A signal is worth reading only when it forms independently of the party reporting it. When that independence cannot be confirmed from outside, the number can be perfectly accurate and still tell you nothing.
In the third quarter of 2025, Amazon reported a $9.5 billion pre-tax gain that came from neither selling goods nor selling cloud services. It came from a markup. Anthropic, the AI company in which Amazon holds a large minority stake, had raised money at a higher valuation, and accounting rules let Amazon record the appreciation of its stake as income. No cash changed hands. A funding round elsewhere lifted a number on Amazon's own statement by nearly ten billion dollars.
Hold that fact next to two others. Amazon has committed, across successive deals, tens of billions of dollars to Anthropic. And Anthropic has committed to spend on the order of a hundred billion dollars, over the coming decade, on Amazon's cloud. The investor, the supplier, and a fast-growing source of reported demand are now bound into a single relationship — and the question this letter is about is not whether anything improper is happening, because nothing in that arrangement need be improper. The question is what happens to the meaning of the numbers once the parties are this entangled.
Begin one step back, with a pattern that looks, at first, like generosity. Across the AI build-out, the most active corporate investors have been placing more bets on more startups than anyone else in the market. Read as a count, that is conviction. Read one layer down, much of it is not capital at all. A great deal takes the form of credits — grants of the investor's own cloud computing, extended to startups in exchange for a relationship, and counted as investment. A grant of cloud credits costs the grantor very little, because it is computing capacity produced at marginal cost rather than cash out the door, and it purchases something valuable: a young company built, from its first day, on the grantor's infrastructure, billing back to the grantor as it grows. Counted as an "investment," it inflates the appearance of conviction while costing almost nothing and seeding future revenue. The cheapest way to look convinced is to invest in a currency you print yourself.
But the equity deals are where the real question lives, and the Amazon–Anthropic structure is the clearest specimen of it. A cloud provider takes an equity stake in a high-growth company. That company spends heavily on the provider's computation — increasingly, on the scarce and expensive computation the current cycle has made the binding constraint, and in this case under an explicit multi-year, multi-billion-dollar commitment to do so. Now the provider holds two linked positions in the same company. It owns equity, which appreciates when the company raises or grows. And it sells the company the compute the company needs in order to grow. The investment line and the revenue line are no longer independent of each other. They are two readings of one loop: the provider's capital helps fund the company's expansion, the expansion is spent in part on the provider's cloud, that spend is booked as the provider's revenue, the growth lifts the value of the provider's equity, and the equity gain — see the $9.5 billion — lands on the provider's own income statement.
Here is the careful part, and it is the whole point. None of this means the demand is fake. Anthropic's growth is, by every available account, real; its product is genuinely consumed; enterprises genuinely pay for it. A hyperscaler executive would be right to insist that the company funded adoption, not demand — that the customer still chose to buy, and the consumption is genuine. Grant all of it. The argument does not need the demand to be fake, and the moment it claims otherwise it picks a fight it should lose. The claim is narrower and far harder to rebut: when the investor, the supplier, and the customer are joined in one loop, an outside reader can no longer verify how much of the reported growth is independent third-party demand and how much is the provider's own capital completing a circuit. The number may be entirely sound. It can no longer be shown to be, from the disclosure available — and a growth signal that cannot be shown to be independent has already lost the property that made it worth reading.
This is what a market actually relies on when it reads a number like cloud revenue. It treats the figure as an independent observation — a reading taken from the outside world, reflecting what unrelated customers freely chose to spend. The figure's entire informational value rests on that independence. Markets work best when their most important signals are independent of the capital creating them; once that independence begins to erode, the numbers can stay perfectly accurate while becoming progressively harder to interpret. The reader is no longer measuring the world. He is measuring a system that includes the measurer.
And this is the durable idea, the one that outlives the particular companies: the problem is never that the number is false. The problem is that it can cease to be independent, and independence is the thing that gave it meaning. The same erosion shows up wherever capital and the metric it is judged by begin to share a circuit. A share buyback mechanically lifts earnings per share without any improvement in the business, because the denominator shrank. A valuation markup priced off the last markup reports momentum that may be nothing but its own echo. Fundraising momentum that attracts capital because it has attracted capital reports demand for an asset partly composed of the demand it already drew. In each case the number stays arithmetically true and stops being an independent witness. Call it, once and plainly, a loss of signal independence — and then notice how much of modern finance quietly depends on signals whose independence no one has checked.
Capital has met one feature of this shape before, and it is worth naming precisely, because the precedent is powerful and the wrong version of it is slander. In the late-1990s telecommunications build-out, equipment makers lent their customers the money to buy the equipment makers' own products and booked the resulting sales as revenue. The specific feature worth carrying forward is not the fraud that touched the worst of it; it is the structural one common to all of it: the capital extended and the revenue recognized had stopped being independent of each other, so the reported demand could no longer be read as a clean account of what the market wanted. Lucent's vendor-financing commitments reached roughly eight billion dollars — about a quarter of its revenue — and when the financed customers failed, the sales booked against their purchases proved to have been, in part, the lender's own money making a circuit. I am not alleging that today's cloud provider is running that playbook; the mechanics differ, the disclosures differ, and the public record shows no improper recognition. The single feature I am carrying across is the one that matters: when the money supplied and the revenue booked lose their independence, the demand figure stops being verifiable from outside, whether or not anyone acts in bad faith. That loss of verifiability — not fraud — is what made the telecom numbers dangerous, and it is the same property at issue now.
