If The AI Bubble Bursts, Who Gets Hurt?
Notes by SDG ·
I came across a headline about Michael Burry replacing bearish stock positions with put options across several AI-linked companies, including Nvidia, Micron, Nebius and Palantir. According to Stocktwits, he now thinks an AI bubble could burst sooner than he previously expected.[1]
Burry’s trade is not proof of an approaching crash. Buying puts also means making decisions about timing, price and the cost of the options. Even a correct concern about valuation does not automatically become a profitable trade.
The main transmission channels of an AI slowdown would likely be supplier orders, project delays, utilization, depreciation and financing. Customers could postpone equipment purchases, developers could delay construction, existing capacity could earn less if demand or rental prices weaken, and heavily financed operators could face higher borrowing costs or refinancing problems. Meanwhile, the rapid obsolescence of AI accelerators can erode asset values faster than traditional depreciation schedules imply. These channels can reinforce one another: weaker utilization can reduce cash flow, which makes financing harder, which can lead to more cancelled orders and delayed projects.
But the question underneath the headline is worth following: if AI infrastructure spending slows, where does the damage actually go?
Not just Nvidia. Not just the companies with “AI” in their investor presentations.
Who else has already spent money on the assumption that the next order is coming?
The same customer, several layers down
Axios reports that Sona mapped 255 public companies across the AI ecosystem.[2] I haven’t independently reviewed that entire roster. The number is a starting point, not a list of 255 companies about to get into trouble.
Still, it helps explain why this is bigger than a semiconductor story.
A data center needs processors and memory. It also needs servers, networking, optical connections, cooling, electrical equipment, a building and power. Somebody has to construct it. Somebody has to finance it.
That takes us from Nvidia and Micron to companies such as Arista, Coherent, Celestica, Vertiv and Eaton, then out toward contractors, power infrastructure, landlords and lenders. They have different businesses and very different levels of exposure. But parts of their growth can depend on the same customer deciding to build the same facility.
Owning five companies in five different parts of that chain can look diversified. Economically, some of those holdings may still be waiting for the same purchase order.
And even Burry’s named targets aren’t one identical bet. A software business, a chip supplier and an operator renting computing capacity do not have the same costs or financing needs. Their shares can fall together without their businesses breaking in the same way.
Imagine the slowdown starts quietly
Here is a hypothetical scenario, not a prediction.
AI usage keeps growing. Employees keep using it. Companies keep experimenting. But the revenue or savings from some deployments take longer to arrive than expected.
At a budget review, a large customer decides to delay two planned data centers. It is not abandoning AI. It wants to use more of the capacity it already has before paying for more.
Nobody announces that the future has been cancelled. The earnings call refers to “optimizing the pace of deployment.”
Now imagine the industry expected spending to rise from 100 to 150, but it reaches only 120. Spending still grew 20%. Yet it came in 20% below the plan.
That gap matters if suppliers hired people, ordered components and expanded factories for 150.
The first reaction could happen in share prices. Investors reduce both their earnings forecasts and the valuation they are willing to pay. A company can remain profitable and still suffer a large stock-market decline.
For a simple illustration: $10 of expected earnings valued at 40 times earnings gives a $400 share price. Change that to $8 at 25 times earnings and the price becomes $200. No bankruptcy required.
Then the revised budget reaches everyone else
The server manufacturer gets a changed delivery schedule. Networking and optical suppliers see orders pushed out. Inventory takes longer to move. A cooling installation that was supposed to happen this year moves into next year.
The impact would not be uniform. Firm contracts, deposits and cancellation penalties can protect a supplier. Service revenue and unrelated customers can soften the hit. Some equipment may be redirected to another project.
But protection from cancellation is not always protection from delay. Cash that arrives six months late still leaves a six-month hole somewhere.
Further out, contractors may hire fewer people. Developers reconsider their next sites. Power companies and equipment suppliers reassess proposed projects, although existing contracts and wider grid demand can provide a cushion.
A postponed project also does not necessarily mean demand has weakened. A facility waiting for a power connection is different from one waiting for customers. In the first case, usable capacity elsewhere may become more valuable. In the second, rental prices may come under pressure.
Confusing those two situations would lead to a very bad diagnosis.
For this bearish scenario to deepen, we need more than construction delays. We need demand or customer funding to disappoint while previously ordered capacity continues coming online.
That is when providers may start competing harder on price.
Hyperscaler balance sheets versus specialized cloud financing
The financial position of the buyer matters as much as the amount of infrastructure being built.
Microsoft, Google, Meta and Amazon can fund much of their AI capital expenditure from operating cash flows generated by large, established businesses: search, advertising, enterprise software, cloud hosting and e-commerce. They can make an investment that looks excessive in the short term, tolerate idle capacity and wait years for utilization to improve.
