The Hyperscaler Capex Stress Test

On 17 September 2025, Moody's revised its outlook on Oracle's Baa2 senior unsecured rating from stable to negative. The text of the action made the reasoning unusually direct. Oracle's debt, the agency wrote, "is expected to increase faster than its EBITDA, adding to a forecast high leverage of 4x," and "it is likely that free cash flow will also be negative for an extended period before reaching breakeven."¹ Two weeks later, S&P affirmed the company's BBB rating but moved its own outlook to negative.²

Rating actions like this happen to leveraged industrials. They are not supposed to happen to large-cap technology companies, the class of issuers that for two decades has been the textbook example of cash-generative, capex-light, financially fortified equity.

This piece steps back to the on-balance-sheet picture. The four largest US hyperscalers, namely Microsoft, Meta, Alphabet and Amazon, have collectively committed enough capital expenditure for 2026 to make the historical self-funded-from-operating-cash-flow narrative no longer hold. They are now structural borrowers, and credit markets have begun to price them as such.

The Scale of the Commitment

In 2020, Big-4 combined capex was roughly $107 billion.³ Five years later, the same four firms spent close to $390 billion. Guidance for 2026 puts the combined Big-4 figure between $620 billion and $700 billion, depending on whether one counts only the disclosed company ranges or the marginally higher CreditSights estimate. Adding Oracle takes the figure to approximately $750 billion on the CreditSights basis.⁴

A useful way to feel that number is to anchor it against macro denominators. The Federal Reserve Bank of Dallas, in a February 2026 research note, calculated that aggregate hyperscaler capex was on track to represent approximately 10% of total US private non-residential fixed investment in 2026, against roughly 3% in 2020.⁵ MUFG puts the same spending at around 2.2% of US GDP.⁶ S&P Global Ratings, writing after Q4 2025 earnings, observed that the combined capex of the Big-4 now exceeds the aggregate capex of the entire S&P 500 industrials sector.⁷

A separate concentration metric: Synergy Research, the most-cited industry tracker, reports that Amazon, Microsoft and Google together account for 58% of all hyperscale data centre capacity globally.⁸ Adding Meta lifts the Big-4 share of hyperscale to roughly 75 to 85% on a capex basis.⁹ Inside the broader US tech sector, S&P notes that the Big-4 share of capex has moved from approximately 50% in 2020 to over 80% in 2025.

Most of that incremental spend has a single source. CreditSights estimates that approximately $450 billion of 2026 hyperscaler capex, roughly 60% of the $750 billion Big-5 total including Oracle, will fund AI-related infrastructure specifically.¹⁰ Microsoft is the only Big-4 firm to publish an explicit dollar split: Brad Smith's January 2025 blog post disclosed that approximately $80 billion of fiscal 2025's $88 billion in capex was committed to "AI-enabled datacenters."¹¹ Read literally, that is 91% of company-wide capex; the marginal dollar, as Satya Nadella put it on the Q1 FY26 call, goes to "a fleet that is fungible across the planet … for inference, for pre-training, for post-training, for RL."¹²

When the Cash Flow Stopped Covering

For most of the last decade, hyperscaler capex was a story of operating cash flow recycled into infrastructure. The defining characteristic of the model was that cash generation grew at least as fast as cash deployment, so free cash flow remained positive and frequently grew with it. In 2025, that relationship broke at three of the four firms.

Amazon was the most dramatic case. Trailing-twelve-month free cash flow collapsed from $38.2 billion to $11.2 billion, a 71% decline, "driven primarily by a year-over-year increase of $50.7 billion in purchases of property and equipment," in the company's own words.¹³ Capex consumed essentially 99% of operating cash flow. Andy Jassy's 2026 guidance of approximately $200 billion in capex, against consensus operating cash flow of roughly $140 billion, means Amazon's 2026 free cash flow is projected by Morgan Stanley and Bank of America to land between negative $17 billion and negative $28 billion.¹⁴

Meta's free cash flow declined 19.4% in 2025 to $43.6 billion, the first annual FCF drop since 2022.¹⁵ The company's Q4 2025 release guided 2026 capex at $115 to $135 billion; in April 2026 that range was raised to $125 to $145 billion, citing "higher component pricing this year and, to a lesser extent, additional data centre costs."¹⁶ Against consensus 2026 revenue of roughly $250 billion, that puts Meta's capex-to-revenue ratio between 50% and 58%. For context, Meta's 2018 ratio, at the time considered unusually high, was around 25%.

