The AI Risk Is in the Balance Sheet

The right question is no longer whether AI works. It is whether the cash flows will arrive fast enough to validate the commitments already being made in anticipation of them.

The wrong question

The debate about artificial intelligence is still framed too often as a referendum: bubble or no bubble. It is the wrong question, because it squeezes two quite separate judgments into one.

A technology can be genuinely transformative and still be financed badly. Railways changed the world and ruined investors. Fibre became essential infrastructure after enormous amounts of telecom capital were destroyed. The internet was real long before many of the prices attached to it proved sustainable.

AI may be following the same pattern. The evidence increasingly suggests that the technology is real, the infrastructure build is necessary and the eventual economic effect could be enormous. None of that guarantees that today's capital will earn an adequate return.

This matters because the risk is moving. Technology risk is falling as models improve, inference becomes cheaper and adoption broadens. Financial-structure risk is rising as more capital is committed against revenues that have not yet arrived.

A better question to steer by is this: not ‘Is AI a bubble?’ but ‘What future cash flow must occur to validate today's financing structure—and what happens if capital markets tighten before it arrives?’

1. The technology has crossed an important threshold

The weakest version of the bubble argument is that the underlying capability is mostly theatre. That view gets harder to hold with every quarter. Model performance has improved sharply across reasoning, coding, multimodal tasks and professional workflows. Reliability is still patchy, but the direction is not in doubt: capability is rising, cost is falling and applicability is broadening.

What matters is the combination, not any one benchmark. Electricity did not transform production because the first motor beat every steam engine. It mattered because electricity eventually became cheap, flexible and general enough for factories to reorganise around it. Computing followed the same path: in 1987 Robert Solow remarked that the computer age was visible everywhere except in the productivity statistics, and the gains arrived only in the late 1990s, once firms had rebuilt inventory, supply-chain and back-office processes around cheap information processing. The shipping container did too. Malcom McLean's first container voyage in 1956 did not transform trade; standardised box sizes, purpose-built ports, gantry cranes and redesigned logistics networks did, over the following two decades. AI may be approaching a similar threshold.

The most important price in the system is increasingly intelligence per dollar. By Stanford's AI Index estimate, the cost of running a model at a given level of capability has fallen roughly 280-fold since late 2022 even as capability improved. If machine cognition becomes dramatically cheaper, companies will go well beyond better chatbots and redesign products, workflows and eventually organisations around an assumption that some forms of intelligence are abundant. [1]

That would be a genuine economic revolution. Cheap intelligence, though, does not mean cheap infrastructure; it can mean the opposite. A basic query consumes little compute; reasoning systems consume more; agentic systems can plan, search, call tools, revise their work and loop repeatedly. The frontier in 2026 shows what that looks like in practice.

In May, OpenAI reported that a general-purpose reasoning model had produced the core ideas to disprove the Erdős unit-distance conjecture, an open problem since 1946, which outside mathematicians then verified. In July, an OpenAI model proved the cycle double cover conjecture in graph theory, and two independent mathematicians published expositions of the argument. [2]

In September, OpenAI said nearly 100 agents worked for about 50 hours to construct a singularity in the three-dimensional Euler equations, delivered with a machine-checkable Lean certificate. The same month, Stanford researchers published the Virtual Biotech in Science: a system in which a virtual chief scientific officer delegates drug-discovery questions to teams of AI agents covering target discovery, safety assessment and clinical development. And METR's January 2026 update estimates that the length of task agents can complete on their own has been doubling roughly every four months since 2023. None of these was a single cheap query; each consumed large amounts of compute. So falling unit costs can end up stimulating far more usage than efficiency saves. [3, 4, 5]

What follows is a familiar pattern: the bottleneck moves. First GPUs, then high-bandwidth memory, then advanced packaging, then data-centre capacity, and increasingly power, grid connections and physical infrastructure. Scarcity has not disappeared. It has changed address.

2. The economy is adopting AI faster than it is reorganising around it

The physical system is being installed very quickly: models, chips, cloud capacity, data centres, training clusters, inference systems, networking and dedicated power. The productive economy is not reorganising at the same speed.

AI use is already broad. Surveys suggest that about 88% of organisations now use AI, and roughly 70% use generative AI in at least one function. Yet much of that use remains shallow. Companies are adding copilots to old processes, automating individual tasks, running pilots and experimenting with agents. They are mostly not redesigning reporting lines, decision rights, workflow architecture, controls or staffing models around cheap machine intelligence. [6]

This is the crux. A company can start using AI tools in weeks. Reorganising a business around them can take years. Productivity therefore lags capability, and realised cash flow can lag infrastructure investment even further.

Hence the central contradiction of this boom: AI can be under-deployed economically and over-invested financially at the same time. The bulls may be right about the technology while the bears are right about the returns.

It also explains why treating ‘AI’ as a single cycle is misleading. Semiconductors already look like a mature infrastructure industry; memory has displayed scarcity economics; data centres are entering aggressive financial expansion; power is still catching up; agents are early; enterprise applications are still searching for dominant designs; organisational redesign is barely underway. Different layers are at different points in the same technological transition.

3. The risk begins when finance stops following demand and starts pulling it forward

Early in a technological boom, finance follows real opportunity. Profits appear, capital follows, infrastructure gets built and new supply arrives. That is the healthy version.

Over time, the direction can flip. Instead of profits leading investment, expected future demand begins to lead capital formation, and capital formation leads capacity construction. Finance is no longer merely funding the expansion; it is helping determine its speed.

