Until recently, renting computing power meant dealing with one of three names: Amazon, Microsoft or Google. The AI boom has created a new type of landlord.

Companies including CoreWeave, Nscale, Nebius and Crusoe have emerged as so-called neoclouds, specialist infrastructure providers built around one increasingly valuable commodity: access to enormous clusters of advanced GPUs. Rather than attempting to compete with Amazon Web Services or Microsoft Azure across every corner of cloud computing, these firms specialise in providing the chips, data-centre capacity and power required to train and run artificial intelligence models.

On paper, the economics look compelling. AI laboratories need more computing power than the traditional cloud providers can always supply, Nvidia's latest GPUs remain highly sought after, and every new generation of models appears to demand more infrastructure than the last. The neocloud becomes the landlord, AI developers become the tenants and increasingly expensive Nvidia chips become the property being rented out.

Investors have understandably taken notice. Crusoe raised $3.9 billion this month at a $30.9 billion valuation1, while Nscale is reportedly targeting a valuation of around $30 billion as it prepares to list in New York2.

Yet railways are useful because their tracks can remain productive for decades. GPUs are rather different. They depreciate quickly, require enormous amounts of power and are being purchased by companies increasingly dependent on debt and a remarkably small collection of customers. The question is therefore not whether demand for AI infrastructure exists. It clearly does. The more interesting question is whether the landlords can make enough money before the buildings need rebuilding.

The Toll Roads of the AI Gold Rush

The bullish argument for neoclouds begins with scarcity. If OpenAI, Anthropic or another AI laboratory needs tens of thousands of advanced GPUs immediately, waiting several years for the traditional cloud giants to build additional capacity is not particularly useful. Neoclouds have stepped into that gap, securing GPUs, land and electricity before renting the resulting computing power back through long-term contracts.

Few companies demonstrate the scale of this opportunity better than CoreWeave. Second-quarter revenue reached $2.58 billion, more than double the same period a year earlier, while its revenue backlog climbed to $104.2 billion. Management responded by raising expected 2026 capital expenditure to between $35 billion and $39 billion3.

Nebius tells a similar story. Its second-quarter revenue reached $582.3 million, while the company said it had accumulated more than $40 billion of customer commitments. Management has argued that demand remains so far ahead of supply that it could sell its entire planned 2027 capacity on current terms4.

But Nscale perhaps provides the most striking example of how quickly enthusiasm has developed. Revenue increased 1,252% year-on-year during the first half of 2026 to $140.6 million, while total contracted value has reportedly grown beyond $103 billion in only two and a half years. That sounds extraordinary until another number appears on the same income statement: Nscale lost $1.02 billion over those six months2.

This is where I become less comfortable with the railway comparison. Railways became valuable because once the expensive track had been laid, passengers could be carried across it for decades. Neoclouds are spending billions on infrastructure whose most valuable component can become yesterday's technology after only a few product cycles.

For the moment, scarcity hides many of those problems. When customers are fighting for compute, providers can charge attractive prices and sign enormous commitments years into the future. But investors buying these companies today are implicitly betting that today's shortage of AI infrastructure remains tomorrow's pricing power. If the hyperscalers catch up, model development becomes more efficient or an oversupply of GPUs emerges, the toll road could suddenly become considerably less exclusive.

When the GPUs Age Faster Than the Debt

The real vulnerability appears on the balance sheet. CoreWeave had $35.6 billion of total indebtedness at the end of June, while interest expense reached $985 million in the first half of 2026 alone5. At the same time, the company warns that keeping its infrastructure competitive requires older equipment to be continually replaced.

That creates an uncomfortable mismatch. The debt used to finance the infrastructure can remain outstanding for years, while the economic advantage of the GPU bought with that debt may disappear much sooner. Nvidia's next generation of chips does not need to make today's hardware useless. It only needs to make someone else's data centre faster and cheaper.

Customer concentration adds another layer of risk. CoreWeave's two largest customers accounted for 40% and 23% of first-half revenue5, while Nscale's largest customer represented 52% of its revenue2. Long-term contracts provide visibility, but there is a difference between recurring revenue and diversified revenue.

None of this means the neocloud model cannot work. Demand is real and the companies controlling scarce GPUs and power may become vital pieces of the AI economy. But investors should be careful not to confuse today's infrastructure shortage with a permanent competitive advantage.

The neoclouds may indeed be building the railways of the AI revolution. I am just not convinced the tracks will last as long as the loans taken out to build them.

Footnotes

  1. Crusoe, Crusoe Raises $3.9 Billion Series F for its Vertically-Integrated AI Infrastructure Platform (opens in a new tab), 17th September 2026.

  2. Reuters, Nvidia-backed AI cloud firm Nscale reveals revenue surge in US IPO filing (opens in a new tab), 18th September 2026. 2 3

  3. Reuters, CoreWeave boosts 2026 spending plan, beats quarterly estimates on AI demand surge (opens in a new tab), 11th August 2026.

  4. Reuters, Nebius powers past estimates as customers race to secure AI computing power (opens in a new tab), 12th August 2026.

  5. CoreWeave, Quarterly Report for the Quarterly Period Ended June 30, 2026 (opens in a new tab), 11th August 2026. 2