Alphabet generated $164.7 billion in operating cash flow in 2025 and spent $91.4 billion on capex.¹⁷ Its 2026 capex guidance, raised in April from $175 to $185 billion to $180 to $190 billion, exceeds 2025 OCF outright before any deterioration. Pivotal Research's published 2026 free cash flow estimate for Alphabet, relayed via Asymco, is roughly $8.2 billion against $73.3 billion in 2025, an 89% decline; it is a single-shop projection and should be read as such.¹⁸ Asymco's Horace Dediu has noted that, if current capex guidance is funded as planned, the combined 2026 free cash flow of the Big-4 "could approach zero for the first time in at least 15 years, at exactly the moment they are also raising record amounts of long-dated debt."¹⁹

Microsoft is the only firm where the regime change is still partial. Free cash flow declined 3.3% in fiscal 2025, the first annual drop since FY2017, but operating cash flow growth narrowly kept pace with the 45% capex jump, and the firm retains, by some distance, the strongest interest coverage in the group.²⁰ Even so, capex now eats nearly 50% of revenue on the most recent quarter and leaves no buffer against either OCF deterioration or another capex acceleration.

The Bond Market Took the Strain

What happens when four companies that historically did not issue much debt suddenly need to fund half a trillion dollars of annual capex from somewhere other than free cash flow? The answer, in 2025, was the investment-grade bond market.

Mellon Investments' January 2026 review of full-year 2025 issuance studied the five largest hyperscaler issuers, namely the Big-4 plus Oracle. Oracle is included here, even though its business mix sits one tier below the four cloud platforms, because the 2025 issuance wave is driven by the same AI-infrastructure financing dynamic, and excluding it would understate the volume the market actually absorbed. On that basis, the five firms issued $121 billion in US corporate bonds in 2025, against an average $28 billion per year between 2020 and 2024.²¹ Bank of America initially forecast $140 billion of further issuance for 2026; in March, after Amazon's $54 billion multi-currency bond offering linked to its OpenAI infrastructure commitment, the desk raised the forecast 25% to $175 billion.²² By mid-March 2026, $110 billion, or 63% of the revised full-year forecast, had already priced.

The marquee deal was Meta's $30 billion six-tranche offering in late October 2025, the largest non-M&A high-grade corporate bond sale on record.²³ The order book reportedly peaked at $125 billion. The tenor structure is informative: tranches at 4.200% maturing in 2030 and 4.600% in 2032 at the short end, scaling out to 5.500% in 2045, 5.625% in 2055, and 5.750% in 2065. Use of proceeds was "general corporate purposes", meaning, fungible into capex, since Meta did not need the cash for working capital. The barbell tenor, anchoring the short end while locking the long end for forty years, reads as a treasury team's response to a refinancing wall it can already see.

The Dallas Fed's framing of this volume of issuance is worth keeping in mind: bond supply from these five firms now accounts for roughly 5% of US investment-grade corporate bond issuance, against 1 to 2% historically.⁵ That share has not yet been priced into the duration risk premium, which is why much of the recent supply has cleared at narrow spreads. The question is whether that holds when the second wave of long-dated issuance arrives.

Where the Pressure Falls: 2027 to 2029

The headline interest expense numbers at Microsoft, Meta and Alphabet do not look stressed. Microsoft's fiscal 2025 EBIT-to-interest coverage was approximately 52x. Meta's 2025 reported interest expense of $1.17 billion against pre-tax income above $80 billion remains a wide margin.²⁴ Alphabet's reported interest is small enough to be a rounding error on its income statement.

