Wall Street is scrutinizing so-called “take-or-pay” contracts for computing power, viewed as the most vulnerable link in the financial chain supporting the AI buildout. These agreements—such as the recent deal between Akamai and Anthropic—bind frontier AI labs to cloud platforms for computing capacity, and can extend upstream to hardware makers and power providers. Under these terms, customers must pay for GPUs, server time, and other inputs regardless of actual usage, with payments recorded as backlog revenue until data centers become operational.
While AI revenue from large language models is expected to grow as adoption widens, models are also becoming commodified and cheaper. There is no guarantee that cash flows will align with fixed payment schedules once they begin. “OpenAI doesn’t know whether it will need that compute or whether it can afford it, yet it committed to paying a set amount monthly for three to five years. That is extremely risky for the compute user,” Gil Luria, head of technology research at DA Davidson, told CNBC. “If they haven’t raised enough capital to cover contracted compute, the entire system collapses.”
Take-or-pay contracts are proliferating across the tech sector. Neocloud CoreWeave derives most of its revenue from such deals, per a Morgan Stanley note, while Alphabet lists take-or-pays within its $811 billion in purchase obligations. Because they represent future payment promises rather than formal debt, they constitute a major component of the sector’s off-balance-sheet commitments. BCA analysts called investor nervousness “overblown,” though noting that off-balance-sheet debt is making markets uneasy.
Pinpointing the exact scale remains difficult, as contracts are often private and buried in footnotes. The Wall Street Journal estimated the total at $3 trillion; Nikkei put it at $1.65 trillion. Anthropic has committed $750 billion for compute through 2030, and OpenAI could spend over $500 billion on inputs in the next decade. The Bank for International Settlements flagged that these projects substitute upfront capital for multi-year operating expenses, leaving Wall Street without clear revenue visibility.
Citizens Securities questioned how Oracle would fund its data center buildouts to meet cloud targets, citing a $153 billion gap needing coverage from compute cash flow or equity. Meanwhile, Nvidia saw a 63% surge in accounts receivable in the first half of the year, with the take-or-pay chain now “running unbroken from the hyperscalers to the foundry,” analyst Greg Miller noted.
Originating in the energy sector to justify long capital-intensive timelines, take-or-pay structures are now standard in tech. Energy legal specialists warn of a “coming wave of disputes” over data center projects. “History suggests that in any large project-financed sector, disputes arise during operations, usually after the first material market dislocation,” said James Barratt of Vinson & Elkins, who called the scale and concentration of commitments “extraordinary.”
Banks support off-balance-sheet vehicles between private creditors and hyperscalers, raising systemic stability concerns. The BIS warned that funding lines could create shock transmission channels through refinancing pressure or procyclical credit shifts. Meta Platforms, for instance, will lease Louisiana’s Hyperion data center through such a vehicle, with rent used to repay investors. Yet AI revenues remain insufficient to service this debt, prompting four Democratic senators—Elizabeth Warren, Richard Blumenthal, Chris Van Hollen, and Tina Smith—to question the assumptions in a January letter to the Treasury.