The financing debate is not just about who borrows for new data centers. According to the dossier, neocloud returns can flatten below roughly 80% utilization, making durable customer commitments a more useful signal than announced capacity alone.
Summary
A Brookings conference draft by Stijn Van Nieuwerburgh examines how leases, joint ventures, project finance, private credit and structured vehicles finance the AI buildout. Its central concern is that these structures can expand funding while moving leverage and demand risk to individual projects. The sharper operational question is whether each facility can sustain enough contracted use to support its economics.
What happened
The draft frames AI infrastructure as a financing problem as well as a construction problem: assets need capital, but the eventual returns depend on demand and cash flows. It describes financing through leases, joint ventures, project finance, private credit and structured-finance vehicles, and warns that liabilities can be less visible in a hyperscaler's headline balance-sheet leverage even as project-level leverage rises. This matters because a data center or compute fleet is not productive merely because it has been announced or built. Its owners and lenders need tenants, utilization and cash flows that last long enough to justify the investment. The available primary excerpt identifies the paper and its conference-draft status, while the supplied source summary reports its estimates and analysis; those projections should not be mistaken for completed capacity or proven returns.
What the archive adds
The dossier attributes to McKinsey’s neocloud analysis the finding that returns can flatten below roughly 80% utilization and that some operators depend on only one or two customers. This is a third-party estimate, not a measurement of every provider or project. It supplies a concrete lens for the paper’s financing concern: tenant concentration and actual usage can matter as much as how a project is financed. The archive contains no distinct earlier case that would establish how this specific buildout will perform.
Trend signal
The trend is toward treating compute capacity as an investable asset whose economics depend on both infrastructure delivery and customer usage. Financing structures can distribute the cost across owners, tenants and capital providers, but they cannot create demand. As a result, capacity announcements may conceal a practical divide between facilities backed by durable workloads and speculative supply awaiting customers. Utilization and tenant quality are therefore useful signals for judging the buildout, although neither alone settles whether a project will earn an adequate return.
Why now
According to the primary source summary, leases, joint ventures, project finance, private credit and structured-finance vehicles are being used to fund infrastructure assets, shifting attention to project-level cash flows and tenant exposure. The dossier’s finding d2 attributes investment scenarios to expected compute-demand growth and notes sensitivity to assumptions including chip supply, enterprise adoption, efficiency and regulation. Finding d3 attributes data-center growth to AI, especially inference, alongside cloud migration and traditional enterprise workloads. Together, these point to both expanding demand expectations and uncertainty about which workloads will occupy the resulting capacity.
Connections to other events
The non-obvious constraint is utilization backed by creditworthy tenants. According to the dossier’s account of McKinsey’s analysis, neocloud returns can flatten below roughly 80% utilization, while some operators rely on only one or two customers. Our read is that a financing vehicle can make construction possible without making the underlying risk disappear: if a large tenant reduces usage or leaves, a project may struggle to generate the cash flows its lenders and owners expected. That makes customer commitments and concentration more revealing than a capacity figure in isolation.
Why it matters
Builders and buyers should distinguish capacity that is contracted and being used from capacity that is merely available or planned. Providers seeking financing may need to show credible tenant commitments; buyers should recognize that a supplier dependent on a small customer base could face pressure if usage changes. For AI teams, this is a reason to ask prospective providers about utilization and customer concentration, not just performance and availability. It does not imply that every concentrated provider is fragile: the terms and durability of commitments matter.
What happens next
- More public reporting on specialist compute providers will discuss customer commitments or utilization as evidence for planned expansion, rather than relying only on capacity announcements. (6 months, 60%). Watch for: A provider, lender or investor disclosure links a specific capacity expansion to named or described tenant commitments or utilization.
- At least one public analysis of a specialist compute provider will identify customer concentration as a material risk to financing or expansion. (12 months, 55%). Watch for: A financing disclosure, investor report or company commentary explicitly discusses dependence on a small number of customers in connection with funding or growth.
Limits and uncertainty
The primary item is a conference draft, and the supplied excerpt does not independently verify the projections or establish that planned projects will be built or repaid. The utilization threshold and customer-concentration observation come from a dossier summary of McKinsey’s neocloud analysis; they may not apply uniformly across operators or facility types. We do not have project-level utilization, contract terms, financing disclosures or evidence of defaults. Strong long-term commitments, changing workloads or different project economics could weaken this take.
Skill to build now
Practice customer-and-utilization diligence. This week, choose one cloud or GPU provider you may use; record its disclosed utilization, largest-customer exposure and whether capacity is backed by commitments. Label each answer as evidenced, asserted or unknown, and use the gaps to shape your next vendor questions.
Reader takeaway
This week, add two questions to one cloud or GPU-provider evaluation: what share of its capacity is actually committed and used, and how dependent is it on its largest customers? Record whether the answers are disclosed or merely asserted. Treat missing utilization and concentration data as diligence gaps rather than assuming announced capacity is dependable supply.
Open question
Will durable tenant commitments become visible enough for buyers and lenders to distinguish productive AI capacity from projects that depend on optimistic utilization assumptions?