AI's Next Bottleneck Is Money

Vantage's search for another $2 billion suggests that the artificial intelligence race is moving beyond chips, models, and electricity into a more complicated contest over who can finance the physical machinery of computation.

By David A. Williams

Artificial intelligence is usually presented as a contest between laboratories. One company develops a more capable model, another secures the latest chips, and governments respond with strategies promising national leadership, digital sovereignty or both. Yet beneath this familiar story lies a rather more prosaic struggle involving loan agreements, credit committees, institutional investors and the increasingly uncomfortable question of how much exposure any lender wants to a single technological wager. Vantage Data Centers' attempt to secure as much as $2 billion in revolving loans from institutional investors including Pimco and PGIM is therefore more than another large financing deal. According to the Financial Times , the proposed facility, known as Project Baja, would allow Vantage to direct capital across several American locations rather than borrow against one designated project. That flexibility is valuable, but the reason it is being sought is more revealing: conventional banks are approaching the limits of what they are willing, or able, to lend to the AI infrastructure boom. The industry has discovered that there is no infinite cloud. Now it may be discovering that there is no infinite balance sheet either.

Vantage has reportedly borrowed approximately $48 billion since 2025 to support hyperscale developments in states including Wisconsin, Texas and Ohio. The number is so large that it changes the meaning of the company's latest request. Another $2 billion is not merely additional capital for another collection of server halls. It represents an effort to widen the circle of institutions carrying the financial burden of the computing race. Banks have traditionally been comfortable financing infrastructure when its economics can be understood through long contracts, predictable demand and assets that retain a reasonably stable value. An airport will still be an airport in twenty years. A pipeline continues to move gas. A toll road does not become technologically obsolete because a rival releases a smarter motorway six months later. Data centers resemble conventional infrastructure in their concrete, cables, cooling systems and power connections, but their commercial purpose is tied to a technological market moving at extraordinary speed. The building may last for decades while the chips inside it age in a few years, or less. That creates an awkward hybrid: a long-lived physical asset supporting a business whose underlying economics may change before the loan has matured.

This is why the arrival of institutional lenders matters. Pension funds, insurers, private-credit managers and bond investors can provide enormous pools of capital, but their participation also makes the financial structure surrounding AI more intricate. Money does not simply move from a bank to a developer and then into a building. It may pass through revolving facilities, construction loans, securitisations, private-credit vehicles, joint ventures and arrangements supported by long-term commitments from hyperscale tenants. Each layer has a purpose. It distributes risk, matches different investors with different stages of a project and prevents individual banks from becoming dangerously concentrated. Yet complexity is never the same thing as safety. It can spread losses so widely that no single institution is destroyed, but it can also spread assumptions so widely that everyone is relying on the same optimistic view of future demand. If lenders, developers and investors all assume that AI consumption will grow almost without limit, diversification may amount to little more than placing the same bet in different financial wrappers.

There is an important distinction here between demand for artificial intelligence and demand for every data center currently being planned in its name. AI will almost certainly become a larger part of the economy, but that does not guarantee that every project will be required, that every location will remain competitive or that every operator will earn the returns embedded in today's financing. Computing may become more efficient. Models may require less energy to perform common tasks. Workloads may shift towards smaller systems operating locally. Customers may consolidate their spending among a handful of providers. Electricity prices, grid delays, planning opposition and water constraints may alter the economics of sites that appeared irresistible when their financing was arranged. None of these possibilities means that the infrastructure boom is imaginary. They mean that genuine technological change can still produce excessive investment. Railways transformed the nineteenth century, telecommunications reshaped the twentieth, and both generated periods in which investors correctly understood the importance of the technology while badly misjudging which assets would make money.

