The question arrives roughly weekly and is almost always posed as a request for a verdict. Is this a bubble. The people asking are not being unreasonable, because the numbers involved are large enough that an ordinary saver with a pension has a legitimate interest in the answer.
The problem is that bubble is a conclusion rather than an analysis, and conclusions delivered without a method are worth precisely as much as the confidence of whoever delivers them. Every previous cycle was called a bubble by some observers and a new paradigm by others, and in most cases both descriptions turned out to be partly accurate, which tells you the terms are doing very little work.
There is a more useful approach. Five structural conditions distinguish a capital expenditure supercycle, which wastes money and self corrects, from a credit bubble, which wastes money and damages people who were not involved. Each condition can be assessed against public disclosure. Applied honestly, they produce a mixed answer, and the shape of the mix is considerably more informative than a verdict.
Test One: Where Does the Money Come From
This is the test that matters most, and it is frequently skipped entirely in favour of arguing about valuation.
Overinvestment funded from a company's own operating cash flow produces a specific and contained outcome. The capital is wasted, returns disappoint, assets are written down, and shareholders bear the loss. That is genuinely unpleasant for the shareholders and largely irrelevant to everybody else. No lender is impaired, no credit contraction follows, and the wider economy continues.
Overinvestment funded with borrowed money produces a different outcome entirely. When asset values fall below the debt written against them, lenders take losses, lenders that take losses reduce credit to unrelated borrowers, and businesses with no involvement in the original cycle find themselves unable to refinance. This is the mechanism by which a sectoral disappointment becomes a general contraction, and it is the only reason anybody outside the sector should care.
On this test the current build out passes clearly at the top of the stack. The largest purchasers of artificial intelligence compute are mature, highly profitable companies funding these purchases substantially from operating cash flow. If every dollar of that spending turned out to be wasted, the consequence would be a severe repricing of some very large equities and no banking crisis.
It does not pass as cleanly further down. The compute lessor layer is debt funded by design, and some model laboratories are funded by capital raising rather than by operations. The build out is therefore not uniformly equity funded, and the parts that are not are the parts to examine.
Overinvestment funded with your own cash costs your shareholders. Overinvestment funded with somebody else's costs everybody.
Test Two: Do the Buyers Convert Earnings Into Cash
The second test asks whether reported growth is being earned or financed, and it is answered by comparing reported earnings with operating cash flow over several periods.
A company whose profits arrive as cash is generating real economic surplus. A company reporting strong profits while operating cash flow lags persistently is recognising something the cash has not confirmed, and the gap is where accounting judgement lives.
The hyperscale buyers pass this test decisively. Their cash conversion is high, has been high for a long time, and comes from mature businesses unrelated to artificial intelligence. This is the second strong argument against the bubble framing, and it deserves to be stated plainly rather than buried.
The model laboratory layer is a different matter, and the honest position is that outside observers cannot fully assess it. Most of the relevant entities are private, disclosure is partial, and the revenue figures that circulate are frequently annualised run rates rather than audited results. What can be said is that revenue at this layer is real, growing quickly, and generally insufficient to fund compute purchases at the scale being contracted, which is precisely why the circular financing arrangements described in the companion piece exist at all.
Test Three: How Long Do the Assets Actually Earn
The third test asks whether what is being built will still be productive when the financing against it comes due, and this is where the sector splits in two.
Roughly half of the physical investment consists of things with very long lives. Buildings last decades. Electrical substations, transmission connections and grid interconnection rights last decades and are becoming scarcer rather than more abundant. Cooling infrastructure, networking and fibre routes have long lives and broad applicability. If demand for artificial intelligence compute disappointed severely, these assets would be repurposed for other computing workloads, and in a constrained grid the interconnection rights alone would retain considerable value.
The other half is accelerators, and their competitive life is short. Not their physical life, which is long, but the period over which they command their original rental economics. Each new generation improves performance per watt and per dollar substantially, and pricing for previous generations adjusts accordingly within quarters.
The result is an asset base with a genuinely bifurcated durability profile: long lived infrastructure wrapped around short lived equipment, financed as though the whole were homogeneous. That is the most legitimate structural criticism available, and no amount of enthusiasm about end demand addresses it, because it is a statement about schedules rather than about value.
Test Four: Are the Contracts Enforceable and Renewable
The fourth test concerns collateral quality. Take or pay contracts are what make the debt in this sector possible, so what those contracts are actually worth is the whole of the credit question.
Two properties matter. The first is enforceability, meaning whether the customer is contractually obliged to pay regardless of usage, and whether that obligation would survive commercial pressure. The second is renewal behaviour, meaning what happens at expiry, since the value of a lease book beyond the initial term depends entirely on renewals at comparable prices.
Enforceability is generally strong on paper. Renewal is where the uncertainty concentrates, and it is not a legal question. A large anchor tenant approaching renewal with newer hardware available elsewhere at lower cost has considerable negotiating leverage, and lessors dependent on that tenant have very little. The likely outcome in a softer market is not default. It is renegotiation, and renegotiation is the event that reprices collateral across every comparable book simultaneously.