The pressure is not in the income statement. It is in the maturity profile. Much of the legacy hyperscaler debt was issued at sub-3% coupons between 2017 and 2020. As that paper rolls between 2027 and 2029, it will refinance into a 5 to 6% environment. The same is true of the new long-dated 2025 to 2026 supply, which is locked in at higher coupons but will itself come due in a yield environment whose direction is not knowable. Moody's view on Meta, expressed in its February 2026 affirmation, illustrates how the agencies are framing this. The company's Aa3 rating was kept at stable outlook, but only, the analysts wrote, because Meta's "exceptional liquidity with more than $81 billion of cash and marketable securities" mitigates "limited to no free cash flow generation over the next two years."²⁵

That formulation is significant. The rating is liquidity-supported, not earnings-supported. If the $81 billion cash buffer is drawn down through 2027 to 2028, by capex and by the $30 billion bond's coupon payments combined, the support structure thins, and the agencies will be looking again. The Oracle situation gives a glimpse of what that looks like: interest coverage already at approximately 5x and falling, JP Morgan equity research flagging Oracle's debt-to-equity ratio as "much higher than AI peers," net debt to EBITDA around 4x against peers below 1x.²⁶

The Microsoft stress arithmetic, by contrast, looks comfortable. If the firm refinances current debt at 6% and adds $50 billion of new 6% paper, incremental interest expense is roughly $3 billion per year. Even doubling reported interest expense from $2.4 billion to roughly $4.8 billion leaves EBIT-to-interest coverage above 25x, still investment grade, still well clear of any agency trigger. The implication: the credit stress is not uniform. Microsoft and Alphabet bear it most lightly; Meta is liquidity-cushioned but exposed; Amazon's earnings buffer is large but its free cash flow position is the weakest; Oracle is the one already at the line.

The Counter-Case

The strongest objection to the stress thesis is that AI revenue will compound faster than capex. If it does, the capex-to-revenue ratios that CreditSights describes as "seemingly untenable", namely 47% at Microsoft, 54% at Meta, 46% at Alphabet, 25% at Amazon, converge back toward historical norms.²⁷ Microsoft has the strongest demonstrated revenue base in this regard. Amazon's $15 billion AWS AI revenue run-rate and 1.4 million Trainium2 chips deployed argue that AI-tied revenue is materializing in real numbers. Both observations are correct and they are the right counterweight to a one-sided framing.

A second objection is structural. The most-cited historical analogue, the 1996 to 2001 telecom boom, was debt-financed from day one. Issuers like WorldCom, Global Crossing and Qwest were not generating the operating cash flow to support the buildouts they were undertaking. The Big-4 hyperscalers are in a different position. They remain solidly investment grade, with cash buffers measured in the tens of billions and core businesses generating hundreds of billions in operating cash. The Couper-Hejkal-Wolman paper from the Richmond Fed, which is the cleanest primary source on the telecom buildout's magnitude, documented overcapacity driven in part by "the dramatic increase in the capacity of a given strand of fibre," namely a supply-side overshoot.²⁸ Whether AI infrastructure has the same property depends on questions about algorithmic efficiency, model architecture and inference economics, namely questions taken up in detail in earlier pieces in this series.

A third objection is about duration. Paul Kedrosky's August 2025 essay, comparing AI capex to historical infrastructure booms as a share of GDP, places AI datacenter capex at roughly 1.2% of US GDP, already above the dot-com peak telecom equipment spend (around 1%) but well below the railroad peak of approximately 6% in the 1880s. Kedrosky's most useful observation is the asymmetry: rails were century-class assets, and even dark fibre had a 20-year residual life that ultimately found use. "Datacentres are short-lived," he writes. "We aren't building century-long infrastructure."²⁹ That is a real cautionary, and it cuts both ways: the buildout is smaller as a share of GDP than the historical precedents, but the asset life is also dramatically shorter.

A fourth objection is about composition, and it is the most uncomfortable for the stress thesis. The headline framing treats hyperscaler AI capex as a homogeneous asset class depreciating on a GPU schedule. The actual asset stack is more heterogeneous. Microsoft's FY25 10-K discloses useful-life ranges of three to six years for server and network equipment, against fifteen to thirty years for buildings and land improvements.³⁰ Meta's 10-K reports a similar split. If, conservatively, GPUs and server-class equipment account for 50 to 60% of the AI-tied capex dollar and shells, power substations, fibre and cooling account for the remainder, the asset-life-weighted average of the AI capex stack sits closer to eight to ten years than to four to six. The maturity-mismatch argument softens accordingly. It does not disappear, because the bond tenor on the long end, namely the 40 to 50 year tranches of the marquee deals, still exceeds even the long-life portion of the asset stack, and the GPU portion still depreciates faster than the debt that funds it.

