AI infrastructure is becoming bankable, but not every AI project deserves financing

The announcement that NVIDIA is working with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilise more than $500 billion of third-party capital for AI infrastructure is not simply another large number in an industry that has become accustomed to announcing increasingly spectacular investment programmes. It marks a more profound shift in the way capital markets are beginning to look at compute infrastructure.

For years, the data center industry has progressively moved from a specialist telecom asset class into a mainstream infrastructure market. The first generation of facilities was deeply connected to telecommunications, carrier hotels, network interconnection and enterprise hosting. The cloud then changed the scale of the market by introducing hyperscale campuses, long-term capacity commitments and investment-grade counterparties whose credit quality allowed lenders to underwrite data center revenues with increasing confidence.

AI infrastructure is now forcing the industry through another transition, but the financing logic is more complex. Until recently, financing AI-oriented infrastructure was materially more difficult than financing a conventional hyperscale data center because the physical asset was only part of the credit story. Lenders were not simply being asked to finance a building with power. They were being asked to finance a building, high-density electrical infrastructure, cooling systems, GPUs with relatively short technological cycles and, in many cases, revenues generated by counterparties with limited balance-sheet history.

That distinction matters because lenders do not finance megawatts. They finance contracted cash flow, recoverable collateral and credible counterparties.

From telecom infrastructure to financeable compute

This is precisely why the market has changed so quickly over the last three to four years. As companies such as CoreWeave, Lambda, Crusoe, Nebius and Fluidstack have scaled, the contractual architecture supporting AI infrastructure has started to look more like traditional infrastructure finance. Contracts are becoming larger, longer and increasingly backed by hyperscalers, AI laboratories or other investment-grade entities. At the same time, banks and private credit providers have become much more comfortable underwriting assets whose value is linked not only to the building but also to compute equipment and long-term customer commitments.

CoreWeave is one of the clearest examples of this transition. In May 2026, the company closed a $3.1 billion delayed-draw term loan facility backed by high-performance computing infrastructure and dedicated customer contracts. The transaction was meaningfully oversubscribed and pricing tightened by 50 basis points during syndication. Moody’s rated the facility Ba2 and Fitch BB+. What matters here is not that CoreWeave suddenly became equivalent to Microsoft or Google from a credit perspective. It did not. What matters is that institutional lenders were prepared to finance AI infrastructure at scale despite the absence of investment-grade corporate credit, because the financing structure, the infrastructure and the underlying contracted revenues were considered sufficiently robust.

Lambda provides another useful example. In May 2026, it closed a $1 billion syndicated senior secured credit facility intended to finance new NVIDIA infrastructure and additional data center capacity. Less than a year earlier, the equivalent facility had been $275 million. The increase is significant because it shows how quickly debt markets are adapting to the economics of AI infrastructure.

The same evolution can be seen on the real estate and development side. JLL has reported that traditional enterprise data center leases historically tended to run for approximately five to seven years, whereas hyperscale leases are increasingly extending beyond ten years. For lenders, that difference is critical because longer contractual duration allows a larger proportion of the development cost to be financed with senior debt. JLL has also reported financing levels reaching up to 85 percent loan-to-cost for some pre-leased developments with strong credit counterparties, with pricing in the low 200 basis points over SOFR.

These are infrastructure finance metrics, not venture capital metrics.

This is why the current wave of institutional interest is so important. NVIDIA’s initiative with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR is not simply about finding more equity for data centers. The more significant development is that the financial market is beginning to create mechanisms that treat compute infrastructure itself as an investable and financeable asset. Reuters has reported that the contemplated financing structures may use NVIDIA hardware as collateral, while other reporting has suggested that mechanisms linked to residual equipment value could be part of the financing architecture.

That is a fundamental shift because residual value has always been one of the major obstacles to financing GPUs with conventional debt. A building may remain economically useful for several decades, whereas compute hardware depreciates much more rapidly and is exposed to technological obsolescence. The more comfortable lenders become with the ability to redeploy, refinance or recover value from GPU infrastructure, the easier it becomes to expand the debt market supporting AI development.

In my view, this is the beginning of the financialisation of compute.

We have seen this process in other infrastructure sectors. Aircraft became financeable because lenders could underwrite lease income and recover an asset that could be redeployed. Telecom towers became infrastructure because long-term tenancy produced predictable cash flow. Renewable assets became highly financeable once power purchase agreements transformed variable production into contracted revenue. Data centers followed the same path when hyperscale leases made future income sufficiently predictable for banks, insurance companies and infrastructure funds.

