AI infrastructure is increasingly being financed by companies whose balance sheets look nothing like those of Microsoft, Amazon or Alphabet. Independent operators such as CoreWeave, Nscale, Lambda and IREN are building data centres, acquiring GPUs and signing long leases in an attempt to capture demand that hyperscalers cannot satisfy alone.
The obvious bull case is that AI compute remains scarce for years. The less comfortable comparison is the telecom fibre buildout before the dot-com crash, when new networks were financed on the assumption that internet traffic would grow faster than capacity could ever catch up.
Leverage changes the AI infrastructure trade
A hyperscaler can fund a data centre from a diversified cash-generating business. A neocloud often has to finance hardware, property and power before customer revenue arrives. That makes customer concentration, financing cost and asset depreciation much more important.
Reuters Breakingviews notes that CoreWeave's total liabilities reached $72 billion at the end of June, 56% higher than at the start of the year. The sector also increasingly relies on GPU-backed bonds, special-purpose vehicles and vendor support from Nvidia.
Our view: the danger is technological overcapacity, not simply too many data centres
Global Markets Review's view is that the most useful lesson from the fibre boom is not that AI demand must collapse. It is that technology can increase effective capacity faster than investors expect. More efficient chips, smaller models, better inference software or new architectures can change the economics of installed hardware even while AI usage keeps growing.
Investors should therefore track utilization, customer concentration, refinancing costs and depreciation assumptions rather than headline megawatts alone. The boom can be real and still produce weak returns for the companies that finance the wrong generation of infrastructure.
| Risk | Why it matters | What to watch |
|---|---|---|
| Leverage | Large fixed financing burden | Debt and refinancing costs |
| Customer concentration | A few AI labs can dominate demand | Top-customer revenue share |
| Hardware depreciation | GPUs may lose economic value quickly | Useful-life assumptions |
| Technology change | Efficiency can create effective overcapacity | Utilization and pricing |
Frequently asked questions
What is an AI neocloud?
A neocloud is an independent cloud or data-centre operator focused heavily on AI compute, often using large fleets of GPUs or other accelerators.
Why compare neoclouds with 1990s telecom firms?
Both sectors involve capital-intensive infrastructure financed ahead of uncertain future demand and exposed to rapid technological change.
Does this mean AI data centres are a bubble?
Not necessarily. The comparison highlights financing and capacity risks even if underlying AI demand remains strong.