This test therefore returns a mixed result. The contracts are good, the counterparties are largely good, and the renewal assumptions embedded in the financing are the part that has not yet been tested by anything other than a rising market.
Test Five: How Dependent Is the Structure on Refinancing
The fifth test is the one that converts a slowdown into a failure, and it is the reason the compute lessor layer matters so disproportionately.
A borrower that can service debt from operations is exposed to its own business. A borrower that must raise new debt in order to repay existing debt is exposed to the mood of the credit market on a particular date, which is a risk it does not control and cannot hedge. This distinction has determined the outcome of every credit cycle in modern history.
The hyperscale buyers have no meaningful refinancing dependence for this spending. The lessor layer has substantial refinancing dependence by construction, because the business model requires debt and the assets are shorter lived than the schedules.
This is the fragile joint. Not the technology, not the valuations, not the enthusiasm. A layer of the market that must return to credit markets on a schedule, holding collateral whose value depends on renewal assumptions that have not been stress tested. The simulator in this piece exists to make that sensitivity concrete: adjust leverage, contract coverage and demand, and the point at which debt service ceases to be covered appears immediately.
Set the terms of one financing cycle. The model traces the money around the loop and reports how much of it was ever external, and which participant runs out of room first.
38%
of compute spend funded by the seller of the compute
$0.62
genuinely new money behind each headline dollar
1.70x
$30.7bn revenue against $18.1bn obligations
41%
fall in revenue the lessor absorbs before it cannot pay
Loop is self supporting. External cash dominates and the lessor covers obligations with room to spare.
Fixed assumptions, stated so they can be argued with: 75 per cent of third party capital becomes compute spend, hardware debt costs 8 per cent blended, hardware depreciates over four years, and capacity revenue runs at 42 per cent of hardware cost annually at full utilisation. Outputs are arithmetic consequences of the inputs above. This is an illustrative structural model and not investment advice.
The Honest Scorecard
Assembling the five tests produces a picture that neither camp will find fully satisfying, which is usually a sign that the assessment is close to right.
On funding source, the build out passes at the top of the stack and is mixed lower down. On cash conversion, the hyperscale buyers pass decisively and the model laboratory layer cannot be fully assessed from outside. On asset life, the result is genuinely split between long lived infrastructure and short lived equipment. On contract enforceability, the contracts are sound and the renewal assumptions are untested. On refinancing dependence, the top of the stack is insulated and the lessor layer is exposed.
Two clear passes, two mixed results and one identified strain point. That is not the profile of a credit bubble, which requires debt funded purchases by buyers without cash flow against collateral nobody wants. It is the profile of a genuine capital expenditure supercycle containing one credit sensitive joint, which is a much more specific and much more useful conclusion than a verdict.
What History Actually Suggests
The most frequently cited precedent is the fibre optic build out of the late 1990s, and it is cited by both sides, which should be a clue that its lesson is being simplified.
Enormous quantities of optical cable were laid in anticipation of internet traffic growth. The traffic growth arrived, exceeded even aggressive forecasts, and continues today. The companies that laid the cable were nonetheless destroyed, because the debt matured in 2002 and the traffic arrived in 2008. Subsequent owners acquired the assets at a fraction of construction cost and built profitable businesses on infrastructure somebody else paid for.
The lesson is not that the demand thesis was wrong. It is that being right about demand and wrong about timing produces the same outcome as being wrong about demand, if you are the one holding the debt. Infrastructure gets built, capital gets destroyed, and society gets the asset. That has been the pattern for railways, canals, electrification and fibre, and there is little reason to expect artificial intelligence infrastructure to break it.
What differs materially this time is who is paying. In the fibre cycle the builders were leveraged new entrants with no other business. In this cycle much of the spending comes from companies with enormous existing cash flows and diversified operations, which changes the credit consequences even if it does not change the possibility of overshoot.
What to Watch Instead of Predicting
Four indicators carry more information than any valuation multiple, and all are publicly observable.
The first is the terms on which new contract backed debt is issued: spreads, tenors, required equity contribution and covenant tightness. Credit markets adjust before equity narratives do, and this is the earliest honest signal available.
The second is capital expenditure guidance from the largest data centre buyers. The infrastructure layer is priced against the slope of that spending rather than against end demand, so a deceleration in guidance reprices the supply chain long before anything changes in usage.
The third is any change in depreciation policy or impairment at the lessor layer, which would confirm that residual values are moving.
The fourth is the gap between reported earnings and operating cash flow at the model laboratory layer, to the extent disclosure permits, because that is where financed growth would first become visible.
The Position Worth Holding
The useful answer to the bubble question is not yes or no. It is that this is a real capital cycle building genuinely useful infrastructure, funded predominantly by entities that can afford it, containing one identifiable financing structure that is more fragile than the rest and that would be the first thing to break.
That position accommodates both of the observable facts that the maximalist and the sceptic each hold half of. Real demand exists and is growing. Real financing strain exists in a specific layer. Neither cancels the other, and an investor who understands where the strain sits is considerably better equipped than one who has simply chosen a side.