The $690B AI Race Is Quietly Killing Big Tech’s Best Asset

Warren BuffettA great business funds its growth and still hands cash to owners. Free cash flow — not revenue, not even reported profit — is the scorecard.

• Buffett’s entire framework rests on owner earnings: reported profit, plus depreciation, minus the capex truly required to maintain the business. The AI buildout is the most capex-heavy episode in a generation, which makes this adjustment decisive rather than academic. When capex runs at 90%+ of operating cash, owner earnings can be a fraction of the headline profit.
• The brutal test he applies: would this business still gush cash for its owners if it stopped expanding tomorrow? For a genuine franchise — a search monopoly, a dominant cloud, a social network with pricing power — the answer may still be yes once the spending normalizes. The question is how many years of depressed free cash flow you must underwrite first, and whether the returns on that capital ever show up.
• His warning is not "sell." It is "don’t confuse a temporarily-spending compounder with a permanently-impaired one — and don’t pay the old free-cash-flow multiple for a business that has, for now, stopped throwing off free cash flow." The premium was earned by the cash. If the cash pauses, the premium has to be re-underwritten.

Seth KlarmanMargin of safety comes from skeptical analysis of exactly the assumptions management would prefer you not examine — starting with how costs are recognized.

• Klarman lives in the footnotes, and the footnote that decides a lot here is the depreciation schedule. Over how many years is a GPU server depreciated — four, five, six? Several hyperscalers have extended assumed useful lives in recent years, which mechanically lifts reported profit by spreading the same cost over more years. The economic truth — that this hardware obsoletes brutally fast — does not change. Only the accounting does.
• That is the iceberg. The bright tip above the waterline is the reported earnings the market trades on. The dark mass below is the real, accelerating cost of replacing chips that may be outdated in a few years, and the gap between depreciation charged and cash actually spent. When the cash going out vastly exceeds the cost flowing through the income statement, "profit" and economic reality diverge.
• His response is never "avoid the sector." It is "widen the discount you demand." Where numbers lean on generous useful-life assumptions and self-reinforcing demand, uncertainty is not a reason to be more careful at full price — it is a reason to require a much lower one. A margin of safety is what lets you survive being wrong about the part of the iceberg you couldn’t see.

Peter LynchGrowth is only worth paying for at a price. Compare the multiple to the durable, self-funded growth rate — not to the excitement.

• Lynch’s discipline is almost rude in its simplicity: what are you paying, and for what growth? A rich multiple can be perfectly rational for a business compounding cash quickly and cheaply. It is far harder to justify for one whose growth now requires it to spend nearly all of its cash flow — and possibly to keep doing so for years — just to defend its position.
• So he splits the field. Which of these AI-spenders is buying durable, high-return capacity it will own and monetize for a decade — and which is in an arms race where the spending is effectively a toll it must pay every year, with the returns accruing mostly to the chip supplier? The first deserves its multiple. The second is paying a growth premium for returns it is, in effect, renting.
• The practical Lynch test for today: take the company’s reported earnings growth, then ask how much of it survives once you charge real, fast depreciation and the cash actually leaving the building. If the growth that justifies the multiple only exists before those adjustments, you are not buying a compounder at a fair price — you are buying a story at a generous one.