47 AI data centers stalled in 2026 as grids, towns, and lenders push back
At least 47 large US data center projects have been delayed, withdrawn, or blocked so far in 2026, according to regulatory-intelligence firm RegLayer — a count the firm itself calls conservative since it only tracks the biggest sites. More than 1,000 US communities have tried to impose moratoriums or restrictions on AI infrastructure this year, and data center executives report utility regulators increasingly refusing grid connections to projects that can't supply their own power. The pattern is consistent enough to name: the constraint on the AI buildout has moved from chip supply to permits, electricity, and financing.
The money at stake is genuinely unusual
Columbia Business School economist Stijn Van Nieuwerburgh, in a paper for the Brookings Institution, projects US AI infrastructure spending exceeding $10 trillion between 2025 and 2032 — roughly 3.6% of GDP annually, about one and a half times the peak of 19th-century railroad construction and more than triple the 1996–2003 internet buildout. Bernstein estimated this month that only about 35% of the currently planned US data center pipeline is credible, meaning the gap between announced capacity and capacity that will actually get built and connected to power may be the largest in the industry's history.
Where the strain is actually showing up
The clearest cases involve companies that pre-committed capital or signed contracts against capacity they haven't delivered yet. Nscale, a London-based neocloud, had 25,000 advanced chips online in August against nearly 500,000 planned — a 5% delivery rate against its own roadmap. Anthropic's $45 billion deal with Nscale for West Virginia capacity includes a walk-away clause if Nscale misses financing or delivery deadlines, which converts a construction delay directly into a contract-cancellation risk. Nscale is still targeting a New York IPO at a $35 billion valuation despite the shortfall.
Nebius, the Amsterdam-based cloud provider, illustrates the financing math at its starkest: CEO Arkady Volozh told a Citi conference in September that the company's planned 5 gigawatts of capacity would require roughly $250 billion, of which only about a fifth had been financed. SoftBank-controlled SB Energy needs approximately $174 billion in capital to deliver $430 billion in contracted revenue — mostly from OpenAI — and has delayed its own listing at a $50 billion valuation. Separately, Firmus shelved its IPO entirely, OpenAI pushed back its own listing plans, and several sites tied to Oracle's $500 billion Stargate program have slipped their schedules.
Delays are compounding into credit problems
When a data center project slips, the cost overruns don't disappear — they get renegotiated, usually in the builder's favor and the financier's expense. Cipher Mining agreed last month to absorb the first $359 million of cost overruns at a Texas site leased to Fluidstack in exchange for more construction time, a concession that required about $750 million in additional equity, partly funded through an $850 million convertible bond priced on October 14. TeraWulf has separately reworked lease terms on Anthropic-linked projects. These are not distressed-company workarounds; they're becoming a standard part of how neocloud financing works in 2026.
Credit markets are starting to price that risk in directly. CoreWeave shares lost almost a quarter of their value in a single week, one of its unsecured bonds slid into distressed territory, and a $1.1 billion bond tied to a Virginia data center leased to CoreWeave fell roughly 5% within three weeks of its original pricing. None of that is a cloud-industry-wide collapse — it's a repricing of the specific companies that borrowed heavily to build capacity ahead of confirmed long-term demand.
Consolidation is now the consensus expectation
Data center and neocloud executives told The Information they expect only seven to ten neoclouds to survive over the next few years, down from the hundreds currently attempting to compete, with consolidation potentially starting as early as late 2027. Gartner's own forecast, published the same week, predicts more than 80% of independent neoclouds will cease operating independently by 2030. Both numbers describe the same underlying mechanism: undercapitalized builders with uncertain power access losing ground to a small number of players who can actually secure grid connections and financing at scale.
Who is actually exposed, and who isn't
It's worth being precise about where the risk concentrates, because “AI buildout is slowing” gets reported more broadly than the facts support. The strain sits primarily with debt-funded developers and neoclouds — companies that built or are building speculative capacity to lease out — rather than with the largest cloud providers, who have more diversified balance sheets and existing power relationships. AI labs themselves are largely protected by the walk-away clauses in their own contracts, which shift delivery risk onto their infrastructure partners rather than absorbing it directly. Two of the most exposed publicly visible names, Nscale and Nebius, are both European companies whose valuations depend specifically on delivering US capacity on a schedule that increasingly looks unrealistic.
What this means going forward
If you're evaluating a neocloud or data-center-linked investment, the diligence question that matters now isn't “how much capacity is planned” but “how much of the required power has an actual interconnection agreement, and how much of the capital stack beyond equity has committed financing.” Those are the two inputs that determined which projects among the 47 delayed sites actually stalled. For enterprises planning AI infrastructure dependent on third-party cloud capacity, the practical takeaway is to treat vendor capacity timelines with real skepticism and build contractual protection — similar to the walk-away clauses AI labs have already secured for themselves — rather than assuming announced capacity will arrive on the announced date.