Sunday Essay — Borrowing Long Against Chips That Live Short

Free cash flow at the four big US hyperscalers — Amazon, Alphabet, Meta, Microsoft — is on track to land at its weakest level since 2014, and on aggregate the cohort has just gone negative for the first run of consecutive quarters in roughly thirty-five years, according to a recent analysis of consolidated capex against operating cash flow. The line bent in 2024, broke decisively in 2025, and is still moving the wrong way in 2026. The combined 2026 capital-expenditure guidance from those four firms now sits around $700–725 billion, per company guidance compiled at the end of the most recent earnings season. One analyst on a CNBC segment last month dismissed the bear case as "garbage." I am not so generous.

This essay is about the financial geometry of that decision. Not whether AI works, not whether the models keep getting better, not whether enterprises buy it. Those questions are settled enough. The question that interests me — the one I keep getting asked by audit committees and risk officers and the occasional family-office principal — is whether the cash-flow arithmetic adds up if rates stay where they are, if GPU lives turn out shorter than the disclosed schedules, and if AI revenue compounds at the pace the capex curve is implying.

The honest answer is: it might. It also might not. And the spread between those two outcomes is wider than the people writing the checks seem to acknowledge.

the capex line bent

The simplest framing: in 2023 the four hyperscalers spent roughly $145 billion on capex collectively. In 2024 they spent around $230 billion. In 2025 they spent roughly $410 billion. In 2026 the guidance points to $700–725 billion — Amazon at about $200B, Alphabet at $175–185B, Meta at $115–135B, Microsoft at $110–120B — and the sell-side is already modelling 2027 above one trillion dollars, as CNBC reported in late April. Roughly three-quarters of the 2026 envelope is going directly into AI infrastructure — GPUs, accelerators, the buildings to house them, the power infrastructure to feed them, the optical and copper interconnect to lash them together.

That is a five-fold step-up in three years. The dot-com peak by comparison, properly inflation-adjusted, is small. Telecom build-out at the end of the 1990s, again properly adjusted, is also small. There is no comparable analogue in modern listed-company history for what is happening to the capex line at the top of the S&P. Allianz Trade has been calling this an AI capex super-cycle. The Dallas Fed has flagged in its research note that the share of long-dated investment-grade issuance driven by AI-related firms has become large enough to influence the supply curve of duration in the broader corporate-credit market. The thing is no longer just a tech story; it is a credit story, a rates story, and an accounting story all moving at once.

the maturity mismatch

Here is the structural problem nobody at the analyst-day wants to spell out. The capex being deployed is being financed against asset lives the issuers themselves are uncertain about.

Take CoreWeave, the cleanest example because the structure is naked. In March, CoreWeave closed an $8.5 billion GPU-backed financing facility — the first investment-grade GPU-backed paper — maturing March 2032. So: six years out, secured by chips most of which will be obsolete in three to four. The economics work only if the cash-flow contracts backing the deal carry the loan through and into amortisation before the underlying collateral is worth scrap. The agencies rated it A3 / A(low). Read that twice. Investment-grade paper, against silicon that depreciates aggressively, with single-counterparty revenue risk concentrated at the top.

The hyperscalers do not look like that on the balance sheet, because their revenue is diversified and their cash generation from non-AI businesses is still vast. But the underlying geometry is the same. They are funding multi-year GPU clusters with bonds that mature in five, seven, ten years; and the depreciation schedule they apply to those clusters has been quietly extended over the past two years. Microsoft moved server and network useful lives from four years to six. Alphabet sits at six. Amazon went to six in 2024, then pulled back to five for some categories in 2025. Meta went to 5.5.

Industry critics — and I count myself among them on this specific point — argue that the working life of an actual training cluster is closer to two to three years before the next generation makes the prior one uneconomic to run at the margin for frontier work. NVIDIA's own product cadence supports the shorter number: Blackwell 300 is ramping, Rubin is in the pipeline, and the Q1 fiscal 2027 data-centre revenue print of $75.2 billion, up 92% year-on-year, is itself evidence of how fast last year's chips are being displaced at the frontier. If hyperscalers were forced to shorten GPU lives toward the actual replacement cycle, the cumulative earnings hit across 2026–2028 could exceed $176 billion. That is not a rounding error.

I will stake out a position. Within eighteen months at least one of the four hyperscalers will shorten its useful-life assumption by half a year on some asset class, and the quarter it lands in will not be one anybody massages gracefully. I would bet on Meta first, because Meta has the least defensible AI revenue line and the most exposure to investor patience. I might be wrong. I have been wrong before. But the pressure is structural, and structural pressure tends to find an outlet.

the accounting choreography

What makes this fragile is not the spend itself. It is the dance between three numbers — the depreciation rate, the cost-of-capital assumption, and the AI-revenue ramp — each of which is plausible in isolation and increasingly implausible when stacked together.

If you assume six-year GPU lives, AI revenue compounding at the current pace, and a long-term cost of capital around 6%, the implied NPV on the build-out is positive. Drop GPU lives to four years and the NPV halves. Lift the discount rate by 150 basis points and a meaningful chunk of the marginal investment goes underwater. Slow the AI revenue ramp by even 25% over the next three years and the picture turns negative on a fully loaded basis for everyone except possibly Microsoft, whose enterprise position is the most defensible.

Each one of those three sensitivities is individually a maybe. The probability that all three stay favourable simultaneously for three years is markedly lower than consensus assumes. Morgan Stanley's research team has, to its credit, flagged this directly: the key 2026 risk to its constructive AI thesis is that the capital boom fails to translate into productivity gains in the real economy at the pace the capex curve requires. I would put that risk higher than they do.