That transition is visible in AI. The strongest hyperscalers still possess enormous operating cash flows. Outside that core, however, the build increasingly depends on external capital: bonds, private credit, special-purpose vehicles, GPU-backed loans, long-term capacity agreements, minimum-use commitments and strategic supplier investment. By my tally, roughly $590 billion was raised across major AI financing channels in 2026. Outside evidence points the same way. Alphabet, Amazon, Meta, Microsoft and Oracle had issued about $220 billion of bonds by 10 August, according to BNP Paribas data, against roughly $108 billion for the whole of 2025; Amazon alone sold about $54 billion in a single March deal. Off balance sheet, Apollo and Blackstone closed a $35 billion private-credit package in June to fund compute for Anthropic through a special-purpose vehicle that owns the chips and leases them back. GPU-backed lending has become mainstream: CoreWeave's $8.5 billion loan was the first GPU-collateralised financing to earn an investment-grade rating, and IREN followed with a $3.65 billion facility secured on GPUs and its Microsoft contract. Morgan Stanley expects around $570 billion of AI-related debt issuance globally this year. [7, 8, 9, 10, 11, 12, 13]

None of those structures is inherently irrational. New infrastructure often requires new underwriting methods. The danger is quieter than that: each layer makes it easier to spend against expected future demand rather than current cash flow. Slowly, a gap opens between what has been built and what has been economically justified.

That is why I think public-equity multiples may be the wrong place to look for the main excess. Nvidia's earnings have grown rapidly while its trailing valuation multiple has compressed materially. In the quarter to 26 July 2026, Nvidia reported revenue of $96.2 billion, up 106% from a year earlier, yet its trailing price-to-earnings ratio fell from about 57 times in July 2025 to about 31 times a year later. By late September its forward multiple was around 19 times, the lowest in a decade and well below its longer-run average of about 32. The more revealing signal may be the quantity and duration of capital being committed to physical capacity on the assumption that scarcity and utilisation will persist. S&P Global Ratings expects Alphabet, Amazon, Meta, Microsoft and Oracle to spend about $750 billion on capital expenditure in 2026 alone. [8, 14, 15]

4. The decisive distinction is between cash-generative assets and claims on future cash flow

The AI stack contains businesses with radically different financial characteristics, even when all of them are labelled ‘AI’.

At the strongest end sit the hyperscalers and semiconductor suppliers. Microsoft, Meta, Amazon and Alphabet possess large cash-generating franchises outside any single AI investment. In the June 2026 quarter, Alphabet generated $39.1 billion of operating cash flow, Amazon $45.4 billion and Meta $31.9 billion, while Microsoft produced $182.9 billion over its fiscal year to June. Yet even here the margin has narrowed: Alphabet reported negative free cash flow of $5.9 billion, its first since its 2004 listing; Amazon's trailing twelve-month free cash flow swung to an outflow of $7.6 billion; and Meta's capex absorbed about 98% of its operating cash flow. The franchises still cover the build, but with far less to spare. Nvidia, TSMC, SK hynix and the broader memory and packaging ecosystem are being paid now for scarce inputs. TSMC's revenue in the quarter to September 2026 rose 41% in dollar terms to $46.7 billion, at a net margin of 51%. SK hynix's second-quarter revenue rose 257% from a year earlier and its operating profit 557%, at an operating margin of 76%. Their economics depend far less on distant assumptions about eventual AI monetisation. [16, 17, 18, 19]

The picture changes in the infrastructure middle. Neocloud operators monetise GPU capacity through rental contracts, but their asset bases are much more heavily financed. The model is simple to describe. A neocloud such as CoreWeave, Nebius or Lambda signs a multi-year contract to supply GPU capacity to a hyperscaler or AI laboratory. It then borrows against that contract and the chips themselves to buy the hardware, and earns the spread between rental income and the combined cost of debt and depreciation. The economics work only if three things hold: the customer is creditworthy, utilisation stays high, and the GPUs keep earning premium rents for longer than the loan takes to repay. Lenders price the risk accordingly. GPU-backed facilities tied to investment-grade customers have priced at roughly SOFR plus 2.25–3.00%, while private credit for AI infrastructure more broadly yields 10–15%. Because most neoclouds lack strong balance sheets of their own, their credit quality is largely borrowed from their customers, and when that support looks thinner the market reprices quickly: in mid-October about $4 billion of CoreWeave's notes traded below 80 cents on the dollar. Oracle still has a large cash-generative enterprise franchise, yet the scale of its incremental AI build is beginning to run ahead of the cash internally generated by that programme. Its model is different: Oracle uses the cash flow and credit rating of its database and applications business to build data centres for a small number of very large customers, recovering the outlay over multi-year contracts recorded as remaining performance obligations. That turns a high-margin, capital-light software company into an infrastructure provider with lower margins, heavy upfront spending and growing reliance on debt, a shift Section 6 examines in detail. [20]

At the frontier-model layer, the gap widens again. The laboratories generate genuine revenue from subscriptions, APIs, enterprise licences and consumer products. But the infrastructure and compute commitments being built around them are much larger than current operating cash flows.

Put plainly, there is a hierarchy. Some parts of the system can fund both investment and obligations from existing cash. Some can service the near-term burden but depend on refinancing or future contract conversion for the rest. At the most aggressive end, continued access to fresh capital is structurally important because current receipts are not sufficient to validate the full stock of commitments.

That hierarchy does more work than the word ‘bubble’ ever could. It tells us where disappointment can be absorbed and where it must be refinanced.

5. Contractual obligations matter more than headline valuations

The most revealing numbers in the next phase of the AI cycle are fixed commitments: capex, leases, purchase obligations, power contracts, debt service, minimum-use agreements and long-duration compute contracts.

