AI slowdown justified, debt-fuelled bubble is the bigger risk

Estimated reading time: 5 minutes · Last updated: 20 September 2026

AI slowdown calls are justified, but the immediate financial risk may be a debt-fuelled bubble underwriting the current AI buildout, not only runaway models. As first reported by The Guardian, hyperscalers including Google, Amazon, Microsoft, Meta and Oracle have been linked to an estimated $132bn (£99bn) of debt issued this year to fund datacentre expansion. That borrowing sits against 10-year US treasury yields hovering at about 5% and a set of contingent “compute commencement” obligations analysts say add up to far larger future payments. The scale and timing of those obligations, and falling prices for AI usage, sharpen the risk of a sudden market reappraisal.

The price of AI is collapsing, while the cost of building it is not.

Bloomberg

Why the safety debate doesn’t erase the economics

Calls for a measured slowdown in AI development follow recent incidents and high-level warnings about model behaviour and misuse. Those public-safety questions matter on their own terms. But the financial plumbing that runs under the AI push is distinct: firms have committed capital and future payments on the assumption that customers will generate far higher revenue than they do today.

Heather Stewart highlights a separation between the safety debate and the financing model that underwrites rapid capacity expansion. Even if regulators impose tighter guardrails on model development, that would not by itself resolve contractual and balance-sheet tensions between hyperscalers and the frontier labs that need compute.

Debt issuance and a fragile bond backdrop

The hyperscalers’ datacentre push has been backed by sizable bond and loan issuance: one estimate puts the total at $132bn this year, with the same source giving a sterling equivalent of £99bn. That issuance comes as a global reference for borrowing costs — the 10-year US treasury yield — hovers at about 5%, raising the cost of rolling and servicing long-term debt.

Higher financing costs make aggressive buildouts harder to justify on narrow unit margins. If credit markets tighten or investors reprice the risk of large technology borrowers, routines that have supported rapid capacity expansions could become much more expensive overnight.

Unit economics and the compute‑commencement problem

Customer prices for processing AI workloads are falling even as the hardware and facility costs remain elevated. A Bloomberg report quoted the blunt summary: “The price of AI is collapsing, while the cost of building it is not.” Silicon Data’s LLM token expenditure index, which tracks what customers pay for a million tokens, has more than halved since June and now sits at less than $1 per million tokens.

Compounding that mismatch is what Groundbreaker calls a $1.5tn compute‑commencement wall. Many datacentre deals are structured on take‑or‑pay terms: the hyperscaler books expected revenue up front while the buyer defers accounting for the later payments. Groundbreaker’s tally implies roughly $700bn of such obligations coming due next year and more than $800bn in 2027 — a concentrated timing risk if end users don’t generate matching revenue or find cheaper options.

How a financial shock might ripple beyond tech

The Groundbreaker comparison to the 2007 teaser‑rate mortgage cliff is deliberately stark: when contingent payments rise and underlying cashflows fall short, defaults and renegotiations spread through borrowers, lenders and investors. If major labs or hyperscalers must renegotiate take‑or‑pay deals or if bondholders demand higher spreads, the effects would extend into debt markets and suppliers of semiconductors and real‑estate services.

There is also a political angle: some observers in the material note that a government‑backed slowdown or regulatory pause could act as a moat that preserves incumbents’ lead and market share. Others warn such protection would merely postpone a reckoning of the underlying unit economics. The practical upshot is that market, policy and technology risks are now tightly interlinked.

The case for and against a disorderly unwind

The bottom line

The safety risks from AI deserve urgent attention, but they do not cancel the financial questions that sit beneath the buildout. Large, timed obligations and heavy borrowing — estimates show $132bn of hyperscaler issuance this year and a potential $1.5tn compute‑commencement wall — expose tech, suppliers and lenders to a coordinated shock if revenues fall short. Policymakers and market participants should therefore watch both the regulatory path for models and the calendar of contractual payments; each can tip whether the current expansion stabilises or unwinds disorderly.

What to watch

• Watch for the 2027 spike in compute‑commencement obligations; Groundbreaker projects more than $800bn will fall due in 2027.
• Watch for the wave of roughly $700bn in compute‑related payments expected next year, as identified by Groundbreaker.
• Watch how 10‑year US treasury yields — currently hovering at about 5% — move through 2027, since they are a key comparator for hyperscaler financing costs.

This article is information, not financial advice. Anyone acting on it should do their own checks.