And once again the company borrowing to sit at the incumbents' table is not one of the incumbents. In 1999 it was Global Crossing, laying fiber on credit so it could stand beside the telephone giants, and it filed for Chapter 11 in January 2002 while the giants survived to write down their own excesses at leisure. Its name today is Oracle. Behind Oracle stands a financing stack of private equity, private credit and insurance balance sheets whose contractual clocks do not care how the AI story ends.
Michael Burry's October 8 note in Cassandra Unchained reads, on its surface, like a political meditation on technocracy. Underneath the essay sits a capital-structure argument. Burry calls the hyperscaler buildout "a want, not a need." The market has not finished the thought. A want can be cancelled. Debt service cannot. The stress in this cycle will not land on the monopolists who choose to spend. It will land on the joiners who must spend, and on the lenders who financed them.
Why Does a Monopoly Spend Like Prey?
Burry's opening observation is the right place to start. "Monopolies almost by definition need not spend wildly." A firm with durable pricing power harvests. It does not race. Yet Microsoft, Amazon, Alphabet and Meta are spending as if a competitor were at the gate, and Apple, the one giant with no cloud silo to defend, is the conspicuous exception.
The consensus explanation is commercial. Demand for compute is real, the backlog is real, and these companies print enough cash to fund the build without strain. That part of the consensus is correct. Microsoft can afford its data centers. Alphabet can afford its data centers. Neither faces a margin call if utilization disappoints for six quarters.
Burry's explanation is political, and here we part company with the consensus framing on a narrower point than he does. He argues the incumbents expect AI to become one more monopoly rent, collected under a government that will "tolerate and fear" the oligopoly. Whether or not that is the motive, its financial implication is the useful part. Spending aimed at defending a rent rather than earning a return has no hurdle rate. As Burry puts it, it has no defined goal or economic end-point.
Spending without an end-point is also spending without a floor. The incumbents can slow it the moment the political or competitive calculus shifts, and their shareholders would cheer the buyback that replaces it.
The incumbents hold an option to stop. The joiners do not.
Who Actually Needs the Oligopoly?
The joiners do. Burry names them: OpenAI, Anthropic, Oracle. None is a monopoly. None can imagine breaking Microsoft, Amazon, Google or Meta inside their own silos. Their only route into the dominant tier is to spend their way there "under cover of a new paradigm," and Burry notes that the prize was large enough that "even Larry Ellison was seduced to bet his company on it."
That phrase deserves a literal reading. Oracle is the one public company in the AI cloud set whose spending runs ahead of its own cash engine rather than beneath it.
The incumbents fund capex from franchises that throw off cash in every regime: search advertising, enterprise software, retail logistics, social advertising. Oracle has a franchise too, in databases and applications, but it is the smaller franchise attempting the larger leap. Its capital intensity puts it in hyperscaler company without hyperscaler cash generation behind it. The published cash flow statements show the consequence. Operating cash flow is still healthy. Free cash flow, after the build, is not.
When the gap between operating cash and capital spending is closed by borrowing, the spending changes character. It stops being discretionary. A Microsoft data center that sits half-utilized is an embarrassment on an earnings call. An Oracle data center financed against a lease, a project loan or a customer's take-or-pay commitment is a fixed obligation with a maturity date. The customer, in several of the most publicized commitments, is one of the other joiners, which is itself burning cash to stay in the race.
This is the Global Crossing shape. The incumbent overbuilds from strength and writes it off. The challenger overbuilds from leverage and does not survive the write-off. History does not guarantee the repeat; the 2000 telecom bust also produced challengers that restructured and lived. The structure is the same here, and the structure is what matters for who absorbs the loss.
Whose Money Is It?
Burry's subtitle is "Other Peoples' Money," and the second half of his note explains why. He describes "creaks and groans" in valuation marks across private equity and private credit, and adds the line the equity market has not digested: "We've moved shadow banking offshore and reinvented the role of insurance companies."
That sentence describes where the AI buildout's marginal dollar now comes from. Data center construction is increasingly financed outside the hyperscaler balance sheet. Special-purpose vehicles, private credit funds, and insurers buying long-dated paper to match annuity liabilities sit between the capex announcement and the concrete. The hyperscaler books a lease or a service commitment. The lender books an asset marked to a model.
Here the guidance language and the disclosure diverge. Capex guidance tells the market how much is being built. It does not tell the market who owns the risk if the buildings outlive their purpose. The answer, increasingly, is a private-credit fund whose investors expect quarterly marks to move gently, and an insurer whose regulators expect the paper to be money-good across decades.
Burry has argued across several pieces, by his own account, that this leverage "fundamentally conflicts with financing timelines." The mechanism is a duration mismatch with three layers:
- The equipment inside the data center, GPUs and networking, has an economic life measured in a few product cycles, because each new generation reprices the old one.
- The building, the power interconnect and the debt that financed them run on schedules measured in a decade or more.
- The tenant commitments that justify the debt are only as good as the tenant, and several tenants are joiners funding their own losses with outside capital.