The OpenAI ecosystem is the most striking example. On my count of publicly reported contracts, identified obligations associated with Oracle, Microsoft and CoreWeave approach roughly $540 billion, against about $73 billion of cash and securities and perhaps $36 billion of planned 2026 revenue.

Those are not directly comparable accounting categories, and the obligations do not fall due at once. They stretch across years and will be serviced partly by future receipts. But that caveat is exactly why the question matters: how much future income must arrive, and how much future financing must remain available, for the structure to validate itself?

The simplest way to think about it: compare fixed cash commitments with the cash receipts available to service them. When receipts comfortably exceed commitments, financing is robust. As commitments consume more cash flow, the margin of safety narrows. When obligations exceed what the business can support internally, the system becomes dependent on refinancing, new equity, asset sales or very rapid growth.

The risk is that perfectly productive assets can be financed with liabilities that need the future to show up on a timetable.

6. Oracle shows how fragility can enter through the marginal investment programme

Oracle is revealing because it sits between an old, cash-generative software business and a new, capital-hungry infrastructure business.

The demand is real. Remaining performance obligations have risen dramatically and cloud infrastructure is growing quickly. Yet trailing free cash flow stands at around negative $29 billion, after negative $23.7 billion for the fiscal year to May 2026, with guided capex approaching $90–95 billion and FY26 operating cash flow of roughly $32 billion. [21]

The company as a whole remains far more resilient than a standalone infrastructure start-up because its installed enterprise franchise produces substantial recurring cash. But that does not mean every incremental dollar of AI expansion is financed from internally generated cash.

The new build increasingly depends on debt markets, conversion of backlog into realised revenue, continued customer performance and the credit quality of a concentrated group of buyers. The point is worth dwelling on: financial fragility often enters a strong company through its marginal investment programme long before the legacy business itself looks weak.

So the real question is not ‘Is Oracle safe?’ It is ‘How much disappointment can the incremental AI programme absorb before investment has to slow, financing costs rise or contracts have to be re-priced?’

7. GPU finance hides a maturity mismatch inside the hardware

CoreWeave takes the problem further. Its assets are productive, demand is real and revenue is growing, yet the liability structure is much more dependent on refinancing. Total debt rose from about $21 billion at the end of 2025 to $35.6 billion by 30 June 2026. In the June quarter, net interest expense of $640 million was covered only about 0.2 times by adjusted operating income of $128 million and about 2.4 times by adjusted EBITDA of $1.5 billion, before the capital expenditure required to build and refresh the asset base. Those figures sit against genuine demand: quarterly revenue more than doubled to $2.6 billion and contracted backlog reached about $104 billion. But in the same quarter CoreWeave spent $6.4 billion on property and equipment against $679 million of operating cash flow, and it guided third-quarter interest expense to $860–940 million against $200–260 million of adjusted operating income. With about $10.6 billion of principal due by the end of 2027, the business depends on refinancing markets staying open. [22, 23, 24, 25]

The core problem is the mismatch between the economic life of a GPU and the duration of the liability attached to it.

A GPU can continue to function long after its ability to earn premium rents has deteriorated. Physical life and economic life are different things. If the loan runs longer than the high-value earning period of the asset, the lender becomes increasingly dependent on residual value, contract renewal and refinancing. The rental market shows why. H100s rented for $8–10 an hour in early 2024; by the second quarter of 2026 on-demand rates had fallen to roughly $1.80–3.50, even though one-year contract prices rose about 40% between October 2025 and March 2026 as capacity tightened again. Operators, meanwhile, typically depreciate GPUs over five to six years (CoreWeave uses six, extended from five in 2023) while Nvidia now releases a new generation roughly every year. Each generation lowers the rent its predecessor can command, so a GPU earns most of its premium in its first two or three years, while most of the loan is still outstanding. Utilisation matters as much as price: one estimate puts the all-in cost of an owned H100 at about $1.77 an hour at 70% utilisation and about $2.79 at 40%, so a modest fall in either rent or utilisation can erase the margin. [25, 26, 27]

A simple example shows the mechanics. An H100 bought for roughly $30,000 at 75% loan-to-value can still carry about $11,300 of debt by year three under a six-year structure, while a stressed collateral value could be only about $10,300. The sums are easy to check. At 75% loan-to-value the lender advances $22,500; under straight-line amortisation over six years, about $11,250 is still owed after year three. Yet used H100s now trade anywhere from about $6,000 to $22,000, against $25,000–40,000 new, as early three- to five-year contracts expire and hardware returns to the market. At the low end of that range the collateral covers barely half the remaining debt. Unless the loan amortises faster than the rent decays, or a creditworthy customer keeps paying under contract, the lender is relying on a resale price that each new chip generation pushes down. The machine may work perfectly and the financing can still become impaired. [28]

If I had to pick one mechanism to watch in this cycle, it would be this one. Nothing has to break. The cash the asset throws off just has to arrive more slowly than the debt requires.

8. Frontier laboratories are borrowing the future into the present

At the frontier-model layer, the mismatch is larger because enormous infrastructure commitments are being made ahead of the cash flows that will ultimately have to validate them.

OpenAI is the clearest case. As section 5 showed, roughly $540 billion of identified obligations sit against around $73 billion of cash and securities and perhaps $36 billion of annual revenue. Again, those obligations stretch across years; they are not a single near-term bill. Still, the cash OpenAI generates today is nowhere near enough to support the full stock of commitments without very rapid revenue growth and continued external finance.

