The AI buildout is now being financed with debt, not cash flow

The bill outgrew the cash flow

The numbers behind the AI buildout no longer fit on a corporate balance sheet. Annual AI infrastructure spending is projected to grow to $1.6 trillion by 2031, and hyperscaler capital spending alone is projected at about $697 billion in 2026. Alphabet guided 2026 capital spending of $195 billion to $205 billion, and Amazon about $220 billion. Even for the largest companies in the world, that is a financing problem, not just a spending plan. m:…6479

The gap is being filled with borrowed money. Up to one-third of AI capital spending is now being financed in bond markets, and outstanding private credit to AI-related companies has grown to more than $200 billion. A further $800 billion in data center financing could come from private credit over the next two years. The pattern is visible at the project level too: CoreWeave used a $7.5 billion GPU-backed debt facility, with the chips themselves as collateral. m:…6479

The demand side of the ledger explains why the borrowing keeps growing. Anthropic reported $16.3 billion in revenue during the first half of 2026, but it also incurred nearly $9 billion in computing and infrastructure expenses in the same period, and it has outlined nearly $800 billion in spending commitments over the next decade. Nominal global AI investment is forecast to rise to $1.8 trillion in 2027, adding 0.4 percentage points to global GDP growth. The spending is real, and so is the financing need behind it. m:…6479 m:…5773

Power and memory set the pace

Borrowing can pay for the buildout, but it cannot fix the two physical limits on it. The US data center power shortfall is projected to reach the mid-30 gigawatt range by 2028, and permitting delays in Texas and Ohio are causing a pause in data center growth rather than project cancellations. Memory is currently the most significant component bottleneck for hyperscalers, and semiconductor supply chain tightness is expected to increase in 2027. m:…6115

The market is currently pricing in a 2028 slowdown for AI chips, which is likely too bearish, and the dissonance between Silicon Valley optimism and Wall Street pessimism regarding AI is currently at an all-time high. That disagreement is itself part of the financing story: lenders are being asked to fund a buildout whose end demand is priced very differently on the two sides of the market. m:…6115

Some builders are answering the financing question with discipline rather than volume. TensorWave, a cloud provider building on AMD hardware, requires a signed customer contract before financing a new deployment, so capacity buildouts are tied to actual demand rather than speculation. Its customers are requesting more CPUs in their next AI deployments and higher CPU-to-GPU ratios, because AI agents need CPUs to handle tasks between model calls. Piotr Tomasik identifies power as the industry's ultimate constraint, followed by memory supply. m:…0303

What the bond market is being asked to believe

The shift from cash flow to debt changes who carries the risk if the buildout slows. When hyperscalers funded data centers from earnings, a pause cost them growth. When bondholders and private credit funds finance them, a pause costs lenders, and the collateral is hardware that loses value fast. That is the trade behind the GPU-backed facility and the private credit surge: lenders are betting the demand is durable enough to outlast the hardware. m:…6479

Bond markets look willing, for now. Elevated starting yields have improved the appeal of the traditional 60/40 balanced portfolio, and the firm rates high-grade and investment-grade bonds as attractive, favouring short- to medium-duration maturities. UBS also points to private infrastructure as a source of inflation-linked income with exposure to long-term investment in electricity networks and data centers. The money is available; the open question is what happens to it if power and memory keep the buildout slower than the financing assumes. m:…7663