Big Tech's four largest hyperscalers - Microsoft, Alphabet, Amazon, and Meta - are set to spend somewhere between USD 720 billion and USD 760 billion on capital expenditure in 2026, up roughly 77% from about USD 410 billion in 2025.
This is, by any measure, the largest single-sector capital deployment in the history of American business - and it was originally supposed to be a story about corporate cash flow, not sovereign credit. That is no longer true. As of September 24, 2026, the 10-year Treasury yield is holding around 5.12%, near its highest level since July 2007. Two weeks earlier, on September 14-15, the 10-year yield briefly surpassed 5% for the first time since 2023 - a level analysts had flagged for months as the threshold at which the AI trade and the fiscal trade would collide.
Advertisement - Scroll to continue
It has not meaningfully receded since.
It is important to note that if long-term interest rates decisively breach the 5% threshold, the impact could derail the AI boom. Global investors have been giving Washington a "free pass" on those numbers till now because of an implicit bet that AI-driven productivity growth will eventually outrun the debt.
The 10-year yield crossing 5% is the market beginning to test that bet.
From Free Cash Flow to Debt Markets
For most of the AI buildout's first two years, hyperscalers funded it the easy way: out of operating cash flow. That era is over. Consensus estimates show AI capex hitting 94% of operating cash flow minus dividends and buybacks in 2025-2026, up from 76% in 2024 - and the shortfall is being plugged with debt. The cost of that debt is rising in lockstep with the Treasury curve it's priced off. Of 91 hyperscaler bonds issued in 2026 with comparable pricing data, 78 were trading at higher yields by late July than when they were issued - a sign that the market is repricing hyperscaler credit risk in real time, not just riding the Treasury curve passively.
Advertisement - Scroll to continue
Oracle is the cautionary case. S&P downgraded Oracle to BBB in July 2026, one notch above junk, citing elevated debt levels, and its fiscal 2026 free cash flow came in at negative USD 23.7 billion - the gap between what its cloud business earns and what it is spending on data centres, chips, and leases. Oracle's credit default swaps hit a record high on August 4, 2026. There's also a shadow layer that doesn't show up in on-balance-sheet debt at all. The five largest hyperscalers held USD 969 billion in total undiscounted future data-centre lease commitments at year-end 2025, of which USD 662 billion sit entirely off-balance sheet, equal to 113% of the same companies' combined on-balance-sheet debt. Whatever the headline debt figures say, the real leverage embedded in the AI buildout is considerably larger.
The 5% Yield Threshold and Earnings Yield Compression
The equity risk premium is the whole ballgame for a high-multiple growth sector. As the risk-free rate (the 10-year yield) rises toward and past 5%, the ERP compresses unless earnings yields rise in step - which, for AI infrastructure names trading on hyperscaler-adjacent multiples, they generally have not.
A stock trading at 30x forward earnings has an earnings yield of roughly 3.3%; against a 5%+ risk-free rate, that arithmetic implies a negative ERP unless investors are pricing in extraordinary future growth to compensate. That is precisely the bet embedded in AI infrastructure valuations today.
The market has already demonstrated how sensitive it is to this threshold in practice. When yields retreated, so did the pressure - on September 21, the Nasdaq hit a record close as yields fell back below 5%, with AMD's market cap crossing USD 1 trillion for the first time. But that relief was short-lived: by September 24 the 10-year was back above 5.1%, and as one strategist observed of the mid-September episode, the AI trade was "already broken before yields hit 5%" - the yield spike just widened the selling.
The Revenue Gap vs. the Capex Sinkhole Sequoia partner David Cahn's "USD 600 billion question," first posed in 2023 at USD 200 billion and updated through 2024 and 2026, calculates the annual AI-industry revenue required to justify the capex being deployed - and that gap is widening, not closing, in 2026. On the revenue side, the growth is real but dwarfed by the spending it's meant to justify. AWS is running at roughly USD 150 billion annualised, up 28% year-over-year; Google Cloud is around USD 80-90 billion, up 48-63%; Azure AI is at a USD 37 billion run rate, up 123%. Impressive growth rates on a revenue base that is a fraction of the roughly USD 750 billion in annual capex it's supposed to underwrite.
Allianz Research put a number on the mismatch: the divergence between AI capital expenditure and revenue growth is running at roughly 46% in 2026 - already exceeding the 32% divergence observed during the 2001 telecom capex bust, a cycle that ended in one of the more brutal multi-year tech corrections on record.
Not every hyperscaler is equally exposed. Amazon's free cash flow plummeted 71% to 11.2 billion on the back of its roughly USD 200 billion capex plan, while Google Cloud revenue rose 48% to $17.7 billion backed by a USD 240 billion order backlog - a much stronger revenue-to-spend ratio. Microsoft sits in between. Azure's growth of 40% remains robust, but the company is capacity-constrained and unable to meet total demand through fiscal year-end, which cuts both ways - it suggests real, monetisable demand, but also that the capex isn't yet translating into realised revenue.
The evidence assembled here doesn't support a clean verdict either way, but it does sharpen the terms of the bet. The 5% line has already been crossed, repeatedly, and the market's reaction each time - sharp semiconductor selloffs, followed by relief rallies when yields dip even briefly - shows that this threshold isn't a future risk but a present, live tripwire that traders are now watching daily.
The financing mix is shifting in the wrong direction at the wrong time. Hyperscalers are moving from cash-funded to debt-funded capex just as sovereign issuance is pushing term premia higher. The revenue side has not caught up, and by some measures, the gap is widening faster than capex itself.
None of this means the AI boom "pops" on a specific date. Sovereign debt dynamics move on a scale of years, not weeks, and productivity gains from AI adoption - even partial ones - could yet outrun the fiscal arithmetic, which is the scenario the bond market has arguably been pricing in as a "free pass" up to now. Yet, the mechanism we describe here is no longer hypothetical: it is visible in real time in hyperscale bond spreads, in Oracle's credit rating, and in semiconductor stocks trading in lockstep with the 10-year's daily moves. The question for investors is no longer whether the debt-AI collision could happen, but whether the market has yet fully priced a world where it already has.
(Deepanshu Mohan is Dean and Professor of Economics, OP Jindal Global University. He is a Visiting Professor at the London School of Economics and a Visiting Research Fellow at the University of Oxford. Srisoniya Subramoniam studies Law and is a Research Analyst with Jindal Global University's Centre for New Economics Studies)
Disclaimer: These are the personal opinions of the author