The conventional banks now approaching their exposure limits may therefore be sending the AI industry a signal more useful than another enthusiastic market forecast. A lending limit is not necessarily a verdict that the boom is ending. It is an acknowledgment that even attractive assets can become dangerous when too many of them depend upon the same customers, the same electricity systems, the same chip suppliers and the same expectations of exponential growth. The largest technology companies may possess formidable balance sheets, but much of the physical expansion taking place around them is being financed through a broader ecosystem of developers and investors. That creates a subtle transfer of risk. The companies buying computing capacity can preserve flexibility through contracts, while developers must construct permanent facilities and lenders must decide whether those commitments remain dependable throughout the life of the debt. The celebrated agility of the digital economy rests on some remarkably inflexible foundations.

Project Baja is particularly interesting because a revolving facility usable across several locations would give Vantage greater freedom to allocate capital as requirements change. That may be sensible management, especially when projects are being developed at a scale that makes individual financing packages cumbersome. It also marks a shift away from the comforting simplicity of lending against a particular building with an identifiable tenant and revenue stream. The closer AI infrastructure finance moves towards pools of capital supporting portfolios of projects, the more investors will need to judge the strategy and judgment of the operator itself. They are no longer financing only a site. They are financing an organization's ability to anticipate where computation will be needed, which customers will remain creditworthy, how quickly equipment will become obsolete and whether sufficient electricity will arrive when promised. In effect, lenders are being asked to underwrite a view of the future architecture of artificial intelligence.

That should concern European governments, including Denmark's, because ambitions for digital sovereignty are often discussed without an equally serious conversation about capital. A country may possess suitable land, renewable electricity, cool weather and political stability, yet still find that the decisive choices about its computing infrastructure are made in New York credit markets, American technology companies and global investment funds. Europe cannot achieve meaningful technological independence merely by attracting data centers to European soil. It must understand who finances them, which contracts support them, where losses would fall and who controls the assets if projections fail. A server campus located in Denmark but financed, supplied and commercially directed from abroad may contribute to employment and the electricity bill without giving Denmark much strategic authority over computation. Ownership is not the only measure of sovereignty, but dependence becomes difficult to manage when the financial architecture supporting critical infrastructure is barely visible to the public.

The next phase of the AI race will consequently be fought partly through the price and availability of credit. Companies with strong tenants, reliable grid connections and access to deep pools of institutional capital will be able to build. Others may possess land, planning permission and ambitious presentations but remain stranded between the substation and the loan agreement. This could accelerate concentration because the largest developers will be better placed to absorb delays, negotiate with utilities and assemble complex financing. It could also give private-credit firms and institutional investors growing influence over the geography of computation. Decisions that appear technological will increasingly be shaped by required returns, covenant structures and the appetite of investment committees thousands of miles from the communities expected to supply the land, water and power.

Vantage's proposed $2 billion facility does not prove that an AI debt crisis is approaching, and it would be careless to make such a claim from one transaction. It does, however, expose the point at which the computing boom is becoming inseparable from the financial system underwriting it. Chips determine how much computation can be performed, electricity determines whether the machines can operate, and credit determines whether the buildings exist in the first place. The question is no longer simply whether artificial intelligence can deliver the productivity, revenue and strategic power promised on its behalf. It is whether those benefits will arrive quickly enough to service the financial machinery already being constructed around them. The most consequential algorithm in the AI economy may soon be the one calculating interest.

How This Affects Denmark

For Denmark, Vantage's search for new sources of credit is a warning that data center expansion depends not only on available electricity and grid capacity, but also on decisions made by international lenders. As Denmark seeks to attract AI infrastructure while protecting scarce power resources, tighter financing could favor the largest foreign developers, leaving the country to provide land, energy and grid investment without gaining equivalent control over the computing capacity built on Danish soil.

Sources

Financial Times, “Vantage Data Centers seeks $2bn in loans from Pimco and PGIM”, 11 September 2026

Vantage Data Centers, $5 billion in incremental green-loan financings, 3 June 2025

Vantage Data Centers, $9.2 billion equity investment, 13 June 2024

Federal Reserve Bank of Chicago, analysis of potential bank risks associated with generative AI investment