidia committed up to $100 billion to OpenAI, and OpenAI’s own CFO, Sarah Friar, has said publicly that “most of the money will go back to Nvidia” because that capital is earmarked to buy Nvidia chips. Nvidia has done a version of this with CoreWeave too: an equity stake, paired with a $6.3 billion commitment for Nvidia itself to buy any cloud capacity CoreWeave can’t otherwise sell. OpenAI, separately, signed a $300 billion cloud deal with Oracle, and Oracle turned around and spent billions of that on Nvidia chips for the very data centers built to serve OpenAI. OpenAI also became one of AMD’s largest shareholders around the same time it committed to buying tens of billions of dollars of AMD chips.

None of this is fabricated or fraudulent — it’s publicly reported, and reasonable people disagree about how alarming it is. Some economists call it ordinary vendor financing, the kind of financing that capital-intensive industries have always used to help customers afford big-ticket purchases. Others call it round-tripping, a way of making demand look organic and independent when it’s really the same handful of players financing each other’s revenue statements. What’s not in dispute is the mechanism: cash is moving, but it’s moving in a loop between a small number of counterparties who all have an interest in the loop continuing to look like growth.

TSMC, notably, sits one layer outside most of this. It gets paid in real cash by whoever shows up — Nvidia, Samsung, Apple, the hyperscalers, regardless of how those buyers financed their own orders. That distance is part of why its numbers hold up to more scrutiny than the labs’ do.

Strip away the “AI” label and just ask the Buffett question of each business in this chain: if a competitor opened up next door tomorrow with unlimited capital, who would still be standing in five years, still able to raise their price?

TSMC would. There is no unlimited-capital competitor who can simply decide to match a decade of yield engineering at the 3-nanometer node. That’s a moat, and the market is pricing it, mostly, like one.

The AI labs are a harder case. Their product is replicable in months, their pricing is capped by whoever’s cheapest, and a real share of the demand propping up their valuations is being financed, in part, by their own suppliers. That’s not a moat. That’s a very well-funded farm, on rented land, paying rent with a loan from the landlord’s tractor company and hoping the harvest comes in before anyone asks too many questions about whose money is whose.

The chips will keep getting more expensive to make because the maker has a real moat and knows it. The question this whole industry still hasn’t answered is whether the businesses buying all those chips actually have one too or whether they’re just very good tenants, for now, on land that was never really theirs.