Just one-third of US data centre pipeline seen credible as AI funding strains rise

Markets have so far repriced only parts of the AI trade rather than the whole complex. The toughest pressure is on firms that depend on outside capital to fund construction — Oracle, SoftBank, and Cipher and TeraWulf, two former bitcoin miners now hosting AI. By contrast, cash-rich hyperscalers such as Microsoft, Alphabet, Meta and Amazon are much less vulnerable, and some of their investors may even be happy to see capital spending slow. The next layer of risk sits with Nvidia and the broader chip supply chain, because a portion of future demand relies on customers that still need to line up financing. In Asia, order delays would surface first at SK Hynix, Samsung and TSMC, a real concern for a chip-heavy Kospi index after its recent rally. At the index level, AI construction has been a major support for US earnings and growth, which means even a deceleration rather than an outright collapse would remove a large prop from global equities and, through risk appetite, the Australian dollar. Stress that extends beyond the neoclouds — for instance, hyperscaler capex cuts or wider credit spreads — would signal a move from a sector repricing to a broader market problem.

The binding constraint on the AI boom has moved away from chips and toward permits, power and credit, and investors have begun to price the gap between announced data centre capacity and what can realistically be financed and built.

Key points:

• No fewer than 47 sizeable US data centre projects have been delayed, withdrawn or blocked this year, a count the data provider deems conservative.
• Analysts consider only around a third of the planned US pipeline credible, while more than 1,000 municipalities have pursued curbs.
• This week saw a sharp drop in AI infrastructure debt: a CoreWeave bond moved into distressed territory, and CoreWeave’s stock fell nearly 25% during the week.
• Listings planned by SB Energy and Firmus have been postponed or parked, OpenAI has delayed its IPO, and Nscale’s $35bn listing target appears difficult to reach in the near term.
• Clauses in some contracts let Anthropic and similar tenants walk away when financing or delivery targets are missed.
• Spending on US AI infrastructure is forecast to surpass $10tn from 2025 through 2032, roughly 3.6% of GDP annually.

The Financial Times reports that the US AI infrastructure boom, expected to cost over $10tn, is starting to strain as building hold-ups, community resistance and a sudden tightening in credit hit the firms hurrying to erect data centres.

A regulatory intelligence firm counts at least 47 large US projects delayed, withdrawn or blocked this year, and says the figure is conservative because it records only the largest facilities. Analysts estimated this month that only about a third of the planned US data centre pipeline is credible, and more than 1,000 communities have asked for moratoria or other curbs on AI infrastructure in 2026.

Pressure has now spread to the funding side. AI-infrastructure-related debt fell sharply this week: an unsecured CoreWeave bond sank into distressed territory, and CoreWeave’s market value shrank by nearly a quarter during the week. Bonds backing facilities leased to CoreWeave also softened, among them a Virginia issue down about 5% since pricing three weeks ago. With interest rates higher and inflation climbing, there is a sharper debate about whether OpenAI, Anthropic and other leading labs will generate enough revenue to meet their costs.

A large share of construction relies on the neoclouds — young data centre operators that expanded quickly on abundant equity and debt — and their plans to go public are stalling. SoftBank-backed SB Energy has not completed a data centre and has shelved a planned $50bn listing, despite needing around $175bn of capital to deliver on roughly $430bn of contracted future revenue, most of it from OpenAI. If Nscale misses financing or delivery deadlines on its $45bn Anthropic arrangement at a West Virginia project, the deal can be cancelled; the company had just 25,000 advanced chips online in August, against a goal of close to 500,000. Nebius reports that it has secured financing for only about a fifth of the roughly $250bn needed to back its planned capacity. OpenAI has likewise postponed its own stock market debut.

The setbacks come with a large price tag. Last month, Cipher agreed to absorb about the first $360mn of cost overruns at a Texas site in return for an extended schedule; that adjustment needed roughly $750mn in additional equity, some of it raised through an $850mn convertible bond this week. Some Oracle facilities for OpenAI’s Stargate programme have also fallen behind, and Stargate ceded its computing lead this week.

Rather than accept a shortage of computing capacity, AI labs are willing to put part of their investment at risk, and their contracts give them the option to exit if partners miss deadlines. Permits are now harder to win, and utilities are increasingly rejecting grid connections for operators that cannot self-supply power; the months ahead will decide how much of the announced capacity becomes reality.