Headline revenue figures need care, because the industry quotes three different numbers. Annualised run rate (ARR) usually takes the latest month's revenue (MRR) and multiplies it by twelve, so it describes a moment, not a year. Recognised revenue is what has actually been delivered and booked. Cash is what remains after paying for the compute that delivered it. In a business growing this fast the three can diverge sharply. OpenAI's run rate was reported at about $50 billion at the end of September 2026, after an earlier figure near $70 billion was corrected for counting partner revenue gross. Yet its recognised revenue was $5.7 billion in the first quarter and $6.7 billion in the second, an annual pace under $27 billion, and it reportedly burned $3.7 billion of cash in the first quarter alone. Leaked accounts for 2025 show an operating loss of about $20.9 billion on $13.1 billion of revenue. Even where growth is strongest, booked revenue trails the headline: Anthropic's run rate was reported at $65 billion in July, while it recognised $11.5 billion of revenue in the second quarter and $16.8 billion in the third. [29, 30]

The bridge from run rate to free cash has several further steps. Usage revenue consumes GPU time as it is earned, so gross margins sit well below those of conventional software. Compute commitments are largely fixed and often prepaid, while much usage revenue is billed only as it is consumed. And training the next model is a recurring cost, not a one-off. Run rate measures momentum. Only recognised revenue, less the cost of serving it and of staying at the frontier, measures the capacity to meet obligations.

For the model to work, several things have to stay true at once: revenue must compound quickly; equity markets must remain receptive; strategic investors must continue supplying capital; infrastructure partners must keep extending workable terms; and realised demand must remain strong enough to prevent large-scale renegotiation.

On current plans, I estimate a further funding requirement of roughly $160 billion in 2028–30. That turns the usual order of things around. In a conventional build-out, a company generates cash, proves demand and then adds capacity. Here, parts of the system are building capacity today against revenue that must materialise tomorrow. Backlog is where this shows up on the supplier side. A lab's compute obligation is its supplier's remaining performance obligation (RPO), the contracted revenue not yet delivered. Oracle reported RPO of $664 billion at 31 August 2026, of which roughly $300 billion is attributed to OpenAI. Microsoft said early in 2026 that about 45% of its $625 billion commercial RPO was tied to OpenAI. Backlog is contracted, but it is not cash. Oracle expects to recognise only about 13% of its RPO over the next twelve months, 37% in months 13 to 36 and 34% in months 37 to 60. Meanwhile, the supplier must build first: in fiscal 2026 Oracle's free cash flow was about negative $23.7 billion, and it issued $43 billion of senior notes and $5 billion of mandatory convertible preferred stock. Much of the industry's backlog therefore points back to the same small set of customers whose own revenue still has to grow into it. [21, 31, 32, 33]

In effect, the future is being pulled into the present. If demand turns up on schedule, that can work spectacularly. If the capital structure has little room for delay, it gets fragile fast.

None of this means the demand is imaginary. By most measures this is the fastest roll-out of a general-purpose technology on record. Stanford's 2026 AI Index finds that generative AI reached 53% population adoption within three years of ChatGPT's launch, faster than the personal computer or the internet. Revenue has followed at a pace with no precedent in software: Anthropic's run rate went from about $1 billion at the end of 2024 to $65 billion by July 2026, whereas Salesforce, long the benchmark for enterprise growth, took roughly a decade to reach its first $1 billion. [6, 30, 34]

But the J-curve is deeper as well as steeper. Earlier general-purpose technologies spread on capital that was largely built once and paid for over decades: rail, cables, power lines. AI must buy capacity ahead of revenue and then buy it again, because the chips age in three to six years and every new model has to be trained before it can be sold. So the trough runs to tens of billions, not millions. A draft of Anthropic's IPO prospectus reportedly shows an operating loss of more than $8 billion for 2025, even as revenue grew roughly twelvefold, and HSBC estimates that OpenAI will lose about $14 billion in 2026 alone. Their suppliers sit in the same trough: Oracle ran negative free cash flow of about $23.7 billion in fiscal 2026 to build capacity that its backlog says will be paid for later. [21, 35, 36]

What justifies spending through so deep a trough is the belief that the market is winner-take-most. Frontier models benefit from scale in compute, data, talent and distribution, and revenue is already concentrated in a handful of laboratories. If that is right, falling behind costs more than overspending, and every leading player has a rational reason to commit capital before the revenue exists. The difficulty is that the identity of the winner has not been stable. Anthropic's run rate overtook OpenAI's in the spring of 2026, little more than a year after it had been a sixth of the size. Capital is being committed as if the winners are known, in a market where leadership can change within months. That combination, a deep J-curve financed on the assumption of a durable winner, is what turns a fast technology roll-out into a fragile financial one. [37]

9. Some of the demand is becoming financially circular

A second source of fragility is the growing role of supplier capital. Some of the companies benefiting from AI demand are also helping finance the customers creating that demand.

The loop itself is not complicated. A supplier invests in or finances a frontier laboratory. The laboratory uses capital to buy compute. Those purchases become supplier revenue or backlog. Stronger reported demand supports earnings, valuations and financing capacity, which can then support additional investment into the same ecosystem.

By my rough estimate, supplier-related capital may represent 40–45% of laboratory inflows in some cases. The demand is still real: the compute is consumed; GPUs are delivered; data centres are built; contracts are booked.