When the middle layer is financed by insurers and private credit, a mark-down at the top layer does not stay at the top. It travels down the stack as a covenant problem, a refinancing problem, and eventually a capital problem at the insurer. Burry claims a slide in Part VI of his Heretic's Guide maps exactly that path back to the buildout. The slide sits behind the paywall and we have not seen it. The direction of travel needs no slide.
We should be explicit about the unverified links. Which specific alternative-asset managers, which insurers, and which SPVs carry the largest data-center exposure is not established in Burry's public text, and we will not name tickers he did not name. The inference that private-credit marks are the first visible stress point is his observation; that it transmits to equity holders of the lenders is our extension, and the size of that transmission is unknown.
Shareholders chose this risk. The annuity holders underneath did not.
What Breaks First?
Burry offers two answers, and they work on different clocks.
The slow clock is compression. "Ultimately, compression will do its deed," he writes, arguing that "human thought is far too redundant for what is being built." Read as engineering, this is a claim about efficiency: models get smaller, inference gets cheaper, and the compute required per unit of useful output falls faster than the buildout assumed. Each efficiency gain is good news for AI users and bad news for whoever financed capacity sized to the old curve. The towers of Babel were not abandoned because the builders lost ambition. They were abandoned because the builders stopped understanding each other, and a buildout sized for one efficiency curve becomes unintelligible to the next.
The fast clock is rates. Burry notes that "long-term rates are rising before the timelines have a chance to play out at all." This is the one place where macro does causal work in the argument. A data center financed at one long-term rate and refinanced at a higher one sees its equity cushion erode without a single tenant defaulting. Precise yield levels are not summarized here, so we will not put a number on the move. The direction Burry describes is sufficient for the mechanism: rising long rates shorten the runway the financing structure was built to need.
The closest analog in the cross-asset record is 2022, when the fastest tightening in four decades took equities down roughly a fifth and handed long-duration bonds their worst losses in generations. The parallel is imperfect. Inflation started this cycle from a lower base, which limits how hard the Fed needs to lean. That difference cuts both ways. A Fed that need not hike aggressively also need not cut aggressively, which means the long end can drift higher on supply and term premium while the policy rate sits still. For an asset class whose financing assumed the long end would come back down, a drifting long end is the worse outcome, because it never produces the crisis that forces a rescue.
The order of failure, then, runs roughly from the edges inward. First the private marks that cannot be held at par. Then the joiners whose tenants cannot pay, or whose refinancing costs jump. Then the suppliers whose order books were built on the joiners' commitments. The incumbents come last, and they come as buyers.
Why Is the Tape So Calm?
Burry ends his public excerpt with the question that matters most. If Wall Street hears the same "nasty noises," why is the bull run to new highs "so well-affirmed, well-accepted, and well-projected?"
Listen to how the consensus sounds. They print cash. Microsoft prints cash! Alphabet prints cash! Amazon prints cash! Meta prints cash! Oracle... Oracle borrows cash, but Oracle has backlog! The backlog is from OpenAI! OpenAI is funded! By whom? By the next round!
Positioning tells the same story in plainer terms. The crowd is long risk on a soft-landing base case, with cross-asset correlations pricing a coordinated recovery. By available consensus data, the semiconductor complex measured by the Philadelphia Semiconductor Index, and its leaders NVDA, AMD and SMCI, along with MSFT, are treated as the obvious longs of the cycle. Nobody in that crowd is hedged against the joiners failing, because the joiners failing is not a scenario anyone models. It is treated as a tail.
It is not a tail. Most people, asked whether the AI financing stack cracks inside the next two years, would say the odds are low but not negligible, which is their way of saying they have not looked at who holds the paper. The path the evidence supports is not a sudden collapse but a slow repricing: private marks give way first, the joiners' cost of capital rises, and the semiconductor names discover that a share of their demand was financed rather than earned. A faster outcome, where the stress reaches earnings revisions and credit availability before the crowd repositions, is a meaningful minority case. The case where the regime simply stabilizes and the noises fade is the least likely of the three, and it is the one the tape is priced for. I don't know the timing. Nobody does. What can be known is that the distribution is lopsided and the positioning is not.
Second-level thinking applies here. First-level thinking says the incumbents are rich, therefore the buildout is safe. Second-level thinking asks who is not rich, who signed the long-dated commitments, and whose money they used. The incumbents' wealth is precisely what lets them abandon the folly when it stops serving them. Their wealth protects them. It does not protect the people they rent capacity from.
The instruments most exposed to the thesis resolving are therefore not the obvious ones. Oracle is first in line, because it is the public joiner with the widest gap between its spending and its cash generation. Behind it sit the semiconductor names whose forward demand leans on the joiners rather than the incumbents; that share of demand is not disclosed cleanly, and we treat the size of the exposure as unverified. Behind them, invisible in any equity screen, sit the private-credit vehicles and insurers whose marks Burry hears groaning. The incumbents themselves, MSFT included, are the least exposed members of the group. They are also the ones most likely to buy the stranded assets at the bottom.
The reckoning, if it comes, will not look like the end of AI. It will look like a transfer: of data centers from leveraged owners to cash-rich ones, of losses from equity marks to insurance reserves, of an oligopoly's ambitions from the joiners back to the four incumbents who never needed to join anything.
The monopolists can stop spending; the borrowers cannot stop paying.
