Wall Street is paying close attention to so-called "take-or-pay" contracts for computing power as the link in the financial chain where the artificial intelligence buildout could weaken – and maybe even snap. Take-or-pay contracts, such as the one agreed to Friday between Akamai and Anthropic, are agreements that frontier labs sign with cloud platforms to secure computing power. They can also govern upstream deals with hardware makers and electrical power providers. The agreements make customers pay for graphics processing units (GPUs), server time and other inputs regardless of whether customers ever actually use them.

The payments are booked in backlogs as expected revenue but generally don't start coming in until under-construction data centers become operational. AI revenue for top-of-the-line large language models will increase as the technology becomes widely adopted. But models are also expected to get cheaper as AI becomes increasingly commodified, and there's no guarantee that revenues will arrive precisely in time to meet pre-set payment schedules once the clock starts running.

"OpenAI doesn't know whether or not it's going to need that compute or whether it's going to even be able to pay for the compute, and yet they made a three- or five-year commitment to pay a certain amount every month, regardless. So it's very risky for the user of the compute," Gil Luria, head of technology research at DA Davidson, told CNBC. "[If] they haven't raised enough capital to pay for the compute that they've signed contracts for, the whole system collapses," he added. Off-balance sheet debt Take-or-pay contracts are proliferating throughout the tech sector during the AI buildout. Neocloud CoreWeave gets the majority of its revenue from take-or-pays, according to an April 15 note from Morgan Stanley. Alphabet lists take-or-pays as a piece of its $811 billion in purchase obligations.

Since they are future promises to pay as opposed to formal debt, take-or-pay contracts are a big part of the tech sector's so-called off-balance sheet commitments. Analysts with BCA wrote Friday that "off-balance sheet debt is making investors nervous," but concluded that concerns were "overblown."

It's hard to put a definitive number on the level of those commitments — or to disaggregate take-or-pay contracts from other types of obligations — since they're mostly recorded in footnotes of public filings across different companies, and the contracts themselves can be private. One tally by the Wall Street Journal put the total at $3 trillion, and another by Nikkei put it at $1.65 trillion.

Among specific large language model (LLM) labs, Anthropic has committed $750 billion for compute through 2030, according to the Journal , and OpenAI could spend more than $500 billion for various inputs in the next 10 years, The Information reported.

Wall Street does not have a clear line of sight on revenue for some of these projects, which substitute upfront capital spending for multi-year operating expenses, as the Bank of International Settlements (BIS) noted in a March analysis .

For example, Citizens Securities said in a Sept. 17 note that it's unsure where Oracle will get the money for promised building projects aimed at reaching its cloud revenue targets. "Customer prepayments would fund a portion of this requirement, but the size and timing cannot be accurately estimated over the given time frame," analyst Greg Miller wrote, citing a $153 billion shortfall that would "still need to come from compute [cash flow from operations] or equity."

In the case of Nvidia , which reported a 63% surge in accounts receivable during the first half of this year, "the take-or-pay chain now runs unbroken from the hyperscalers to the foundry," Miller said.

Bracing for a 'wave of disputes' Take-or-pay contracts have their origins in the energy sector, where long and capital-intensive production timelines justified a similar legal ultimatum. The U.S. tech sector is now finding the same structure advantageous.

Energy sector legal specialists are gearing up for a " coming wave of disputes " related to data center projects undergirded by take-or-pay agreements. "History suggests that in any large, project-financed sector, disputes often arise in the operational phase, generally only after the first material market dislocation," James Barratt, partner at Vinson & Elkins, a law firm with an energy specialty, wrote in July . He described the scale and concentration of take-or-pay capacity commitments in the tech sector as "extraordinary."

Banks support off-balance sheet arrangements, such as special purpose vehicles, between private creditors and hyperscalers, which is raising concerns about systemic stability, especially in light of contractual activations. "Banks support the vehicles with funding lines, potentially creating new shock transmission channels – eg via refinancing pressures at the vehicle level, procyclical shifts in private credit appetite or the activation of guarantees," the BIS report said.

Meta Platforms , for example, will lease the Hyperion data center in Louisiana from such a special purpose vehicle, and revenue from the rent checks will then be used to repay investors. But AI revenues overall are not substantial enough yet to service these debts, a group of four U.S. senators have noted.

"AI companies assert that while they may not be generating sufficient revenue now, data center buildouts will eventually turn profitable. However, it is unclear whether that assumption is accurate," Democratic Senators Elizabeth Warren (Mass.), Richard Blumenthal (Conn.), Chris Van Hollen (Md.) and Tina Smith (Minn.) wrote to the Treasury last January.