Who can afford the AI buildout? The spending debate turns to balance sheets

The new question

On Bloomberg Surveillance, panelists discussed a shift in how the AI investment story is told. The panel's framing was that the AI story is moving from who spends the most to who can afford to, and that the current infrastructure build-out is in year three of an eight to 10 year cycle. AI demand has accelerated approximately 30% over the last month, according to the panel discussion. That combination — a long cycle, accelerating demand, and a new affordability test — reframes the debate. The question is no longer whether the buildout continues, but whose balance sheet carries it. m:…0757

The panel also drew a quality line inside the market. Profitable Russell 2000 stocks are up 18% year-to-date, while non-profitable stocks are in negative territory, according to the discussion. Hyperscalers have seen a significant drop in year-over-year growth in cash flows, leading to increased debt financing, the panel said. And OpenAI has lowered its valuation target to $125 billion in recent funding talks, according to the panel. These are panel claims, not audited figures, but the pattern they describe is consistent: cash flow is tightening while spending is not, so the funding mix is changing. m:…0757

Debt fills the gap

A separate Bloomberg Surveillance discussion made the financing point more directly. AI investments are increasingly turning to debt financing as free cash flows are exhausted, the speakers said. Peak AI investment is forecast to reach 3% of GDP, compared with 1.5% during the internet era, according to the discussion. Lazard believes the current AI capital expenditure is necessary to build full-stack capabilities for long-term productivity, the speakers said. That is the bull case for the debt: the spending builds a durable platform, not a speculative asset. m:…5592

The same discussion added two context points that cut in opposite directions. Japanese corporate governance reforms have improved return on equity from 5% to 10% over the last decade, according to the speakers, which is a reminder that governance reform can unlock returns without new spending. At the same time, the rate of zero-day vulnerability exploitation has risen to 80% in 2026 from 20% in 2018, the speakers said. The implication for an executive is that the AI stack is getting more expensive to finance and more expensive to defend at the same time. Debt that looks affordable in a growth model looks different when security costs and rates are both rising. m:…5592

The measurement problem

Fast Money panelists put the enterprise version of the problem bluntly. Less than 15% of enterprise chief financial officers in North America can measure the return on their AI investments, according to the panel. At the same time, the share of open-source AI models used by North American companies has grown to 48% of total token usage. Those two numbers belong together. If most CFOs cannot measure returns, cheaper open models are an attractive way to keep experimenting without committing to the most expensive contracts. Open source becomes the measurement strategy: run it cheaply, learn what works, and only then pay for scale. m:…0473

The panel also discussed the surrounding market signals. American Express is cutting 5% of its global workforce as part of an AI overhaul, according to the panel, and investment grade credit spreads are as wide as they have been since March of last year. Williams-Sonoma has outperformed major tech stocks including Meta, Apple, and Microsoft over the last three years, the panel said. None of those claims proves an AI bubble or disproves one. Together they describe a market that is rotating: capital is asking harder questions about who benefits from AI spending, and rewarding companies whose earnings do not depend on the answer. m:…0473

What it means for budgets

For an executive approving an AI budget this quarter, the three discussions converge on a practical test. First, can the company afford the spend from cash flow, or is it adding debt to a balance sheet that is already carrying a long buildout? Second, can the finance team measure the return, or is the spend justified only by a competitor's spending? Third, is there a cheaper open-model path that delivers the same learning? The Surveillance panel's affordability framing, the Lazard and IMF debt warning, and the Fast Money measurement gap are not three separate stories. They are three ways of asking whether the AI buildout is being paid for by earnings, by debt, or by hope. The next earnings season will start to answer that, one cash flow statement at a time. m:…0757 m:…5592 m:…0473