CoreWeave and the neoclouds: the AI bubble's leveraged edge

What are neoclouds, and how do they earn money?

A neocloud is a cloud company built around AI computing. It buys accelerator chips, houses them in data centres, and rents the capacity by the hour or under long contracts. Large cloud providers sell many services, while a neocloud sells mainly this one, so its fate follows the demand for AI computing closely.

Customers include model builders and large companies that need capacity faster than they can build it. When demand is strong, a neocloud fills its machines and earns well on each one. When demand softens, the machines still cost money, which is why the structure behind the revenue matters as much as the revenue. Read each quarterly report with both questions in mind: how full are the machines, and who pays for the machines.

Why does borrowing make neoclouds the leveraged edge of the AI bubble?

Chips and data centres cost a great deal before they earn anything. Many neoclouds borrow against the hardware or against signed contracts to pay for it. Borrowing lets them grow faster than their own cash would allow, and it fixes a schedule of interest and repayments that continues in good and bad quarters.

That is the meaning of leverage here: a small change in revenue causes a large change in what is left for owners. Upside rises when customers pay on time. Pressure rises when a few of them delay, renegotiate or leave. The two checks that follow measure how large the cushion is. A cushion is the space between what the business earns and what it owes in the same period.

How does customer concentration raise the risk at a neocloud?

Annual filings state the share of revenue that comes from the largest customers. When a few buyers account for most of it, one decision changes the picture. A customer who is itself a young model builder depends on outside investors, so its ability to pay rests on future funding rounds.

The EQOS Asia AI study records 95% of AI projects failing in production (MIT 2025) and about $200B a year invested (April 2026). Many buyers are still searching for returns. Established customers with steady workloads and long contracts give a firmer base than a handful of fast-growing, cash-hungry ones. Ask how long the contracts run and whether they carry firm minimum payments.

How fast does AI hardware lose value, and why does that matter?

Accelerator chips improve with each generation, so older machines lose rental value over time. The accounting life chosen for the equipment affects profit: a longer life lowers yearly cost, and a shorter life raises it. Compare the stated life with the pace at which new chips arrive.

Debt secured by hardware depends on what the hardware is worth. If prices for rented capacity fall, the value of the collateral falls with them, and lenders may ask for more protection. Ask what share of loans rests on equipment and how the lender values it. Loan terms that tie repayment to hardware value add a second source of pressure when rental prices fall.

Which 5 checks measure the room neoclouds have for a slow year?

Check 1: backlog of firm contracts against debt. Check 2: interest cover, which is operating profit divided by interest cost. Check 3: share of revenue from the 3 largest customers. Check 4: debt due in the next 2 years against cash on hand. Check 5: accounting life of equipment against the pace of new chips.

Mark each check green, amber or red on the date you read it. Three reds in one quarter is the signal to treat the risk as high. Record the answers every quarter, so the trend becomes visible and a single odd period does not mislead you. A written record also makes it easy to explain your view to colleagues.

What does trouble at a neocloud mean for the wider AI market?

Neoclouds are customers of chip makers and power suppliers, and suppliers of model builders. A neocloud that cuts orders reduces sales for the chip makers behind it, while one that fails forces its customers to move their work. The effect travels both ways along the chain.

Because the group is small, it works as an early signal. Stress here, such as delayed projects or renegotiated contracts, can appear before it shows in larger names. Read it as a prompt to check the wider map of exposed companies, and avoid reading it as a verdict on the whole market.

How can a buyer of AI capacity use the neocloud reading?

If your company rents capacity from a neocloud, run the five checks on it as a supplier. Ask for service commitments, keep your data portable, and plan a second source for critical work. The EQOS Asia method page explains how to combine such checks into a score.

This page is general information, not investment advice. It frames risk and leaves prices and dates open. For a larger commitment in Singapore, ASEAN or Hong Kong, the Λ simulator gives a first score in 10 minutes, and the 30-minute pre-diagnostic with EQOS Asia applies it to your case.

Frequently asked questions

About the author

Stéphane Jambu

President, EQOS Forecast Asia

Stéphane Jambu leads EQOS Forecast Asia and works as a consultant in e-commerce, search visibility, and AI. LinkedIn profile