The Crossover
Big Tech AI capex crosses cash flow this quarter. From here, the build is borrowed.
The prospectus landed on the desks of institutional bond buyers on a Monday morning in February. It was several hundred pages, it carried the Alphabet Inc. name, and it meant investment-grade paper backed by the largest advertising business on the planet. Twenty billion dollars in seven tranches, with maturities spanning decades and a century bond that would not come due until 2126.
By Friday, the total had climbed to thirty-two billion dollars across dollar, sterling, and Swiss franc denominations. It was the largest corporate debt raise of 2026 to that point. It had a single purpose: data centres.
Alphabet had just told the market it would spend between $175 billion and $185 billion in capital expenditures this year, and operating cash flow, formidable as it was, could not cover the full cost. Morgan Stanley projected global AI-related bond issuance would approach $570 billion in 2026, nearly double the prior year. The era in which hyperscalers funded their infrastructure from earnings had ended. The era in which they funded it from capital markets had begun.
That transition is the threshold this article is about. When Google, Amazon, Microsoft, and Meta collectively guide to $725 billion in capital expenditure for a single fiscal year, up 77 percent from the roughly $410 billion they spent in 2025, the number stops being a corporate planning metric and becomes a macroeconomic variable. The combined spending of four companies exceeds the GDP of all but roughly twenty countries, larger than the entire economic output of Sweden, Poland, or Belgium. Bridgewater Associates estimates that AI-related capital expenditure will contribute 140 basis points to US GDP growth in 2026, accounting for half of the country’s projected 2.8 percent real growth.
The Federal Reserve cannot model the path of inflation without accounting for what these four companies decide to build.
Jim Covello, head of equity research at Goldman Sachs, has been making a version of this argument since mid-2024. His position has sharpened over two years, not softened. “Companies are losing more money today implementing this technology than they were two years ago,” he told Fortune in June. The hill has gotten steeper because the spending has accelerated while the revenue base has not kept pace.
In 2024, Sequoia Capital’s David Cahn calculated that the AI industry needed to generate $600 billion in annual revenue to justify the infrastructure investment. In 2026, with capex nearly $200 billion above that figure and climbing, the implied revenue requirement is substantially larger. The gap is widening, not closing.
The bull case is not imaginary. Microsoft’s AI business reached a $37 billion annualized run rate in Q3 FY2026, growing 123 percent year over year. Google Cloud posted $20 billion in revenue for Q1 2026, up 63 percent from the prior year, with a backlog that nearly doubled quarter on quarter to more than $460 billion. These are real numbers, and the thirty percent probability that the build validates itself belongs to this world, the one where the revenue curve bends sharply enough to close the gap before the debt burden becomes structural.
If you are long the hyperscaler trade and watching Q2 earnings in late July, a second consecutive quarter of 60-plus percent cloud growth from Alphabet and another upward revision to Microsoft’s AI run rate would be the signals that the capex-to-revenue ratio is stabilising rather than diverging.
But here is what even the optimists must hold in their heads simultaneously. Microsoft attributed $25 billion of its 2026 capex increase to memory chip inflation, not to new capacity. DRAM contract prices rose roughly 95 percent quarter over quarter in Q1 2026, with a further 58 to 63 percent increase projected for Q2, according to IDC. When Meta raised its full-year capex guidance from a range of $115 billion to $135 billion up to $125 billion to $145 billion in late April, the company cited higher component prices as the primary driver.
A meaningful fraction of the 77 percent year-over-year capex increase is not buying more compute. It is buying the same compute at higher prices.
This is the arithmetic Covello keeps returning to. The economics are more questionable today than they were two years ago, he has argued, because enterprise buyers, model companies, and hyperscalers have still not demonstrated returns on what they are spending. All economic value in the AI supply chain continues to flow to semiconductor companies, a dynamic he calls historically unprecedented and unsustainable. The question is not whether AI revenue is growing. It is whether AI revenue can grow fast enough, on a large enough base, to justify a capex programme that is itself inflating due to input cost pressures the capex programme created.
The most probable path, at thirty-five percent, is the one the bond market is already underwriting. Revenue grows at 40 to 60 percent but the gap between what the build costs and what it earns does not close. Hyperscalers borrow more, because the competitive cost of stopping is higher than the financial cost of continuing. Amazon, Alphabet, Meta, Microsoft, and Oracle collectively issued $121 billion in US corporate bonds in 2025, up from an annual average of $28 billion between 2020 and 2024, according to Bank of America.
Morgan Stanley projects $570 billion in AI-related bond issuance for 2026, with $236 billion already issued through May. UBS notes that hyperscaler capex is on pace to consume nearly 100 percent of operating cash flow this year, compared with a ten-year average of 40 percent. If you are a fixed-income allocator, this is the scenario that rewrites your credit book. The weight of four companies’ capital needs is large enough to compress spreads across the entire investment-grade universe.
