The AI buildout gets riskier if it works.

In the second quarter Alphabet spent $44.9bn on capital expenditure, roughly double a year earlier, against $39.1bn generated by operations. The result was negative free cash flow of $5.9bn, the first cash-negative quarter since the company listed in 2004. The CFO raised full-year capex guidance to between $195bn and $205bn at the same time.

Keep it in proportion. On a trailing twelve-month basis Alphabet's free cash flow remains positive at $53.3bn, and one negative quarter against those sums is a choice rather than a constraint. Nothing here is distress.

What makes it worth a closer look is that three separate forces are now working on these accounts, and none of them requires demand for AI to disappoint. They arrive because of when equipment was bought and how accounting rules recognise it. That is a different kind of risk from the one usually discussed, and it is more predictable.

The first force: depreciation arriving late

Buying a server does not reduce profit. The cash leaves immediately, but the cost is spread across the years the machine is expected to earn. That is what depreciation is. Buy a $100,000 accelerator and assume it works for five years, and roughly $20,000 hits the income statement each year, whatever the machine actually did.

Two things follow. The profit hit arrives years after the decision to spend, and the size of it depends entirely on a guess about how long the hardware lasts.

Both are now visible. Meta's revenue grew 28% last quarter to $60.8bn while its operating margin fell from 43% to 31%, with depreciation and amortisation up 46% year on year to $6.4bn. The honest qualifier is that legal charges and severance sit inside that margin move, and excluding them operating income would have grown 9%, so depreciation is not carrying the whole decline. Anyone presenting 43-to-31 as a pure depreciation effect is overstating it.

But the direction is not in dispute, and it is the most mechanical of the three forces. Equipment bought in 2024 and 2025 is only now reaching income statements, on schedules fixed at the moment of purchase. The cost has already been incurred. The only open question is which quarters absorb it.

Which brings up the guess. The large hyperscalers moved to six-year useful lives for servers from 2023; several specialist operators run shorter policies. The informative data point is not either camp's advocacy but a disagreement: in 2025 Amazon shortened the assumed life for part of its fleet while Meta extended. Two companies looking at broadly similar hardware reached opposite conclusions, which is a reasonable sign that nobody knows.

That guess carries more weight than it appears to. Goldman Sachs, modelling cumulative AI capital expenditure at roughly $7.6tn across 2026 to 2031, identifies the useful life of silicon as the single largest variable in its own model, larger than demand. The reason is that the assumption does not just set the annual charge; it sets how often everything has to be replaced. Assume four years instead of six and the same fleet needs buying again a third sooner, which moves cumulative spending by hundreds of billions.

The loop this creates runs the wrong way. A shorter life means more replacement spending, which means a faster annual depreciation charge, which raises the return the spending has to earn to be worth doing. The assumption that makes the accounting look better is the same one that makes the replacement bill arrive later, and a company cannot choose both.

The second force: leases that are not on the balance sheet

This one needs the accounting explained before the number means anything.

When a company signs a lease, it does not immediately record a debt. Under GAAP the liability appears when the lease commences, meaning when the landlord hands over a building that is finished and connected and the tenant starts paying. Signing is a promise. Commencement is when the promise becomes a number on the balance sheet.

For an office lease that gap is a few months and nobody cares. For a data centre it is years, because the building has to be constructed and, more slowly, connected to the grid. So a company can commit today to two decades of rent on a facility that does not exist, and show nothing.

Moody's counted the size of that gap: $662bn of leases signed but not yet begun across the five largest US hyperscalers. Total future lease commitments at the end of 2025 were $969bn, so more than two thirds of what these companies have promised to pay is not on any balance sheet a shareholder can read.

The timing is the part that matters. Leases signed through 2026 commence largely into 2028 and 2029. That is when they convert from invisible promises into recorded liabilities with cash rent attached, and it is the same window in which the growth rate used to justify signing them is most likely to have flattened. The obligations peak on a schedule fixed years earlier, independent of how the business is doing when they land.

A second layer sits underneath. The standards leave room on whether renewal options and backstop payments get disclosed, so $662bn is a floor rather than a measurement. Residual value guarantees, which are promises to make a landlord whole if an asset is worth less than assumed at handback, become liabilities only once payment is judged probable, and most have not crossed that line.

None of this is concealment. It is the rule working exactly as written, and auditors sign it. But it does mean that the question every credit analyst asks first, how much does this company owe, cannot be answered from the balance sheet alone. The invisible portion is large relative to the debt these companies do report.

The third force: the tax reversal

The tax books and the accounting books are allowed to disagree, and here they do so deliberately.