Vendor commentary is, predictably, sunnier. Jensen Huang is on record talking about $3–4 trillion of annual AI infrastructure capex by the end of the decade. NVIDIA's data-centre business is at a $300 billion annualised run rate. Take both numbers seriously, but apply the discount you would to any single-supplier projection of its own market. NVIDIA has every reason to publish the high case. The job of a board is to make sure somebody around the table is publishing the low case.

where the rate sensitivity actually bites

The Federal Reserve meets on June 17, with futures pricing roughly a 96–98% probability of no change, May CPI still uncomfortably warm, and a policy path the market currently sees near 3.8% by end of 2026 and 4% by mid-2027. That is the higher-for-longer regime made concrete.

The textbook framing — long-duration assets get punished hardest when discount rates rise — applies cleanly to this cohort, but the mechanism most people miss is on the financing side, not the equity-valuation side. A hyperscaler issuing ten-year paper to fund AI infrastructure today is paying roughly 130–180 basis points more for that duration than it would have in 2021. The annual interest burden on, say, $400 billion of incremental AI-related debt across the four — a reasonable order-of-magnitude estimate over the next twenty-four months — is approaching $25 billion a year of additional cash interest expense by 2027 versus the 2021 baseline. That number does not destroy any of these businesses. But it shaves perhaps two to three percentage points off operating-margin trajectories the equity story is already pricing as expanding.

There is also a talent-compensation overhang the public conversation underweights. Median engineer total comp at OpenAI is reportedly running at roughly $555K, with 90th-percentile packages at the frontier labs clearing $1.28 million, according to Pin's 2026 AI compensation benchmark. That number is structural, not transient. A frontier lab with two thousand technical staff is now carrying a compensation cost that comfortably exceeds a billion dollars a year in cash terms before any equity expense. None of this is in the headline capex line. All of it is in the operating-cost trajectory.

This is where, last week, a CFO at a European industrials group I have worked with for years stopped me mid-sentence in a budgeting review and asked: "Are we really sure the cost of training a frontier model halves every nine months from here?" I had said something casual along those lines. I corrected myself. The honest answer is that the cost of training a given level of capability halves on roughly that cadence; the cost of training the frontier keeps going up because the frontier keeps moving. Two different curves, often blurred. He was right to catch it. CFOs are usually right when they catch you blurring two curves.

the board memo

If I were on the audit committee of one of the four — and I am not, though I sit close to a couple of analogous boards in Europe — the memo I would push for would be three pages long and would say roughly this.

One: the depreciation assumption on AI infrastructure is the single most material accounting estimate on the balance sheet, and external auditors should be asked to document why management's assumption is reasonable in light of the observed replacement cycle and competitor disclosures. Make them write it down. Make them sign it. The discipline of having to defend the assumption in writing every quarter changes how loosely it gets adjusted.

Two: the AI revenue line should be broken out with at least three sub-disclosures — recurring enterprise contracts, consumption-based inference revenue, and one-off training engagements — because aggregating them obscures the part of the revenue most sensitive to a recession. AI revenue in aggregate looks like a beautiful exponential. AI revenue split into its real components looks like a respectable enterprise SaaS business with a large consumption-rate-sensitive tail. The board needs to see the tail.

Three: rate-sensitivity disclosures should be standardised. What does a 100 basis-point shift do to project-level returns on the marginal cluster? What does a 25% slowdown in AI revenue do to the cash conversion cycle? These are well-known sensitivities to management. They are not well-known to most non-executive directors. Closing that gap is the cheapest governance improvement available.

A memo of that shape, in various forms, is the kind audit committees push back on politely — and then quietly ask for in supporting-schedule form a week later. They catch up; they just catch up after the analyst-day cycle, which is the cycle that drives the language management uses publicly.

what this rhymes with, what it does not

The dot-com rhyme is real but only partial. The capex of 1999–2000 was misallocated to fibre and routing capacity that was profoundly overbuilt; what saved the survivors was that demand eventually caught up over a decade, and the capital was largely written off in the meantime by the bankrupt builders. The AI rhyme is different in two important ways. First, the demand is already here, not speculative — enterprise inference and consumer chat-style traffic are real, large, growing. Second, the capex is being deployed by firms with the strongest balance sheets in the listed universe, not by venture-funded entrants of unknown durability.

Where the rhyme is strong: the speed at which the capex curve is bending is comparable to the worst parts of 1999, and the temptation to treat capex as the index of strategic seriousness — rather than NPV — is back in full force. The phrase "we cannot afford not to spend" has been spoken to me, verbatim, by two CEOs in the past six weeks. That phrase is always the moment to start asking the disciplined questions.

a stake to drive into the ground

If you forced me to write down a single bet — and I will, because I owe the reader a position and not just a survey — it would be this. Within the next twelve to eighteen months the equity market will reprice the AI-infrastructure cohort sharply on a single quarter where two things happen at once. A hyperscaler will revise GPU useful lives downward and disclose an unexpected non-cash charge; and a major neocloud will need to refinance into a less generous credit market. The combination — not either alone — will move the names ten to twenty percent inside a fortnight. It will look, at the time, like a confidence shock. It will actually be a duration-mismatch shock that the credit and accounting plumbing finally surfaced. I am not predicting a crash. I am predicting a repricing. The build-out continues afterwards. It just continues from a more honest level.

The macro question for the rest of us is not whether AI is real. It is whether we are pricing the infrastructure underneath it on assumptions that survive a single bad quarter. I do not yet think we are. I would like to be wrong about that. The honest measurement problem here is the same one that shows up in productivity statistics, in benchmark results, in vendor case studies: the numbers we publish are the numbers we want to be true. The numbers we eventually live with are the ones the cash-flow statement forces on us.

Tarry Singh is the founder and CEO of Real AI, an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan for Energy AI, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.