AI data centres stall, neoclouds under pressure: what it means for rented GPU capacity

Two reports from 18 and 19 October paint the same picture of the market for AI compute. On 18 October 2026 heise summarised Financial Times research: building AI data centres in the US is stalling. On 19 October The Decoder followed up with what a shake-out among neoclouds would mean for Nvidia. If you rent GPU capacity, these are the numbers to know.

Sources: heise (18 October 2026, based on the Financial Times) and The Decoder (19 October 2026, based on Gartner, Bloomberg, The Information and the Telegraph). Forecasts come from third parties and are not checked by us. (Quelle: Simple Icons, file nvidia.svg (CC0))

What is being built and what is blocked

According to heise, AI investment is expected to reach 3.6 percent of US GDP, a Columbia Business School figure. Railway building peaked at 2.2 percent, mid-century road building at 1.1 percent, and the internet build-out of 1996 to 2003 at 1.1 percent.

At the same time projects are stuck. Data from the compliance firm RegLayer, quoted by the Financial Times, shows 47 large US data centre projects delayed, cancelled or blocked in 2026. The FT also counts more than 1,000 US municipalities with their own rules for AI infrastructure, some of them moratoria.

What happens on the markets and at the providers

CoreWeave's stock lost about 25 percent in the week of the report, per heise; the FT points to higher interest costs and worries about US inflation. Bloomberg, as relayed by The Decoder, has CoreWeave losing a quarter of its market value and a measure of its credit risk hit a record. The Australian provider Firmus pulled its IPO, and some Nscale investors want the targeted valuation cut from about $30 billion to $16 to 19 billion.

Nscale also shows how much rides on a few customers. More than 40 percent of its contracted revenue of $103 billion comes from Microsoft, the Telegraph reports. Microsoft has also walked away from a project in Loughton in the UK, where it wanted to rent 23,000 Nvidia chips; grid power there is not expected before the 2030s. Per heise, Nscale had 25,000 of a planned 500,000 chips online in August.

Four different quantities in one chart: a share, a forecast, an estimate and and a price drop that cannot be added up. (Quelle: The Decoder, 19 Oct 2026)

Gartner, Bernstein and the consolidation

The forecasts come from The Decoder. Gartner expects 80 percent of independent neoclouds to stop operating independently by 2030 and sees $3.5 trillion of AI infrastructure spending exposed to impairment by the end of 2030 (the figure is in the Gartner release, not in The Decoder text). Bernstein estimates that only 35 percent of announced North American capacity will actually be built. Two managers told The Information that of hundreds of providers only seven to ten will remain in a few years. Cirrascale CEO David Driggers expects consolidation too. CoreWeave CEO Michael Intrator locates the turbulence at smaller rivals such as Firmus, Lambda and Nscale.

Nvidia has a lot at stake. Customers outside the hyperscalers, neoclouds included, brought in $40 billion in the July quarter, 45 percent of data centre revenue and more than twice the year before. Nvidia props up the sector with stakes, chip allocations and credit guarantees, and, according to an unnamed manager, it changed the terms of those guarantees after larger neoclouds turned down a first offer.

What follows for renters and builders

Neither report quotes GPU rental prices or the debt of individual providers. Satya Nadella is quoted expecting prices to fall with oversupply from 2027, but no source documents a GPU rental price path, only a contract risk: if you book capacity from a small, debt-financed provider, you depend on its funding and on the power supply at its site. That is our reading and appears in neither source.

We run our own cards in our own hosts: four RTX 5090 in one host. That protects against a provider failing, but not against the market shortage we describe in our post on the RTX 5090. If you rent, ask about term, termination rights and data export, and keep your models stored so that switching provider takes days.