It is fair to object that this is how strategic investment has always worked: corporates invest to advance their own platforms, everyone knows it, and bundling equity with compute commitments is simply an efficient version of an old and legitimate practice. Largely true, and not the point. The objection defends the practice; the letter concerns the reading. There is nothing wrong with a cloud provider seeding its ecosystem, or with an AI company choosing the infrastructure of an investor who understands it. The hazard appears only when the resulting consumption flows into a figure the market reads as independent third-party demand — when an activity that is partly internal to a relationship is reported through a line item everyone treats as external. The practice is fine. The conflation is the cost, and the conflation lives in the disclosure, not in the strategy.
There is a stronger objection: that the entangled portion is small relative to a number as large as a hyperscaler's total cloud revenue, and therefore immaterial. Perhaps, today. But the binding constraint of this cycle is exactly the expensive computation these arrangements supply; the companies being funded are exactly the fastest-growing consumers of it; the commitments are measured in tens and hundreds of billions; and the structure is expanding, not contracting. A dynamic that is immaterial at small scale becomes material at the scale these relationships are heading toward — and the time to understand a loss of signal independence is before it is large enough to move the headline number, not after, when it has been read as organic for years and the correction arrives all at once. The telecom arrangements looked immaterial early, too.
So the investment count is not a clean measure of conviction, and the cloud-growth line is not, by itself, a clean measure of independent demand. Each is a figure through which a single party's capital may be circulating and being counted as it passes — conviction priced in a currency the investor prints, growth partly composed of consumption the investor helped fund. The market reads both as external signals taken from the world. They are, in part, internal readings of a loop, and the disclosure does not separate the two. An investor who wants to understand a hyperscaler's position in this cycle has to ask the question the income statement is built not to answer: how much of this growth did you pay for yourself?
The correction is not to assume the worst, and it is certainly not to call the growth fake. It is to refuse to read a number that has lost its independence as though it still had it. It means treating investment counts as marketing until the currency is known; reading a cloud-growth figure with the question of self-funded demand held open rather than closed; and remembering that the most dangerous numbers in financial history were rarely the false ones. They were the accurate ones that had quietly stopped being independent, and were trusted as though they hadn't. The investors who saw the telecom reckoning coming were not the ones who detected fraud. They were the ones who noticed that the money supplied and the revenue booked were no longer separable, and stopped reading the growth as an independent fact before the market did.
When a company can pay in its own coin, every number it reports has to be read twice — once as stated, and once for whose capital is actually moving through it. The number may be entirely true.
The costliest error is to assume that, because it is true, it is also independent — and to learn only later that the most-watched line in the market had, for years, been partly measuring itself.
—QED—
Sources & Methods. The argument is structural; the contemporary figures are verified public reporting, used to illustrate a mechanism, and the historical parallel is offered as a single shared structural feature, not as equivalence. No claim of improper conduct by any company named or unnamed is made or intended. The empirical record the analysis is set against:
■ The Amazon–Anthropic structure — Amazon's Q3 2025 results, which included a ~$9.5B pre-tax gain recorded in non-operating income from the markup of its Anthropic stake following Anthropic's September 2025 funding round (no sale or cash event); Amazon's successive equity commitments to Anthropic (an initial program building to $8B, and a further commitment announced April 2026 of up to $25B, the latter tied to stated commercial milestones); and Anthropic's commitment to spend $100B+ on AWS over the coming decade, alongside multi-gigawatt Trainium capacity commitments. Sources: Amazon earnings coverage (GeekWire, CNBC), company statements, contemporaneous reporting (Network World, Motley Fool). Anthropic's annualized revenue (reported to have surpassed ~$30B) and AWS segment growth (~24% YoY in Q4 2025) are likewise from public reporting. All figures should be cited as reported; the entanglement reading — that the equity and revenue lines are no longer independent — is the letter's structural argument, not a claim asserted by any of these sources.
■ The credit-grant pattern — CB Insights, AI Agent Bible (2025): the account of cloud providers seeding agent startups heavily through non-equity accelerator grants of cloud credits and technical enablement rather than capital, with equity deals a smaller subset. Used to support the "currency you print yourself" point about investment counts.
■ The compute constraint — the report's treatment of specialized inference and training capacity (and in-house silicon such as Trainium/Inferentia) as the binding cost of the current cycle; corroborated by reporting that the Amazon–Anthropic capacity commitments are aimed at relieving exactly this constraint.
■ The vendor-financing precedent — the late-1990s telecommunications vendor-financing episode. Verified specifics: equipment makers (Lucent, Nortel, Cisco, others) lent customers funds to purchase the makers' own equipment and booked the loans' value as revenue while holding the debt as a balance-sheet asset; Lucent's vendor-financing commitments reached ~$8.1B, roughly 24% of ~$33.6B revenue (1999–2000), including ~$2B committed to WinStar, which filed bankruptcy after Lucent declined a final ~$90M extension; Nortel extended billions to CLECs, most of which failed. Distinct from the round-tripping fraud of the era (e.g., Global Crossing/Qwest reciprocal capacity "sales"); the letter deliberately carries forward only the vendor-financing feature — loss of independence between capital extended and revenue booked — and explicitly disclaims the fraud framing. (Note: earlier drafts referenced a specific causal line to Sarbanes-Oxley; that linkage has been removed as imprecise — SOX followed the broader accounting scandals of the period and should not be attributed specifically to telecom vendor financing without a source establishing the connection.)
■ Signal independence and reflexivity — general principles of information quality and reflexivity in markets (the dependence of a signal's value on its independence from the party reporting it); the buyback/EPS, markup-on-markup, and fundraising-momentum examples are structurally definitional rather than empirical claims and require no company-specific sourcing.