That does not make their spending automatically rational. It does mean they have more room to absorb a forecasting error.
Specialized GPU cloud providers operate under a different constraint. They often depend on debt and equity markets to purchase accelerators, build facilities and bridge the period before customer receipts arrive. Their financing may include maintenance covenants, collateral requirements, minimum utilization assumptions or restrictions on additional borrowing.
A hyperscaler may be able to overbuild and carry underused capacity for years. A levered operator may not survive a six-month decline in utilization rates, particularly if rental prices fall at the same time. Interest, leases, power commitments and other obligations continue even when the machines are idle.
This is why the same demand slowdown can be a manageable capital-allocation problem for a cash-rich platform and a solvency problem for a specialized provider.
Depreciation can accelerate the pressure
AI infrastructure also has a less obvious vulnerability: the useful economic life of the equipment may be shorter than the accounting schedule suggests.
Traditional real estate and many categories of physical servers are often modeled over five- to seven-year depreciation periods. AI accelerators can lose value much faster when a new generation delivers substantially better performance per dollar, per watt or per unit of floor space.
If utilization falls while the equipment is rapidly becoming obsolete, the operator faces two pressures at once. Revenue declines because fewer customers are renting the machines, while the assets are losing economic value and continuing to generate depreciation expense. The resulting return on invested capital can deteriorate much faster than a standard cash-flow model implies.
A machine can remain technically functional and still become commercially unattractive. Customers may prefer newer hardware, forcing the operator to cut prices, spend again on replacements or accept lower utilization for the older fleet.
This creates a mismatch between accounting lives, financing lives and economic lives. Debt may have been raised against equipment expected to produce cash for several years, while the market value of that equipment is determined by a much faster technology cycle.
Debt is where a slowdown can become a crisis
A well-funded business may be able to wait.
A heavily financed operator has a harder problem. Its equipment and facilities were paid for under assumptions about utilization, pricing and customer receipts. Interest, lease payments and other obligations do not automatically fall when those assumptions change. Nor does depreciation disappear simply because the machines are earning less.
CoreWeave’s second-quarter 2026 results illustrate why financing belongs in this discussion. It reported approximately $2.6 billion in revenue and $640 million in net interest expense. Its substantial backlog also carries delivery and service-availability conditions.[3]
Those figures do not predict default. They show why revenue growth and financial resilience are different questions.
Nor should every specialized cloud provider inherit the same risk assessment. Cash reserves, customer prepayments, contract protections, debt maturities, remaining construction costs and the age of the installed GPU fleet can produce very different outcomes.
In our severe scenario, weaker rental economics make lenders more cautious. Refinancing becomes more expensive or unavailable. Equity investors demand a lower price. An operator cuts expansion, tries to sell equipment or restructures its obligations.
Selling older accelerators into a weakening market can be especially difficult. If many operators attempt to liquidate similar hardware at once, secondary-market prices may fall, reducing the collateral available to lenders and increasing losses for owners.
Those actions create another round of cancelled orders and weaker asset prices. Suppliers pull back. Other developers find funding harder to obtain.
Now financing conditions are reducing spending, not merely reacting to it.
Some investors could suffer permanent losses. The facilities and machines might continue operating under different ownership, even if the original shareholders do not recover their money.
That is a plausible bubble-burst mechanism. It is not yet an argument that we are heading for another 2008. A broader financial crisis would require evidence about the size and concentration of lender losses, guarantees and funding dependencies—not just a dramatic network diagram.
But the effects could reach people who never bought an individual AI stock: through diversified funds, weaker hiring, construction jobs and local businesses expecting activity around new facilities.
The strongest counterargument: demand may catch up
There is a serious problem with turning that scenario into a confident forecast.
The current operating numbers do not describe an industry already collapsing. Nvidia reported $89 billion in data-center revenue for the quarter ended July 26, 2026, up 117% from a year earlier.[4]
That does not settle the valuation debate. It does mean a bearish argument has to explain how exceptionally strong demand becomes insufficient demand, rather than simply assume it has already happened.
The largest buyers also have resources that many speculative developers do not. Microsoft reported about $40.6 billion in quarterly operating income for its fiscal fourth quarter of 2026.[5] That is not cash available after every investment obligation, and it certainly does not make every project sensible. But there is a meaningful difference between funding expansion from a large established business and needing the next financing round to keep building.
There is also a credible path in which cheaper computing makes more applications economical.
If the cost of an AI task falls, companies may do much more of it. Some work that was too expensive becomes practical. More useful applications could absorb capacity that initially looked excessive.
