AI infrastructure needs $8.5 trillion. Credit markets see a $4.9 trillion gap
Gartner says data centre owners need $8.5 trillion of revenue through 2030 but forecasts at most $3.6 trillion. Oracle's default swap has widened 28 basis points in a week as lenders charge more to wait.
On Thursday, the cost of insuring the debt of America's largest technology companies hit record highs. Samsung Electronics fell 9.9 percent in Asian trading, while SK Hynix and Kioxia fell more than 11 percent.
The move was about the cost of money. Gartner has put a number on the revenue gap, and debt financed capacity is the weak link because its payback arrives late. Oracle's five year default swap, the price of insuring its debt, was quoted at 288 basis points on Thursday, up from 260 a week earlier.
Gartner's release of 14 October gives three separate numbers. Data centre owners need $8.5 trillion of cumulative revenue through 2030. That covers announced capital spending, financing costs and a 12 percent return on invested capital, the hurdle Gartner assumes.
Gartner forecasts no more than $3.6 trillion of AI revenue attributable to data centres. Subtracting that from $8.5 trillion leaves a $4.9 trillion shortfall. Forecast revenue covers 42 percent of what owners need.
The other two figures describe the damage if spending continues at its current pace. More than $3.5 trillion of infrastructure spending could be exposed to impairment, meaning written down, by the end of 2030. More than $4 trillion of asset value is exposed through 2032.
The first figure is roughly half of the nearly $7 trillion Gartner expects providers to spend between 2026 and 2030. The source of the gap is easy to miss: Gartner expects less than half of AI revenue to go to providers that own significant AI infrastructure.
Model developers that rent capacity capture the revenue, while the owners carry the depreciation. That puts the exposure on leveraged owners of a single asset class, who have only rental income to service debt. Gartner expects more than 80 percent of neoclouds to cease independent operations by 2030. Neoclouds are specialist operators that rent out graphics processors.
Hyperscalers, the largest cloud providers, have other businesses to carry their borrowings. The gap therefore lands unevenly.
Gartner also estimates spending on AI optimised infrastructure rental will grow from $66 billion in 2027 to $174 billion in 2030. In its words, that growth is too low for most AI compute providers to recover their investments before the current generation of AI chips is replaced.
The market linked the move to money. The Bureau of Labor Statistics reported on Wednesday that consumer prices rose 0.6 percent in September, with energy up 5.2 percent in the month. Over twelve months, prices rose 3.7 percent.
Charu Chanana of Saxo described the moves as "less like a sudden deterioration in technology fundamentals and more like a repricing of the cost of capital." Chanana said investors "are not questioning whether AI demand exists."
That interpretation matters because Gartner does not expect the 12 percent return before the end of 2032 at the earliest. A payback that far out is a long dated claim, and long dated claims lose the most value when yields rise.
Rates widen the gap directly. The $8.5 trillion includes financing costs, so every point added to borrowing rates raises the revenue owners must earn for the same return. A rates shock widens the gap without a single customer cancelling.
The same swap data shows who pays most. SoftBank's swap widened by 24 basis points to 424. Swaps on Amazon, Alphabet, Meta and Nvidia are also at records, though they have risen far more slowly than Oracle's. The market charges most to insure the borrowers whose revenue is least diversified, which is the order Gartner's own distinction predicts.
For a data centre developer or neocloud with a refinancing inside the year, the choice is concrete. Fixing term debt at widening spreads costs money now. A take or pay contract with a hyperscaler lets lenders underwrite the revenue and lowers the financing cost, but it caps the upside if capacity becomes scarce. Spot capacity pays more in a boom and offers little collateral in a bust.
The IMF's Global Financial Stability Report warned policymakers against complacency over a pullback in AI investment and valuations. Jason Wu told journalists that if questions arise over monetisation, "the equity investments will deplete in value, the debt may be difficult to service."
In the downside scenario of the Fund's World Economic Outlook, a steep fall in AI investment cuts global output by 1.1 percent by 2028. The US loses 1.8 percent.
Governor Pan Gongsheng of the People's Bank of China spoke to G20 finance chiefs in Bangkok on Thursday. He said sharp swings in government bond markets and a heightened risk of correction in AI related valuations have increased financial fragility. Higher sovereign yields raise the discount rate on AI cash flows just as those cash flows are being questioned.
Strong supplier earnings do not answer this. TSMC's profit is up 68.7 percent, but it is paid by the companies doing the spending. Supplier earnings record the outlay, not the return, so they lag.
Telecom and fibre owners learned the same lesson in the early 2000s. The users of the networks prospered, and the builders took the write downs.
Oracle's quarterly results in December provide the next reading. If its five year swap holds at or below 288 basis points, the market is still willing to wait. Gartner does not see the 12 percent return before the end of 2032. A move above that line says lenders have stopped.