The latest contracts show how tight the loop has become. In OpenAI's $122 billion round, which closed in March 2026, Nvidia and SoftBank each committed $30 billion and Amazon $50 billion. Nvidia pledged up to $10 billion and Microsoft up to $5 billion to Anthropic, which in turn committed to buy $30 billion of Azure compute. Amazon has invested $13 billion in Anthropic, with up to $20 billion more tied to milestones, against Anthropic's commitment to spend more than $100 billion on AWS over ten years. Nvidia is now moving from equity towards guarantees: it was reported in July to be weighing a $250 billion backstop for an OpenAI data centre in Ohio, although terms were unsettled. Its own filing for the quarter to 26 July 2026 shows $47.9 billion of private holdings, $25 billion more committed, and $36 billion of agreements to buy back AI cloud capacity from partners that bought its hardware. [38, 39, 40, 41]

Still, gross demand is one thing and independently financed demand is another. If part of a customer's purchasing power originates from the same ecosystem whose revenue that spending helps validate, the system can look more independently confirmed than it really is.

Margins explain why the loop exists. The profit pool sits upstream: Nvidia's gross margin was 75% in its latest quarter, TSMC's 66% and SK hynix's operating margin 76%. Downstream, the buyers earn far less. OpenAI's gross margin reportedly fell from 40% in 2024 to 33% in 2025, missing an internal target of 46%, as inference costs roughly quadrupled to $8.4 billion; Anthropic's was reportedly projected at about 40% for 2025, up from deeply negative a year earlier. Both are improving, and some analysts estimate that the most profitable inference work already earns far higher margins, but the gap remains wide. Supplier capital is, in effect, high-margin profit recycled into low-margin customers so that they can keep buying. That works while the customers' margins rise fast enough to stand on their own. [18, 19, 42]

If they do not, the damage travels through three channels at once. First, revenue: purchases funded by supplier capital slow when that capital slows. Second, balance sheets: equity stakes in the laboratories are marked down just as orders weaken. Third, contingent liabilities: guarantees and capacity buyback agreements that look dormant in good years become claims in bad ones. A supplier that finances its customers is therefore exposed to them three times over, and the headline revenue figure shows only the first.

The mechanism feeds on itself: higher expectations create financing; financing creates spending; spending creates revenue and backlog; those revenues validate the original expectations; the validated expectations support still more financing. The loop can persist for a long time. Its weakness is that it becomes sensitive to the availability of fresh capital.

10. Concentration turns company risk into system risk

Fragility becomes systemic when apparently separate balance sheets depend on the same final source of cash flow.

OpenAI alone supports roughly $300 billion of Oracle obligations, a meaningful portion of Microsoft's remaining performance obligations and approximately $22 billion of CoreWeave commitments. Those contracts then support debt issuance, data-centre construction, equipment purchases and GPU-backed financing. [32]

A financing problem at one frontier laboratory would not stay there. Slower contracting would reduce expected supplier cash flows. Lower expected cash flows would weaken the value of contracts. Lenders would demand wider spreads and lower loan-to-value ratios. Collateral values would come under pressure. Credit would tighten. Infrastructure capex would slow.

The systemic weakness is concentration: several balance sheets increasingly reference a small number of final demand nodes. The more concentrated the cash-flow source, the more quickly a local disappointment can become a capital-market event.

In practice, the test is margin of safety. Working in the base case proves little; a robust project is one that survives disappointment. Cash buffers, conservative leverage, diversified customers, rapid amortisation, strong collateral and long committed funding all create room for error. Fragile structures are those in which the expected case must happen almost exactly as planned.

11. The most important spread is return on new capital minus cost of funding

One more threshold deserves attention: the return earned on incremental AI infrastructure relative to the cost of financing it.

On my estimates, prospective returns on new AI infrastructure fall in a broad range of roughly 8–20%, against funding costs of about 6–10%. At the top end, the economics are exceptional: a 20% return financed at 6% creates an enormous spread. At the bottom end, the logic reverses: an 8% return financed at 10% destroys value before technological obsolescence, execution risk or cost overruns are considered.

The latest data moves both sides of that range in the wrong direction. On the funding side, the 10-year Treasury yield was 5.12% on 20 October 2026, up more than a percentage point in a year, and the investment-grade corporate index yielded 5.99% at the start of October, 121 basis points higher than a year earlier. AI-related issuers paid more than that: by September their credit spreads were around 115 basis points, against about 78 for the broader investment-grade market. Further down the stack, CoreWeave's term loans and senior notes carry effective rates of mostly 9–11%, and 15% on its oldest term loan. The bottom of that 6–10% funding range is now roughly where the strongest borrowers start. [43, 44, 45]

On the return side, Goldman Sachs estimates that realised hyperscaler return on invested capital has already fallen from 40% to 25%, and that the hyperscalers need roughly $300 billion of annual AI revenue simply to break even on their average 2026–27 capex. Reported margins have not yet absorbed the full cost. Microsoft's quarterly depreciation rose 55% to $9.0 billion in the March 2026 quarter, Alphabet's by 44% and Meta's server and network depreciation by 67%; across the four largest buyers, trailing depreciation was only about a third of capex, so most of today's spending reaches the income statement in 2027–29 whether or not the revenue arrives. AWS's operating margin has already slipped 140 basis points in a quarter, to 37.7%. [46, 47, 48]

That spread tells you more than any amount of excitement about the technology. Once realised infrastructure returns fall below marginal funding costs, continued expansion must be justified by future growth, refinancing, higher residual values or eventual capital appreciation rather than the economics of the asset itself.