And then there is the physical ceiling. Nearly half of all US data centres planned for 2026 have been cancelled or delayed, with only about 5 gigawatts under active construction out of 12 gigawatts announced. The bottleneck is not chips. It is transformers.
Lead times for large power transformers have extended to four years, and China controls approximately 60 percent of global production capacity. The binding constraint on the AI infrastructure build is not the willingness of hyperscalers to spend or the willingness of bond markets to lend. It is the lead time on a piece of electrical equipment that weighs several hundred tons and requires more than a year to manufacture.
Twenty-five percent probability belongs to the world where the build slows not because the financial thesis fails but because the supply chain cannot deliver. If you are pricing AI infrastructure equities and watching the next two quarters, the tell is not the earnings call. The tell is the transformer order book. A hyperscaler that guides capex lower in Q3 or Q4 and cites infrastructure delivery timelines rather than demand optimisation is confirming this scenario.
The financial threshold is equally precise. Epoch AI calculates that aggregate hyperscaler operating cash flow is growing at roughly 23 percent per year while cash capex is growing at roughly 70 percent per year. Those two lines cross in Q3 2026, the quarter that has just begun, when aggregate free cash flow reaches zero. Oracle has already crossed. Amazon is crossing now. Alphabet’s crossover arrives around Q1 2027, Meta around Q3 2027, Microsoft not until Q3 2028.
The sequence matters. By the time the last company crosses, the first will have been borrowing for two years.
Ten percent belongs to the tail. A hyperscaler cuts capex by 20 percent or more in a single quarter. The market reads it as confirmation of the Covello thesis, and the AI infrastructure trade unwinds in days. If you are running a tech book, this is the risk you hedge against not because it is likely but because the velocity of the unwind would be faster than any portfolio rebalance you could execute after the fact.
Q2 2026 earnings arrive in late July. Watch for updated full-year capex guidance from all four hyperscalers and, more specifically, for whether Microsoft disaggregates capacity expansion from input cost inflation in its capex disclosure. If the $25 billion memory inflation figure grows, the build is inflating faster than it is expanding.
DRAM contract pricing for Q3 2026 publishes via TrendForce in late August. Server DRAM prices rising at 50-plus percent per quarter would confirm that the input cost spiral is accelerating, rendering every capex guide in the sector stale on the day it is published.
Epoch AI’s free cash flow tracker updates monthly. The aggregate crossover confirmation is expected in the September update. Once free cash flow turns negative for the group as a whole, the financing conversation shifts from how much they can borrow to how long.
The October transformer delivery report from the Department of Energy will reveal whether the physical bottleneck is loosening or tightening. Lead times above 36 months confirm the physical ceiling scenario is binding.
Watch, too, for any hyperscaler that stops disclosing capex by region or by category in its Q2 results. When a company that has been breaking out AI-specific spending stops breaking it out, the disclosure is the disclosure.
For forty years, the macroeconomic variables that central banks, treasury departments, and commodity traders had to model were sovereign in origin: government spending, central bank balance sheets, fiscal deficits, trade flows. The AI infrastructure build has introduced a new category. Four companies, answerable to their boards and shareholders but to no electorate and no central bank, are now deploying capital at a scale that moves interest rates, reprices electrical transformers from South Korea to the American grid, and adds a full percentage point to the GDP growth rate of the world’s largest economy. The crossover is not just the quarter when capex exceeds cash flow. It is the moment when corporate capital expenditure becomes a force that macroeconomists can no longer treat as noise. The four balance sheets are now inside the model.
ANNEX: WHAT HAPPENS WHEN THE BUILD OUTRUNS THE CASH
Four scenarios for the next twelve months of AI infrastructure financing, summing to 100 percent.
Revenue Absorption -- 30%
If you hold hyperscaler equity and you have been waiting for the revenue to justify the capex, this is the scenario where it does. Microsoft’s AI run rate crosses $50 billion by late 2026. Google Cloud sustains 60-plus percent growth through the back half of the year. Meta begins separately disclosing AI-driven advertising revenue, and the number silences the sceptics. The capex-to-revenue ratio stabilises by late 2027, and the hyperscaler trade, which has been pricing on faith for two years, begins pricing on fundamentals. In this world, Alphabet’s thirty-two billion dollar bond raise looks not like desperation but like the smartest financing of the cycle.
The quantitative variable to watch is Microsoft’s AI annualized revenue run rate, disclosed quarterly in earnings. The current figure is $37 billion as of Q3 FY2026 (April 2026). If it crosses $50 billion by the October 2026 earnings report (Q1 FY2027), the absorption is accelerating. Scenarica estimates the probability of crossing $50 billion within one month at 10 percent, within three months at 30 percent, and within twelve months at 65 percent.
