That is according to Jonathan Mauck, senior managing director at Digital Bridge Holdings.
Speaking in a fireside interview at Datacloud USA x Metro Connect Fall 2026 in Austin, Mauck said data centres are built as 20-year industrial infrastructure, while the GPUs inside them have a useful life of five to seven years.
“You’re effectively a 20-year creditor,” Mauck said, describing the risk facing operators who commit capital to a facility on the assumption that a tenant can keep paying rent over two decades, against hardware that needs replacing several times over.
Nvidia announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR in August to establish independent compute financing platforms aimed at mobilising the capital, of which Nvidia may itself backstop up to $125bn. The structures are designed to turn Nvidia compute into a financeable, investable asset class in its own right, separate from the civil and electrical infrastructure that houses it.
Mauck was measured on the mechanics. “That looks like a version of round tripping my capital,” he said, framing the arrangement as Nvidia enlisting financial partners to effectively backstop GPU sales rather than guaranteeing buybacks itself. He argued the structures are best suited to neo-clouds and other operators without investment-grade credit ratings, rather than hyperscalers such as Microsoft and Amazon that already have ready access to capital. A separate, larger example of the same funding mismatch is playing out in Nvidia’s reported talks to guarantee up to $250bn in financing for OpenAI’s roughly $500bn Ohio data centre project.
Inference reshapes the topology
Mauck also pointed to a structural shift as AI workloads move from training to inference. Large training campuses continue to be built, he said, but growth is increasingly concentrated in smaller, 20 to 40MW facilities in tier two and tier three markets, close to end users, a model he compared to the enterprise colocation sites of several years ago.
He linked the trend to a broader globalisation of compute demand, citing growth across Latin America, Asia and Europe, partly driven by capacity constraints and local opposition to new builds in North America, an issue Capacity has continued to track (community backlash; cost, power and pushback).
Bubble concerns and a return-on-capital question
Asked whether AI infrastructure investment constitutes a bubble, Mauck drew a contrast with the 2001 downturn, arguing that today’s capital is backed by cash-generative businesses rather than speculative valuations alone. But he said revenue still needs to catch up to the scale of capital deployed, warning that any hyperscaler pulling back on capital expenditure “would result in a pretty volatile year” across the sector.
He predicted the next major growth driver would be the “industrialisation of AI” through robotics and real-world applications, citing significant investment from the US, China and Japan.
RELATED STORIES
Nvidia’s $500bn Wall Street pact turns AI chips into assets
Notes from Austin: Fibre builders and data centre developers talk cost, power and community pushback
Metro Connect USA 2027
Metro Connect USA is the largest executive-level digital infrastructure event in the U.S. The only one of its kind, this 25-year-strong gathering is where decision makers come together to make deals happen.