There are signs that capital allocation is becoming less price-sensitive. Companies are accepting scarce allocations, prepaying suppliers, expanding capex budgets and tolerating higher input costs even as estimated returns compress. That can be rational when strategic value is high and capacity is scarce. It also suggests that the fear of missing out on capacity is beginning to compete with ordinary marginal-return discipline. The evidence is visible in 2026 guidance. Meta raised its capex range to $125–145 billion in April, and then to $130–145 billion in July, mainly because of higher memory prices and data centre costs rather than more capacity. Nvidia guided its gross margin down to about 74% as memory and wafer prices rose. In Oracle's largest AI contracts, customers either prepay so that Oracle can buy the GPUs or buy the GPUs themselves. Buyers are paying more for the same unit of capacity, and paying earlier. [49]

The warning sign is rising capex while expected returns shrink and funding gets dearer.

On current data, all three conditions are present at once. Goldman Sachs expects the largest US hyperscalers to spend $1.2 trillion on capex in 2027 and $1.4 trillion in 2028; returns on their invested capital are falling; and long-term borrowing costs are the highest in years. That does not mean the spread is negative everywhere. A hyperscaler earning 25% on capital and borrowing at around 6% still has a wide cushion, and Goldman itself argues the hyperscalers are almost certainly earning more than a 15% baseline on existing investment. The thinner margin sits further out. A neocloud whose adjusted operating margin was 5% last quarter, funding itself at 9–15%, has very little spread to lose. There is no single spread for the industry: it is widest where balance sheets are strongest, and thinnest where leverage is highest. That is where the warning signal should be watched first. [23, 46]

12. Credit markets are already drawing a sharper map than the equity debate

Markets are increasingly distinguishing among AI exposures by financing quality rather than by the AI label itself.

The core megacap equities do not look like the centre of a classic mania. Nvidia's trailing valuation multiple has fallen materially even as earnings expanded. Credit markets, meanwhile, have become more discriminatory: CoreWeave financing has repriced sharply, Oracle credit protection has widened, while hyperscaler debt has moved much less. [15]

The lesson is blunt. A Microsoft balance sheet, an Oracle balance sheet, a CoreWeave financing structure and a frontier-laboratory commitment are not economically equivalent merely because all four sit inside the AI ecosystem.

Credit and property markets also seem to disagree, and the disagreement may matter. Weaker AI-linked credits are demanding materially more spread while hyperscale data-centre property cap rates have tightened. The same underlying tenancy can therefore appear riskier to credit investors and safer to property investors.

That divergence is worth watching. If credit demands a larger margin of safety while property capital accepts a smaller one, one market is likely to be too optimistic.

13. We are not at the bust. We are at the point where a bust has become possible

The current system does not yet look like a broad forced deleveraging. Refinancing markets remain open. Infrastructure prices have not collapsed. Defaults are not systemic. Capex is still accelerating rather than being cut.

So we are at the stage where the financial plumbing could carry a bust.

That matters, because financial reversals usually begin when assumptions that were previously harmless become binding: a refinancing takes longer, a funding round becomes more expensive, a customer delays a commitment, a collateral value falls, or lenders reduce advance rates.

The system can remain apparently healthy long after fragility has begun accumulating underneath it. That is why I would watch the next refinancing cycle more closely than today's equity multiple.

Part of the apparent health is just timing. Most of the indicators people watch lag. Depreciation reaches the income statement years after the cash has been spent, so reported margins flatter a build-out while it is under way. Fixed-rate debt raised when money was cheap keeps today's interest bill below today's cost of borrowing. Backlog records a customer's promise to pay, not the payment. Defaults and write-downs come last of all, because a borrower that can still raise new money rarely misses a payment. Each of these can look sound right up to the point at which the old terms expire.

An equity multiple is a poor guide for a different reason. It measures how investors feel about future earnings, and it can move in either direction without anyone having to pay anything. Nvidia shows how loose the link is: its trailing multiple roughly halved over the past year while its earnings more than doubled. Shares have no maturity date, so an expensive or cheap valuation can persist for years. Debt does have a date. When a loan or bond comes due, a lender has to decide afresh, at today's interest rates, today's collateral values and today's view of the customer, whether to lend again and on what terms. A refinancing is where the market is forced to re-price the asset, rather than simply re-rate the company. [15]

That test is already on its way. CoreWeave has about $10.6 billion of principal due by the end of 2027, already pays 9–15% on most of its term debt, and has guided third-quarter interest of $860–940 million against $200–260 million of adjusted operating income. Some of that debt is secured on GPUs whose used market price has fallen to $6,000–22,000, from $25,000–40,000 new. Even if demand holds, the next loan is likely to be smaller relative to the collateral and more expensive. Early signs are already visible in the bond market. Demand for new hyperscaler bonds fell from about five times the amount on offer in February 2026 to about two times in July. Of 91 hyperscaler bonds issued in 2026, 78 were trading at higher yields than when they were issued. Close to half of 2026's data-centre bonds are rated junk. In the week to 16 October, about $4 billion of CoreWeave's notes traded below 80 cents on the dollar. [8, 20, 23, 24, 28, 50]

None of this makes a bust inevitable. A strong borrower facing a harder refinancing can simply roll its debt at a higher cost, and the hyperscalers can still do so comfortably. The real question is whether the cost of rolling the debt rises faster than the return on the assets it finances, the spread described in section 11. Where that spread is wide, a tougher refinancing is an expense. Where it is thin, it forces a choice between raising equity, selling assets or shrinking the build-out. And where new capital keeps arriving to cover the gap, the fragility is only postponed, and the next refinancing becomes the bigger test.

The years 2027–28 look like a critical window because existing commitments, future funding needs and asset refinancing begin to overlap more meaningfully. What counts is the mechanism: yesterday's commitments must eventually be validated by tomorrow's cash or tomorrow's capital.