Full expensing lets a company deduct the entire cost of qualifying equipment from taxable income in the year it is bought, instead of spreading the deduction across the asset's life the way depreciation spreads the accounting charge. So the same server is a five-year cost to shareholders and a one-year deduction to the tax authority.

While purchases are rising, that keeps cash taxes low, because each year of larger spending generates a larger immediate deduction. The shield grows with the buying.

It works in reverse when the buying stops growing. A plateau produces no new deduction, while the deductions already taken finish unwinding, so cash taxes normalise sharply at precisely the moment capital spending flattens. Nothing has to go wrong for this to happen. It is the arithmetic of the same rule running the other way.

Which puts the reported numbers in an awkward light. The cash flow declines already visible are occurring while cash taxes are being held down by a shield that expires. The underlying picture is worse than the reported one, not better.

Why success is the problem

The usual question is whether the buildout will fail. The more interesting question is what happens if it succeeds.

Scarcity is currently holding up three separate things at the same time. It supports pricing power for compute. It supports the rental rates that specialist providers underwrite their borrowing against. And it supports the residual values that lenders are treating as collateral.

All three depend on there not being enough capacity.

Building enough capacity ends the scarcity, which is the entire purpose of building. Not building leaves capacity that has already been financed against a schedule that assumed it would be there. There is no configuration in which both sides work at once.

This is not a prediction that anything fails. It is the observation that in this cycle, success and safety point in different directions. That is unusual, and it is the part of the structure least discussed.

The collateral nobody is tracking

There are three prices for a GPU and they behave differently.

The list price is flat. The rental rate is stable to slowly declining, on the order of fifteen percent a year for previous generations. The resale value is chaotic, and it is the one that matters for credit, because resale is what a lender recovers.

The tracked price is the first. The traded price, once compute futures list, will be the second. The collateral is the third.

How wide is the gap? Used A100 80GB cards have been observed trading anywhere between roughly $4,800 and $18,900, close to four times the price for a card with the same name on it. The variation is not noise. A card that has spent two years at full load training models is a different asset from one that idled in a research cluster, and operating hours matter more than calendar age. That history is rarely documented and almost never verified.

So a pool of ten thousand such cards is worth somewhere between about $48m and $189m depending on facts the lender cannot see. Now recall what an advance rate is: the share of an asset's value a lender is willing to lend against, and the buffer that protects them if they have to seize and sell. Those rates were set against new-card prices. If the recovery value is a quarter of what the advance assumed, the buffer is not thin. It is gone, and the loan was effectively unsecured from the start without anyone intending that.

This also explains an argument that keeps going in circles. Both the optimistic and pessimistic figures in circulation are true, because they measure different curves. When an operator reports re-contracting expiring hardware at close to its original price, that is a rental figure, meaning someone agreed to keep renting it. When secondary market trackers report steep declines, that is resale, meaning what someone paid to own it outright. Renting and owning are different transactions and there is no reason their prices should track. Quoting one against the other is a category error, and it happens constantly in both directions.

The next scheduled test is the arrival of the following GPU generation in late 2026 into 2027, which will mark the current one down as buyers wait for the newer part. That lands before the refinancing concentration of 2027 to 2029, not after it, so collateral gets revalued while the debt against it is being rolled.

What the ratings actually say

Some of this borrowing gets packaged into bonds and sold on. That process is securitisation: a pool of assets and the rent they generate is placed into a vehicle, the vehicle issues bonds against it in layers, and rating agencies grade each layer. Investors buy the grade, not the asset. So what the agencies think the collateral is worth determines how cheaply the whole thing funds.

Which makes one distinction essential, and it is the one most commentary gets wrong.

Data centre securitisations are backed by real estate: land, buildings, mortgages, and the assignment of the customer contracts attached to them. GPU-backed lending is a separate market where the collateral is the hardware itself. Buildings and grid connections hold value; silicon does not. So the securitised paper is the sounder structure of the two, and the pressure sits in the GPU lending. Treating them as one thing overstates the risk in one place and understates it in the other.

Within the securitisation market, the disagreement between agencies is observable rather than inferred, which is rare. S&P has never rated a data centre tranche above A: of seventy rated between 2020 and 2025, forty-two sit at A− and twelve at A, and none higher. Moody's and Fitch have gone to triple-A. In July 2026 a Cloud Capital deal became the first to carry triple-A from three agencies at once.

One transaction shows it cleanly. In CLDHQ 2026-1, notes inside the same class drew Aaa and AAA from two agencies and A from two others. Same deal, same seniority, four notches apart depending on who was asked. When agencies disagree that widely about identical paper, the rating is telling you as much about methodology as about the asset.

