Hyperscaler AI Capex Arms Race Is Breaking Internal Governance and Balance Sheets

Alphabet, Amazon, Microsoft, Meta and Oracle will spend somewhere between $700bn and $900bn on capex in 2026, and for the first time the aggregate exceeds what they generate in cash. The question worth asking is not whether the spend pays back. It is which corporate function gets hollowed out to fund the gap, and which board director is left holding the explanation.

The numbers have crossed a line the operating model was not built for

For two decades, hyperscaler capex was a residual of operating cash flow. Build what the cash allowed; the balance sheet stayed pristine; the equity story was capital-light compounding. That model died this year. CreditSights puts the big five's 2026 capex at $700-900bn, up 36% on 2025, with roughly $450bn tied directly to AI infrastructure. [1] Amazon alone has guided to $200bn, more than double 2025. Alphabet has roughly doubled its own guidance to $175-185bn. [2]

The ratio that matters is capital intensity. Capex as a share of revenue is now running at 45-57% across the group. [3] That is a utility number, not a software number. The equity research community, the compensation committees, and the internal resource-allocation processes at these firms were not designed to evaluate it. When a software business with 30% operating margins reinvests at utility intensity, three things break simultaneously: the multiple the market is willing to pay, the depreciation schedule underpinning reported earnings, and the hurdle rates governing which projects get funded.

The group raised $108bn of new debt in 2025 to plug the gap. [4] Morgan Stanley and JPMorgan model another $1.5tn of tech-sector issuance over the coming years. [5] No rating agency has yet placed a hyperscaler on watch on the basis of AI capex debt load. That silence is itself an open question, and one a CFO should be asking her treasurer about.

The revenue side is more circular than the headline numbers admit

Bulls point to AWS at +28%, Google Cloud at +63%, Microsoft AI at a $37bn run rate up 123% year on year. Those are real numbers. They are also, in part, the same dollar being counted multiple times. TS Lombard's Perkins has flagged that roughly 50% of the combined order backlog at Microsoft, Oracle, Google and Amazon comes from OpenAI and Anthropic, neither of which currently earns enough to meet those obligations from end-customer revenue. [6] Microsoft funds OpenAI, OpenAI buys Microsoft compute, Microsoft books the revenue, the cloud growth rate looks settled. The dollar has not yet touched an external customer.

Sequoia's David Cahn puts the gap between hyperscaler AI spend and industry-wide AI revenue at $600bn annually, and it is widening. [7] Allianz Research has the divergence between capex growth and revenue growth running at 46%, against 32% in the 2001 telecom cycle. [8] Telecom is the wrong analogue for the technology, but the right one for the financing structure: long-lived assets, vendor-financed demand, end-customer monetisation lagging by years. The MIT Project NANDA work cited by Forbes found 95% of enterprise GenAI pilots produced zero measurable P&L impact on $30-40bn of corporate spending. [9] McKinsey's own survey puts the share of companies scaling AI across the enterprise at 13%. [10]

Even Sam Altman, who has the least incentive of anyone to concede the point, told CNBC alongside the Stargate Michigan announcement: "I know some great stuff is happening but there is also a tonne of waste. How long do I have to wait for it to show up in revenue? How long do I have to wait to really get costs under control?" [11] When the principal seller of the product is asking the timing question publicly, the counterparty risk for the enterprise buyer is no longer theoretical.

The second-order risk sits inside the customer, not the vendor

This is where the brief diverges from the standard capex-bubble take. The material exposure is not whether Microsoft's share price corrects. It is what happens inside the Fortune 500 buyer when AI spend starts cannibalising other line items in the IT budget.

McKinsey's Noshir Kaka, surveying roughly 690 executives, found 72% are increasing tech budgets by 6-8% on average, but AI infrastructure costs are crowding out legacy spend. [12] His phrasing is worth quoting: "Every enterprise that I know on average is increasing the spend on technology. And yet, when you look at the trickle down that's actually happening in services…services feels like it's in a recession." [13] The money is going to the GPU layer and being taken from elsewhere in the same technology budget: security tooling, compliance modernisation, data quality, integration work, vendor management.

Bain's April 2026 survey of 951 companies found executives approving incremental AI spend on the basis of projected savings that have not materialised, and identified data access as the single largest reason AI programmes underperform: "Despite a decade of investments in data modernisation running well into hundreds of billions of dollars globally, the No. 1 reason AI programs underperform is that companies cannot reliably get access to their own data." [14] Companies are funding new AI spend by stripping investment from the data and security infrastructure that AI requires to work. Bain's own framing: "Self-funding the next wave from past returns sounds like discipline. In reality, it is a circular bet with a structural leak." [15]

The internal governance evidence is starting to surface. Microsoft terminated Claude Code licences after internal AI usage "ran past the annual AI budget months ahead of schedule." [16] Uber is reported to have exhausted its full-year 2026 AI budget in four months. [17] Amazon shut down an internal employee leaderboard tracking AI activity after discovering staff were running unnecessary autonomous bots to climb it. [18] These are not exotic cases. They are early evidence of a control-environment problem that audit committees have not yet been briefed on, because finance functions lack the telemetry to see token consumption the way they see headcount or T&E.

The counter-case, and why it is weaker than it looks

The bull case is straightforward and has been right for three years. Cloud revenue keeps compounding. Capex turns into depreciation, which becomes the cost base for the next wave of AI services, which monetises through enterprise software and consumer surfaces. Ciena's Gary Smith and hyperscaler management teams argue, plausibly, that the buildout is demand-driven. The cash-generating engines at Microsoft, Alphabet and Meta remain among the most productive in the history of capitalism. Even at $200bn of capex, Amazon's operating cash flow can carry it.

The strongest form of the bull argument is that critics have been consistently early, and being early is indistinguishable from being wrong. Perkins of TS Lombard concedes this directly: "revenues and stock prices surging" has caused the market to declare the sustainability debate over. [19]

Two problems remain. First, the bull case requires the capex-to-revenue divergence to start closing, and it is doing the opposite: Allianz puts the 2026 gap above 2001 telecom levels. Second, even if the hyperscalers themselves are fine, the enterprise customer is not. The bull case conflates the hyperscaler's balance sheet capacity with that of the enterprise buyer who funds the revenue line. Those are separate questions. The PitchBook analysis of B2C earnings calls found 90% of the 186 AI-related analyst questions were "probing and investigative," with ROI displacing product roadmap as the dominant theme. [20] The sell-side has already moved. The buy-side and the audit committees will follow.

What to watch

1. Rating agency action on a hyperscaler before Q2 2026 earnings. No outlook change has yet been published despite $108bn of incremental debt in 2025. [21] A negative outlook from Moody's or S&P on any of the big five, citing capital intensity, would mark the moment debt markets begin pricing AI capex as utility-like rather than tech-like. Falsifiable: either it happens by July or it does not.

2. A Fortune 500 CFO disclosing an AI-related budget overrun in a 10-Q risk factor or MD&A. Microsoft and Uber are the precedents inside leaked reporting. [22] The first explicit mention of token-consumption variance in a public filing will trigger immediate auditor and audit-committee attention across the index. Watch the Q1 2026 filing cycle.

3. Cloud revenue growth rates decoupling between hyperscalers. If AWS, Azure and Google Cloud diverge by more than 15 percentage points in a single quarter, the capex-recycling thesis advanced by TS Lombard gains evidence: the differential will reveal who is selling to end customers and who is selling to the OpenAI/Anthropic pair. [23] Convergence at high rates supports the bulls; divergence supports the governance critique.