14. The cleanest stress test is to close external financing

One way to tell robust businesses from fragile financing structures is to run four simple shocks and see what breaks.

First, raise funding costs by 300 basis points. The strongest hyperscalers absorb the increase comfortably. Oracle faces more free-cash-flow pressure. CoreWeave becomes substantially more exposed because coverage is already thinner. Frontier laboratories face more expensive future raises.

Second, assume AI revenue comes in 20% below plan. Semiconductor margins compress, but the strongest balance sheets remain viable. Hyperscalers retain diversified cash flows. The expected return on newer infrastructure falls, utilisation assumptions weaken and the financing gap at frontier laboratories widens.

Third, reduce GPU collateral values by 30%. Short-amortising loans backed by strong contracts can absorb more of the shock. Longer-duration debt that relies on residual value cannot. Loan-to-value rises, covenant pressure grows and asset sales can push second-hand prices down further.

The fourth test tells you the most: close external financing for twenty-four months. Microsoft, Meta, Amazon and Alphabet still function. Oracle survives but would probably have to reduce the pace of investment materially. CoreWeave becomes vulnerable. The current investment trajectory of frontier laboratories becomes extremely difficult to sustain.

That is the line that matters. Operating risk and dependence on capital markets are separate risks. A business that can survive without new capital occupies a fundamentally different financial category from one whose existing commitments assume that capital markets remain open.

15. What would the actual reversal look like?

A 20% fall in Nvidia, or even a sharp correction in AI equities, would be a symptom at most. The real break begins when refinancing assumptions that previously looked routine stop validating commitments already written into the system.

The sequence would probably begin at the frontier. Revenue growth at a major model company disappoints. The next mega-round becomes harder or more expensive. The laboratory slows new compute commitments. Oracle and CoreWeave bookings soften. Existing backlogs look less bankable. Lenders raise spreads and reduce loan-to-value ratios. GPU refinancing becomes harder. Residual values fall. Collateral coverage deteriorates. Borrowers sell assets. Used-GPU prices weaken further. Lenders tighten again. New data-centre capex falls. Semiconductor orders eventually slow.

Each step feeds the next: falling asset prices reduce collateral values; weaker collateral reduces credit availability; tighter credit slows investment; slower investment weakens cash flow; weaker cash flow creates more forced selling; forced selling pushes asset prices down again.

The trigger can be as mundane as the next financing round becoming materially harder to complete.

So the question that matters for the next phase of the boom is whether the financing architecture can tolerate demand arriving later, returns arriving lower or capital arriving at a higher price than today's commitments assume.

Conclusion: AI can win and today's investors can still lose

The strongest case against calling AI a simple bubble is also the strongest reason to take its financial risks seriously. The technology works. Demand is real. Productivity gains are plausible. Infrastructure has genuine scarcity value. Yet none of those facts guarantees that every financial claim written against the technology will earn an attractive return.

If history is any guide, this is the normal pattern, not the odd one out. Three earlier build-outs show the same gap between being right about a technology and being paid for it, and each failed in a slightly different place.

The railways are the oldest example. In Britain's Railway Mania of the mid-1840s, Parliament authorised thousands of miles of new line in a few years, and investors bought shares by paying only part of the price up front, with the rest due on call. When money tightened in 1847, the calls came due, and share prices collapsed. In the United States, the Panic of 1873 began with the failure of Jay Cooke & Company, the bank that had been unable to sell enough Northern Pacific bonds; two decades later, during the depression of the 1890s, a large share of the country's track passed through receivership. In both countries the railways were built, used and transformative. They were reorganised, not torn up. What failed was the liability side: fixed payments on bonds and calls fell due before the traffic arrived to pay them. [51]

The fibre boom of the late 1990s is the closest modern parallel. After the US Telecommunications Act of 1996, telecom companies invested more than $500 billion, much of it borrowed, laying fibre-optic cable. The case rested on claims that internet traffic was doubling every 100 days, when it was closer to doubling every year. By 2001–02, estimates suggested that 85–97% of the fibre laid was dark, and bandwidth prices fell by something like 90%. Global Crossing filed for bankruptcy in January 2002 with $12.4 billion of debt, and WorldCom followed in July as the largest US bankruptcy to that date, brought down partly by fraud. Carriers had also traded capacity with one another and booked the swaps as revenue on both sides, an early version of the circularity described in section 9. Yet the fibre found its use. It was bought cheaply out of bankruptcy and later carried broadband, streaming and the cloud. The builders were gone by then; the buyers collected the return. [52, 53]

The Internet of Things shows a different outcome. Around 2011, Ericsson and Cisco forecast 50 billion connected devices by 2020; estimates for 2020 came in closer to 9 billion. The forecast missed by more than 80%, yet there was no financial crisis. Most IoT spending came from corporate operating budgets and venture equity rather than long-dated debt secured on the assets, so the miss produced disappointing returns and write-offs, not forced selling. Overestimated demand becomes systemic only when it is financed with fixed obligations. [54]

AI resembles all three, but not exactly. Like the railways and fibre, capacity is being financed with debt ahead of the revenue it is meant to earn, and some demand is circular. Unlike fibre, today's core builders are cash-rich companies rather than WorldCom-style start-ups, and today's capacity is being used rather than sitting dark. The largest difference cuts the other way. Rails and fibre lasted for decades, so their overbuilt capacity was still valuable to whoever bought it after the crash. GPUs lose most of their value in three to six years. If demand arrives late, there may be no later owner to inherit cheap, useful capacity; the overhang simply ages out. The lesson, then, is that a technology can win while the financing built around its first wave loses.