Then there is a horizon mismatch that takes a moment to see. These bonds carry a legal final maturity of thirty or thirty-five years, and that is the date agencies rate against. But they also carry an anticipated repayment date, usually five or seven years out, and no principal is repaid before it. The expectation is that the borrower refinances at that point. If refinancing is unavailable, the investor is left holding a thirty-year claim on a building whose useful contents turn over every two or three years. The rating horizon and the technology horizon are not the same length, and the gap is roughly a decade.

The contract rights deserve more scrutiny than the stress assumptions. Under triple-net leases the tenant, not the landlord, pays the taxes, insurance and maintenance and controls the equipment lifecycle. That stabilises the landlord's cash flow in the near term, which is why lenders like it. But it also means the tenant decides how hard the facility is worked and what condition it comes back in, so the residual value risk moves toward the bondholder. Add termination, reduction and assignment rights, common in single-tenant deals, and "contracted revenue" starts to mean something weaker than it sounds. Those clauses are exercised in exactly the circumstances where the tenant needs them and the bondholder does not.

Where the risk finally sits

Follow the paper to the end and it does not stop at a bank. Much of this lending never went through one.

The chain runs like this. Operators borrow from private credit funds rather than banks, because private funds move faster and take structures a regulated lender would decline. Those funds raise their money largely from insurers, who need long-dated assets to match long-dated policy obligations and have been buying private credit for a decade to get yield that government bonds no longer provide. Insurers, in turn, are funded by the people who hold their policies and annuities. In parallel, a large share of retirement money sits in default investment options that allocate to those same strategies, so the exposure arrives without anyone choosing it individually. And the grid connections these facilities require are paid for through regulated tariffs, which means the upgrade cost is spread across every electricity customer on the network.

So the end holders are insurers, retirement savers allocated by default, and ratepayers. None of the three signed anything, and none of them is being defrauded. This is close to how infrastructure has always been financed. But it does mean the question usually asked, whether the banking system is exposed, is pointed at the wrong place. This is a corporate spending cycle with a slow digestion, not a credit event with a bank at the centre. It will not look like 2008 and it should not be expected to.

The case against this reading

It deserves to be made at full strength, because it has been right all year.

Occupancy is the strongest counter. Data centre vacancy sits near historic lows, capacity is absorbed roughly as fast as it is delivered, and demand has consistently exceeded what the sceptical case assumed. Anyone who positioned for this to unravel during 2026 lost money doing it, and that is not a small piece of evidence.

The useful life argument may also simply be right. If the economic life of this hardware turns out to be longer than the schedules assume, most of the depreciation pressure disperses, and there is real evidence of older accelerators still earning in production.

The balance sheets carrying this are among the strongest in corporate history. Alphabet's negative quarter came alongside raised guidance, which is the behaviour of a company that has decided to spend, not one that has run out of room.

And the securitisation market has been more disciplined than the headline suggests. Rating methodologies published in 2026 apply multiple rounds of tenant default, re-letting delays, haircuts on replacement tenants, and stressed terminal capitalisation rates. Agencies check whether landlords are overselling capacity relative to what a tenant will actually draw.

What we take from it

We read this as a timing problem rather than a solvency problem, and the three mechanical forces are what make it a question of timing rather than of sentiment. Depreciation schedules, lease commencement dates and the tax reversal converge on roughly the same window between 2027 and 2029, and they converge whether or not demand cooperates.

What would change that read is narrow and observable. If a second large operator extends the assumed useful life of its AI hardware and its auditors sign it, the depreciation wave stretches and most of the pressure disperses. One operator doing it is a company decision; two make it a sector convention. That single change would do more to alter the arithmetic than any demand surprise.

The second thing that would change it is evidence. There is still almost no public record of what these assets fetch when actually sold, because very little has been through a full cycle. Efforts to collect executed secondary transactions with provenance have started. If that data becomes standard, residual value stops being a modelling assumption and becomes an observable number, and the collateral question resolves itself in either direction.

Dates that do the work

Late 2026 into 2027 — the next GPU generation ships, which marks the first genuine test of second-hand value for the current one.

2027 to 2029 — lease commencements peak and refinancing concentrates in the same window.

Every quarterly filing — whether the not-yet-commenced lease line keeps growing at the pace it did through the first half, and whether any operator changes its depreciation assumptions.

The first real disposal — a large, disclosed sale of used accelerators with known provenance. It has not happened yet, and it would settle more than any forecast.

A note on sourcing. The figures here come from company filings, from published rating agency research, and from Goldman's capital expenditure work, and each was checked against its source rather than taken from summaries. Where a number appears in circulation that could not be traced to a primary source, it has been left out. Several are missing for that reason, and the argument does not depend on them.