AI infrastructure is now moving in the same direction, but with an additional layer of complexity because the financeable asset is no longer simply land, power and a building. It is increasingly a combination of powered land, electrical and cooling infrastructure, compute equipment, customer contracts and credit support.

More capital does not mean more bankable projects

That combination creates an entirely new capital stack. Land and power can be financed by infrastructure investors. The building can be financed through conventional development debt. GPUs can increasingly support equipment-backed financing or asset-based lending. Long-term compute contracts can support cash-flow-based facilities. Stabilised assets can then potentially be refinanced through institutional debt, securitisation or infrastructure capital.

This is why I do not believe the real story is that AI infrastructure suddenly has access to unlimited capital. That interpretation is too simplistic and, in my view, fundamentally wrong.

The real story is that good AI infrastructure is becoming much easier to finance, while poor AI infrastructure is likely to become much harder to finance.

The distinction will be made by credit quality.

A 100 MW project with secured power and a 15-year lease may look attractive in a presentation, but those numbers are almost meaningless without understanding who ultimately supports the revenue. If the tenant is a recently established AI infrastructure company with limited equity and only five years of credit support from an investment-grade entity, then lenders may effectively underwrite five years of credit rather than a 15-year lease.

JLL has made this point explicitly in its analysis of data center financing, and it is one of the most important issues for the market to understand. Lease duration and credit duration are not necessarily the same thing.

This is where I believe the AI infrastructure market is currently most vulnerable. The extraordinary level of demand has attracted a very large number of new developers, GPU operators, brokers, intermediaries and supposed AI infrastructure platforms. Some of them will become major operators. Others are attempting to secure hundreds of megawatts of land and power without the balance sheet, contractual depth or downstream demand required to support those projects.

The market should therefore stop treating every AI tenant as equivalent.

It matters whether the customer is Microsoft, Meta, Google, OpenAI, Anthropic, CoreWeave or a newly incorporated company with little equity and an ambitious business plan. It matters whether the contract is genuinely take-or-pay. It matters whether there is a parent guarantee, a letter of credit or another form of credit support. It matters whether the end user is identified. It matters whether the compute equipment can be redeployed if the customer defaults. It also matters whether the technical design remains usable for future generations of GPUs.

This is why due diligence on AI infrastructure must now go far beyond the traditional real estate model.

Historically, developers and investors could focus primarily on land, power, permitting, construction cost and tenant covenant. That is no longer sufficient. A serious underwriting process for AI infrastructure now needs to analyse the entire chain from power availability through to the ultimate source of compute revenues.

The relevant sequence is land, power, grid certainty, permitting, building design, cooling architecture, compute equipment, customer contract, tenant credit, end user and residual value. Weakness at any point in that chain can materially alter the financing profile of the project.

Bankability will become the real differentiator

This is also why I believe the market is approaching an important period of differentiation.

Over the past several years, simply controlling powered land was enough to create substantial value because access to electricity was the dominant constraint on data center development. Power will remain critical, but it will no longer be sufficient on its own.

The next stage of the market will reward developers who can combine secured power with technically credible infrastructure, financeable counterparties and contractual structures that institutional capital can actually underwrite.

The numbers already show that this transition is underway. NVIDIA and its financial partners are discussing more than $500 billion of third-party capital for AI infrastructure. CoreWeave has raised $3.1 billion through a syndicated HPC-backed facility. Lambda has secured $1 billion of senior secured financing. JLL has observed up to 85 percent loan-to-cost on certain pre-leased data center developments. Data center ABS issuance increased by approximately 40 percent year on year during the first half of 2025.

These figures point in the same direction. Compute is progressively becoming infrastructure.

But the arrival of institutional capital should not be interpreted as a validation of every project currently being marketed as an AI data center. Institutional capital does the opposite. It introduces discipline, underwriting standards and a much sharper distinction between assets that generate durable cash flow and assets that simply benefit from a fashionable narrative.

My view is therefore straightforward. The next bottleneck in AI infrastructure will not simply be capital. Capital is becoming increasingly available for the right projects. The real bottleneck will be bankability.

The projects that succeed will be those that can demonstrate real power rather than theoretical grid capacity, real customers rather than speculative demand, genuinely long-term credit support rather than headline lease terms and infrastructure that can remain technically relevant as compute hardware evolves.

The $500 billion NVIDIA initiative is important not because it means Wall Street is prepared to finance every AI data center. It is important because Wall Street is beginning to build the financial architecture required to decide which AI infrastructure deserves to be financed and which does not.

That distinction will become one of the defining features of the next phase of the AI infrastructure market.