It also helps to stop treating AI as a single cycle, because its parts are moving at different speeds. Frontier models remain experimental and intensely competitive. GPUs are already a major infrastructure industry, and memory has shown outright scarcity economics. Data centres are moving from construction into increasingly aggressive financial expansion, while power and grids are only beginning to catch up. Agents are early, enterprise applications are still searching for dominant designs, and the redesign of corporate workflows has barely begun. Governance is lagging, and economy-wide productivity effects remain far less developed than headline adoption rates suggest. Put directly: the technological system is still being built, while some of the financing around that build already behaves as if the economic payoff were proven.

The earlier sections trace exactly this transition. At first, finance follows the technology: real profits attract capital, capacity gets built and new entrants arrive. Later, the sequence runs from expected demand to capital raising to construction, and finance begins to set the pace of the build. Outside the cash-rich core, the build now depends on continuous external capital; on the tally in section 3, roughly $590 billion was raised in 2026 across hyperscaler bonds, OpenAI and Anthropic equity, private credit and SPVs, Oracle and CoreWeave. The new structures, from GPU-backed loans and data-centre SPVs to minimum-use commitments and supplier investments, are not excessive in themselves; new assets often need new ways of underwriting them. But each one makes it easier to spend against expected demand rather than current cash flow, and so widens the gap between what has been built and what has been economically justified.

It is also why people may be looking for the bubble in the wrong place. Nvidia's earnings have grown faster than its share price, so its valuation multiple has compressed, which is hardly the picture of a classic equity mania. The speculative signal lies in the quantity of capital being committed against future demand. Excitement has moved from models to GPUs, then to memory, data centres and now power and financing structures, following whichever input is scarcest. The danger is that physical capacity is being financed on the assumption that today's scarcity will last.

The hierarchy from section 4 still holds: a cash-rich core, a more refinancing-dependent infrastructure middle and a frontier layer whose current ambitions depend heavily on future capital and future revenue. The risk rises as we move away from existing cash flow and toward commitments that require the future to arrive on schedule.

That does not imply imminent failure. Strong revenue growth could pull weaker structures back toward self-funding. Equity issuance can repair balance sheets. Contracts can be renegotiated. Assets can be refinanced. A fragile structure can become robust if the expected income actually arrives.

But the burden of proof has changed. The continuation of the boom increasingly depends on AI succeeding and on financial markets continuing to fund the consequences of that success.

The shift boils down to two questions. The first is what future cash flow must arrive to validate today's financing structure. The second is what happens if capital markets close before that cash flow arrives. Asked that way, the debate breaks into a sequence of tests, and the answers change as one moves down it.

Technology: does AI work? Yes.

Demand: are people and companies buying it? Yes.

Economic value: does it create substantial value? Very likely.

Return: does the capital invested earn an adequate return? Increasingly uncertain.

Financing: can today's cash flows service today's commitments? Only in parts of the system.

Refinancing: can the weaker layers survive without new capital? No.

Systemic: would an interruption in financing spread through connected balance sheets? Increasingly, yes.

The first three tests concern the technology, and AI passes them. The last four concern the money, and that is where the answers turn. On the current evidence, this is a late-stage capital boom, not a bust. The technology and the demand are real, but fragility is moving outward: from a core that can pay both interest and principal out of operating cash flow, into an infrastructure layer that can cover its interest but must refinance its principal, and towards a frontier apex that depends on fresh external capital to cover even its running costs. The moment of reversal has not arrived. Credit is being repriced rather than withdrawn, and no one has yet been forced to sell assets to meet obligations. The decisive test is likely to be the 2027–28 refinancing cycle, not today's level of AI equity valuations.

History also separates the people who pay for a build from the people who profit from it. Today the cost of discovering what AI is ultimately good for is spread across hyperscaler shareholders, bondholders, venture and sovereign funds, private-credit lenders, strategic suppliers and infrastructure investors. The investors who funded railway and fibre overcapacity were rarely the ones who later benefited from cheap transport or abundant bandwidth. The question for AI is who is paying for the experimentation today, and who will own the productive system once it is over. They may be very different people.

The reason is that overbuilding can destroy capital without destroying value. Suppose a $10 billion data centre earns 4% on its cost while its funding costs 9%. The investment is a failure. But if the debt is restructured and the facility changes hands for $4 billion, the same income is a 10% return for the new owner. Compute becomes cheaper, and uses that could not justify the original cost can justify the distressed one. Private losses can sit alongside wider economic gains. The caveat from the history above still applies: the building, the power connection and the cooling last for decades, but the chips inside do not. In an AI overbuild, the durable shell may survive a restructuring while much of the value tied up in GPUs is simply written off.

Taken together, four things are true at once: the technological breakthrough is real, the construction is enormous, productive deployment is incomplete, and finance is increasingly setting the speed of the build. The technology works, and society will almost certainly find uses for cheaper machine intelligence. The open question is whether productive demand can catch up with the infrastructure and financial commitments already built in anticipation of it. Once installed capital starts to outrun the cash flows able to support it, the problem stops being primarily a technology question and becomes a financing question.

My conclusion, then, is this: if the AI cycle breaks, it is more likely to break first through the liability side of the balance sheet than through the technology or demand side of the income statement.

The technology can continue improving. Usage can continue growing. AI can transform the economy. And investors can still lose enormous amounts of money. There is no contradiction in that.

Note: Figures are drawn from company filings, market data, published research and press reporting as of October 2026. Some are estimates, and are used to illustrate the financing mechanics rather than as precise forecasts. This essay draws on and forms part of the author's AI Boombustology series [55, 56, 57, 58].