The Leveraged Plateau -- 35%
This is the most probable path and the one the bond market is already underwriting. Revenue grows at 40 to 60 percent annually but the gap between what the build costs and what it earns does not close. Hyperscalers borrow more, not because they must but because the competitive cost of pausing is existential: the company that slows cedes capacity to the company that does not. Bond markets absorb the issuance because the credits are investment-grade and the coupons are attractive relative to sovereign paper. The risk is not default. The risk is that four companies become the dominant borrowers in investment-grade credit, compressing spreads for every other issuer and structurally repricing the cost of corporate debt.
The quantitative variable to watch is aggregate hyperscaler bond issuance tracked against Morgan Stanley’s $570 billion forecast for 2026, reported monthly by Bloomberg and Refinitiv. Through May, $236 billion had been issued globally. If the monthly pace exceeds $50 billion through Q3 2026, the plateau is forming. Scenarica estimates the probability of exceeding $50 billion per month: one month 60 percent, three months 70 percent, twelve months 75 percent.
The Physical Ceiling -- 25%
If you are watching the AI build and focusing on the financial metrics, you may be watching the wrong constraint. Transformer shortages, DRAM price spirals, and grid power limitations force capex guidance lower before the equity market or the bond market tests the financial thesis. Hyperscalers do not cut because they lose conviction. They cut because the infrastructure cannot be delivered on the timeline the capex guide assumed. The market reads the cut as prudence. The structural read is that the build has hit the speed limit of the physical world, and that speed limit is set not in Silicon Valley but in transformer factories in Ulsan and Changsha.
The quantitative variable to watch is the share of announced 2026 US data centre projects that break ground versus those deferred, tracked by DataCenter Dynamics and the Department of Energy. Currently, roughly 5 GW of 12 GW announced is under active construction. If fewer than 50 percent of announced projects break ground by Q4 2026, the ceiling is binding. Scenarica estimates the probability of sub-50 percent groundbreaking at one month 40 percent, three months 50 percent, twelve months 60 percent.
The Repricing -- 10%
This is the tail risk and it belongs to a single trigger. One hyperscaler, most likely the one with the weakest AI revenue narrative relative to its capex commitment, cuts guidance by 20 percent or more in a single quarter. The market reads the cut not as company-specific but as the first crack in the consensus, and the AI infrastructure trade, which has added roughly $3 trillion in market capitalisation since early 2023, begins to unwind in days rather than weeks. If you are running a tech-heavy portfolio, this is the scenario you cannot afford to ignore even at low probability. The drawdown speed would exceed any rebalancing window.
The quantitative variable to watch is individual company capex guidance revisions in quarterly earnings. The specific trigger is a single company guiding 20-plus percent below the prior range. Scenarica estimates the probability of this trigger: one month 2 percent, three months 5 percent, twelve months 10 percent.
Sources:
Alphabet Inc., Bond Prospectus and Q1 2026 Earnings, February-April 2026.
Microsoft Corporation, Q3 FY2026 Earnings Report, April 29, 2026.
Meta Platforms, Q1 2026 Earnings Report and Updated Capex Guidance, April 29, 2026.
Tom’s Hardware, “Big Tech’s AI Spending Plans Reach $725 Billion,” 2026.
Fortune, “Goldman’s Top AI Skeptic Warns the Clock Is Running Out,” June 5, 2026.
Morgan Stanley, AI-Related Debt Issuance Forecast, 2026.
Bridgewater Associates, “The Macro Implications of the AI Capex Boom,” 2026.
Epoch AI, “Hyperscaler Capex to Exceed Cash Flow by Q3 2026,” 2026.
UBS, Hyperscaler Capex and Cash Flow Analysis, 2026.
Bank of America, Hyperscaler Bond Issuance Data, 2025-2026.
IDC, DRAM Pricing Forecast, Q1-Q2 2026.
TrendForce, DRAM Contract Pricing Reports, 2026.
DataCenter Dynamics, US Data Center Construction and Google Debt Raise Coverage, 2026.
The Register, “Microsoft Lifts 2026 CapEx by $25B to Cover Price Rises,” April 30, 2026.
Sequoia Capital, David Cahn, “AI’s $600B Question,” June 2024.
Disclaimer: This report is published by Scenarica Intelligence for informational purposes only. It does not constitute investment advice, a solicitation to buy or sell any financial instrument, or a recommendation regarding any particular investment strategy. Scenarica Intelligence is not a registered investment adviser or broker-dealer. All scenario probabilities and assessments represent the analytical judgment of Scenarica Intelligence and are subject to change without notice. Past performance of any asset or strategy discussed does not guarantee future results. Readers should conduct their own due diligence and consult with qualified financial advisers before making investment decisions.
Scenarica Premium: The full Scenarica suite includes Geopolitics, Economy, Bitcoin, AI, and Sunday Edition.
Scenarica Intelligence
We don’t predict the future. We price it.