The Neoclouds promised to democratize the AI. Right now are the most fragile and indebted link in the entire sector

Coreweave, Lambda Labs, Crusoe and Nebius They represent the most booming and also more fragile link in the AI ​​value chain: Neoclouds.

These companies have raised tens of billions in capital and debt to build Data centers full of NVIDIA GPUSbut its business model rests on an increasingly questionable premise: that the demand for computing capacity to continue to grow exponentially.

Why is it important. The problem is not just that these companies lose money. Is that its financial structure depends on a vicious circle:

1. They raise debt to buy GPUS.
2. They use those GPUS as a guarantee for more debt.
3. And the money they enter comes mostly from the same companies that sell them the chips and lend them the money.

The model. The Neoclouds They came promising GPU infrastructure in months, no years, already prices up to 66% cheaper than AWS, Azure or Google Cloud. The proposal sounded well: companies needed GPUS and Hyperscalers (AWS, Azure and Company) did not supply.

The market responded with enthusiasm: Coreweave went from billing 16 million dollars in 2022 to 5,350 million in the last year. Nebius (which has exploded in the stock market and whose germ is Yandex) grew from 5 million quarterly to 105 million. The segment Neocloud As a whole, 82% per year has grown in the last four years.

The problem of the single client. Coreweave generated 60% of its 2024 revenues by renting capacity to Microsoft for Openai. Only Microsoft. Nvidia represents another 15%. If you eliminate a The magnificent seven already openai of the accounts of the main Neocloudsthere are hardly 1,000 million dollars of combined income, As calculated by analyst Ed Zitron.

Lambda Labs has half of his income at Amazon and Microsoft, plus 1.5 billion in a contract with Nvidia. Almost all Nebius’s growth projection comes from A 19,000 million agreement with Microsoft.

• There is no diversified market of business clients.
• There are a handful of technological giants using these suppliers as an exhaust valve or as a vehicle to move money without inflating their own capital expenses Aka Capex.

The money trail. Coreweave owes 25,000 million with annual revenues of 5,350 million. Its debt-active ratio reaches 85.4%. It is like two times your annual salary. And unlike the property that supports a mortgage, the GPUS depreciate quickly.

Nebius He has just closed a 4,200 million round to build the infrastructure that allows you to fulfill your contract with Microsoft. Lambda Labs and Crusoe have raised hundreds of millions in risk and debt capital.

The model is always the same:

1. You get a large contract.
2. You use that contract as a guarantee to raise debt.
3. Purchases Gpus to Nvidia.
4. Rrena more data centers.
5. Repeat.

The problem arises when the Ancla client decides that he no longer needs so much capacity, or when you cannot build the infrastructure quick enough.

Between the lines. Nvidia has invested directly into several Neoclouds And it is also its largest supplier and, in many cases, its largest client. Coreweave signed a 6,300 million agreement with Nvidia a few days ago For the manufacturer to buy any capacity that cannot be sold to other customers until 2032.

In the end we see an elaborate mechanism of Circular financing:

1. Nvidia needs to sell GPUS to maintain its growth.
2. The Neoclouds They need to buy GPUS to fulfill their contracts.
3. The Hyperscalers They need additional capacity but do not want to inflate their capex.
4. And the Private Equity You need to place tens of billions in something that seems the future.

In figures. Building a Data Center Capacity Gigavatio costs between 32,500 and 50,000 million dollars. Oracle and Crusoe took 2.5 years to complete a gigavatio for Openai. Nebius has promised to build multiple gigawatts in increasingly unrealistic terms.

The alarm signal. Coreweave has reported important operating losses in its last quarter despite explosive growth in income. Nebius plans to reach 1,100 million in annual recurring revenues by the end of 2025, almost exclusively driven by the contract with Microsoft.

• What happens if Microsoft decides that you can build your own cheaper capacity?
• Or if Openai, the final customer of much of that capacity, collapses under the weight of their own losses?

The decisive moment. The consolidation has already begun. Coreweave has just bought Core Scientific for 9,000 million in shares. Only great will survive, and probably not many.

It is a matter of time when the adjustment will arrive. The doubt is how much damage will cause when billions in debt collide against the reality that the real demand for GPU capacity is a fraction of what is assumed.

In Xataka | The PC is mutating: the future is filled with AI work stations so you can have your chatgpt at home

Outstanding image | Nebius