The important word is could. Lower prices do not guarantee enough additional volume to preserve suppliers’ profits. But they are not automatically evidence that demand is disappearing either.
And the benefits of AI need not all show up as standalone chatbot subscriptions. A business might justify spending through better existing products, more effective advertising, faster development or lower operating costs. Those returns still need to be measured; looking only for a new “AI revenue” line may miss part of the picture.
There is another possibility: the boom cools without an industry-wide collapse. Some projects get delayed. Margins normalize. Expensive shares fall. Stronger businesses keep investing, and usage gradually catches up with installed capacity.
Unpleasant for investors who paid too much? Certainly.
The end of AI? Not even close.
Signals and counter-signals to watch
A falling stock price would not tell us whether the infrastructure boom is breaking. Neither would another announcement about a multibillion-dollar data center.
I would watch what customers are buying, what operating capacity is earning, how quickly the equipment is becoming obsolete, and whether the companies building it can meet their obligations. These are indicators to monitor, not claims that they are all flashing red today.
Customer spending and renewals. Several major buyers cutting commitments because returns are disappointing would be a warning. Smaller renewals, cancellations and requests to renegotiate contracts would add weight to it.
The counter-signal would be customers renewing, paying and expanding usage without providers having to offer increasingly generous discounts. A project moving into the next quarter is less concerning if the customer still needs it and the funding is secure.
Rental prices and utilization. Lower GPU rental prices deserve context. Newer hardware can make computing cheaper without indicating weak demand. I would be more concerned if comparable machines were earning less while also spending more time idle.
If lower prices attract enough additional paid work to keep capacity busy and support healthy cash generation, that would argue against a demand collapse. More usage alone is not enough if the economics keep getting worse.
Hardware economics and depreciation. I would also compare the age and performance of the installed fleet with the depreciation and replacement assumptions used in financial models. A provider can report acceptable utilization while returns deteriorate if customers increasingly prefer newer chips or if older accelerators require steep discounts.
The warning would be stronger if equipment values fell, replacement cycles shortened and operators continued carrying debt based on older assumptions about useful lives. A machine that remains operational is not necessarily an asset earning an adequate return.
Supplier orders and cash collection. Weakening orders across chips, servers, networking, optics and cooling would be more meaningful than a disappointing quarter at one company. Rising cancellations, inventory building faster than sales, and customers taking longer to pay would strengthen the warning.
The counter-signals would be repeat orders, deliveries proceeding on schedule and cash collections keeping pace. A large backlog is useful, but it needs to turn into revenue and then cash. Suppliers can still report strong sales while delivering older orders, even as new demand weakens.
Project execution. Repeated delays between announced capacity and customer billing would deserve attention, especially if construction costs and interest keep accumulating.
But a facility waiting for a power connection is not the same as one waiting for customers. Completed sites filling up, customers accepting delivery and revenue starting on schedule would suggest that the buildout is finding buyers.
Financing. More expensive borrowing, tighter collateral requirements, failed refinancing or an urgent equity raise would make a slowdown harder to manage. The concern would be greater if important customers also needed fresh funding to keep paying their computing bills.
The counter-signal would be enough cash and committed, accessible funding to cover remaining construction, debt payments and operating needs. Customer prepayments can help, but they come with obligations to deliver future services. Raising money is not automatically a sign of trouble; the terms and the use of that money matter.
I would become more concerned if these signals appeared together across several companies: customers cutting commitments, existing capacity earning less, GPU values declining faster than expected, and lenders becoming less willing to finance it.
If renewals remain healthy, operating sites stay busy, newer hardware continues to earn attractive returns and funding holds up, a spending pause could be exactly that—a pause. Shares could still fall because investors paid too much for the growth they expected.
There is no single number that settles this. But checking orders, cash receipts, utilization, equipment values and financing terms should give us more to work with than counting how often management says “unprecedented demand.”
This is scenario analysis, not a forecast or investment recommendation. Company examples illustrate different business models, not predicted failures.
Sources and reporting notes
1. Stocktwits, September 28, 2026: Burry’s reported move from stock shorts to puts. The positioning is attributed to this report; I did not independently inspect his portfolio.
2. Axios, September 21, 2026: Sona’s AI ecosystem research. The full underlying 255-company roster was not retrieved. This blog does not claim a company-by-company audit.
3. CoreWeave, Q2 2026 earnings release filed with the SEC.
4. Nvidia, Q2 fiscal 2027 results, August 26, 2026.
5. Microsoft, Q4 fiscal 2026 results, July 29, 2026.
The slowdown sequence and valuation arithmetic are illustrative, not company guidance.