Leadership1 publisher3 min readPublished
Datacentre deals let hyperscalers book revenue years before payment is due
A Guardian column by Heather Stewart argues the nearer risk in AI is financial, counting $132bn of hyperscaler debt issuance this year and a $1.5tn compute bill that lands within about two years.
The Board Room · Leadership desk
What happened
- Heather Stewart argued in the Guardian that a collapse of the AI bubble, with repercussions far beyond the US, is the more immediate threat. It is worth asking, she wrote, whether the sector is sailing towards a financial iceberg.
- Debt issuance funding the datacentre rollout by the hyperscalers Google, Amazon, Microsoft, Meta and Oracle comes to $132bn this year alone, on one estimate cited in the column.
- A research note from the financial analyst Groundbreaker sets out a $1.5tn compute commencement wall facing the AI labs over the next couple of years. The note likens it to teaser mortgage rates expiring in 2007 and 2008.
- Anthropic told investors its adjusted operating income was positive, on a measure that effectively excludes many of the company's costs.
Compiled by The Board RoomSomething wrong?How this is made
Why it matters
- exposure A roadmap resting on five vendors inherits their refinancing calendar, so the terms on which capacity is offered move with bond markets rather than with model quality.
- constraint A token price under $1 per million limits how much of the buildout sellers can fund from revenue, so the discount a buyer wins this year is the same number squeezing the seller who has to keep serving it.
- decision Contracts signed this budget cycle mature inside the two-to-three-year window when take-or-pay deadlines land. Post-deadline price and capacity belong in the clause set at signature.
- contradiction A vendor's safety advocacy says nothing about its balance sheet: the same proposal can be governance to a regulator and protection to an incumbent carrying multibillion-dollar debts.
The debt total on its own is the smaller worry for a buyer. At the roughly 5% benchmark Stewart cites for 10-year US treasury yields, this year's hyperscaler issuance carries about $6.6bn of annual interest before any corporate spread [1]. Five firms of that size can service that. The $1.5tn that the analyst Groundbreaker puts in front of the AI labs is a different order: roughly eleven times a single year of that issuance, and Groundbreaker dates it inside the next couple of years [2] [11].
Behind those commitments sits a contract structure that connects a systemic worry to a procurement one. In Groundbreaker's account, as set out in the column, datacentres are built and kitted out on take-or-pay contracts. Payment starts only when a deadline is hit, often two to three years out, and the site starts running [12]. In the interim, the hyperscaler that built it books the value of the contract as expected future revenue [13].
Prices are moving the other way. A Bloomberg report quoted in the column said: "The price of AI is collapsing, while the cost of building it is not." [5] Silicon Data's index of what customers pay per million tokens has more than halved since June, to less than $1 [7]. Demand for semiconductors and other real-world datacentre components keeps costs elevated [8]. OpenAI has repeatedly cut its fees to hang on to customers [6]. A buyer renewing this quarter takes the benefit of that; Stewart wrote that the maths works only on an assumption of "epic revenue growth" [18].
The vendor's own profitability is the part a buyer can least verify. Cory Doctorow, the digital rights campaigner quoted in the column, wrote: "These companies are claiming that they are so cool that their profitability can only be measured using a novel, secret form of mathematics." [10]
Stewart holds two positions at once, and both matter to a procurement team. She wrote that there is ample evidence AI urgently needs regulating, pointing to Meta's smart glasses filming people without consent and to the lack of safeguards that allowed swarms of chatbots to go on a hacking spree [16]. She called suggestions such as independent analysis of AI models important improvements on the ungoverned status quo [17]. She also warned that a small number of intricately linked megafirms carrying multibillion-dollar debts may be hoping the state will throw a regulatory moat around them [14]. A government-backed pause, she warned, could for instance keep cheaper Chinese options from encroaching on Silicon Valley's market dominance [15].
Stewart is arguing systemic risk, not enterprise procurement. Her numbers still leave a buyer one narrower question to put to a vendor: which entity carries the debt on the capacity being sold, and when the take-or-pay deadline on that capacity falls [12] [13].
What to watch
- Whether Silicon Data's token price index keeps falling below $1 per million tokens or steadies.
- Whether any of the five named hyperscalers restates how it recognises take-or-pay datacentre contracts as expected future revenue.
- Whether independent analysis of AI models becomes a rule, and which indebted incumbents it shelters.