What to Watch

The honest framing of the current position is that this is not 1999 telecom, but it is not the cloud build of 2015 either. Capex-to-revenue at the Big-4 in 2016 to 2020 averaged 10 to 15%. The 2026 guidance puts it at 25 to 50% plus. The financing mix has shifted from cash-only to roughly half-debt at the marginal dollar. The asset stack, weighted by composition, depreciates closer to eight to ten years on average than to fifteen to twenty, and the GPU-heavy slice depreciates faster still. None of those facts individually breaks the business, and none of them require imagining a 1999-style equity collapse. They do mean that the credit profile is qualitatively different than it was even three years ago.

Three things, with concrete thresholds, are worth watching across the next four quarters.

The first is the trajectory of capex guidance. Watch for any Big-4 firm raising 2026 or 2027 capex guidance by more than 10% of its spring 2026 midpoint at the Q3 2026 or Q4 2026 print. 2026 guidance has already been raised twice at Meta and once at Alphabet within a single year, which is not how mature companies typically forecast. A further mid-cycle raise of that magnitude would be the cleanest signal that the buildout is still re-pricing upward in real time.

The second is the cash buffer trajectory at Meta and Amazon. The Moody's logic on Meta is explicit about liquidity, and the buffer is what is keeping the outlook stable. Meta's combined cash-and-marketable-securities position falling below $50 billion would put the liquidity-supported rating logic at material risk. For Amazon, a cash position falling below $80 billion against the $200 billion 2026 capex guide would imply the firm is consuming its buffer faster than the bond market can refinance it.

The third is the depreciation curve. Useful-life assumptions have been lengthened at every Big-4 firm in the last two years, which has flattered reported margins. Watch for any Big-4 firm shortening useful-life assumptions in the 10-K depreciation footnote for fiscal 2027 or fiscal 2028, particularly on server and networking equipment. A reversion to shorter useful lives would compress earnings in 2027 to 2028 in a way that the current consensus does not assume, and would re-couple reported earnings to the GPU-class depreciation curve that the lengthening has masked. The depreciation question was examined at length in the April piece in this series.

Personal View

The body above is the maturity arithmetic. What follows is what the arithmetic leaves me uneasy about.

The body above frames the credit stress as a grind, not a bubble, and on the maturity-mismatch evidence I think that framing is the right one. What I am less comfortable with is the implicit assumption sitting underneath both the grind thesis and the counter-thesis, namely that the economic model the buildout is betting on stays roughly intact. The Big-4 are pricing roughly $700 billion of 2026 capex against a per-token revenue curve that has been declining steadily but predictably, on assets that depreciate over four to fifteen years depending on composition, financed by debt that comes due over five to forty. The structure works as long as the decline in per-token revenue stays predictable.

In my view, that is the weakest assumption in the stack. The historical precedent that worries me is not 1999 telecom or 1880s railroads. It is the way algorithmic and architectural improvements have repeatedly compressed LLM inference costs by factors of three to ten in short order. The DeepSeek R1 episode on 27 January 2025 took roughly $590 billion off Nvidia's market capitalisation in a single trading day before the market re-priced and partially recovered. That was not an outlier event. It was a sample from an ongoing pattern that includes the GPT-4 to GPT-4o transition, the Claude 3 to Claude 3.5 generation, and the Llama-class open-weight compounds, each of which delivered order-of-magnitude per-token cost improvements on 12-to-18-month cycles. If something of similar magnitude happens again without warning between 2026 and 2028, the asymmetry of the position is unforgiving. The capex is sunk, the debt is fixed, the revenue assumption collapses faster than the depreciation schedule can absorb. The willingness of US capital markets to issue 40-year paper against an asset stack with a four-to-ten-year weighted life is itself the expression of a high tolerance for that kind of tail risk.

I do not know where the algorithmic-efficiency curve goes from here, and I do not believe anyone with credibility in 2026 does either. What I do know is that the asymmetry is the wrong way around for the people writing the cheques. That is what keeps me cautious, even inside the grind scenario.

This is part 4, the closing piece of a four-part series on AI infrastructure financing. The earlier pieces map the layers feeding into the headline credit profile read here: Part 1 on the SPV structures financing the GPU clusters; Part 2 on the vendor-financing loop that books their demand; Part 3 on the off-balance-sheet compute commitments that read as backlog on the seller's books and as nothing on a private buyer's.

END · 04