I keep coming back to one simple thought. What happens when the machines that power the biggest technology boom in decades start losing value faster than the loans used to buy them can be repaid? That question sits at the heart of the AI infrastructure boom, and it is not getting enough attention yet.
Private credit has stepped in to finance the massive buildout of data centers. The collateral? Graphics processors that can lose most of their economic usefulness in four to six years. The tenants? A small handful of the largest technology companies. When one of those tenants decides not to renew, lenders may suddenly find themselves owning rows of aging chips in remote warehouses. I believe Washington will eventually step in, just as it did with the mortgage giants after the last crisis. The pattern feels familiar, even if the assets look completely different this time.
Why Private Credit Is Funding The AI Infrastructure Boom
The scale of spending is hard to grasp. Forecasts put global data center capital expenditure near three trillion dollars through the end of the decade. Roughly half of that money goes into chips and servers. The rest covers buildings, power, cooling, and networking. Hyperscaler cash flow covers a large portion, yet a sizable external financing gap remains. Private credit is expected to fill a meaningful share of that gap.
Public markets have shown clear signs of fatigue. Demand for hyperscaler bond deals has dropped sharply over recent months. Coverage that once ran near five times has fallen below two times in some cases. Money is moving toward private credit funds and off-balance-sheet structures. That shift is not accidental. It reflects both the size of the capital need and the preference of some companies to keep certain obligations away from their main balance sheets.
One high-profile example involves a major campus development in Louisiana. A private credit manager holds the majority equity stake in the joint venture while the technology company holds a minority share. Billions of dollars in bonds were issued against the project. The technology company leases the facility on an initial four-year term with renewal options. In its own disclosures, the company reports tens of billions in maximum exposure through residual value guarantees on similar unconsolidated ventures. The debt sits outside the company’s primary balance sheet, yet the economic risk remains closely tied to its willingness and ability to keep paying.
This structure is not unique. Technology firms have shifted well over one hundred billion dollars of AI-related data center spending into special purpose vehicles. Bondholders believe they are lending against solid infrastructure. In practice, the credit quality often depends on the ongoing lease payments of a small number of powerful tenants.
The Short Useful Life Of The Core Collateral
Buildings can last decades. The processors inside them do not. Rental rates for leading-edge GPUs have fallen dramatically since the early days of the boom. What once commanded high hourly rates now trades at a fraction of those levels on comparable contracts. Companies that own large fleets have revised their depreciation assumptions more than once. Some started with four-year lives and later stretched them. Others use six years. When the owners themselves keep changing their view of useful life, lenders face genuine uncertainty about residual value.
Credit markets have already begun to reflect this reality. Early investment-grade facilities for GPU-backed loans priced at relatively tight spreads. Newer facilities for similar assets have come at much wider levels, sometimes double or more. Some of the latest deals are backed by customers that carry speculative-grade ratings. The GPU itself matters less than the strength of the tenant’s promise to keep writing checks. One major cloud infrastructure provider reported tens of billions in principal outstanding, with a substantial portion tied to GPU-backed delayed-draw facilities. Pricing on recent paper has moved into territory more commonly associated with higher-risk credit.
I’ve found that markets often price the asset first and the tenant second when the asset feels tangible and high-tech. In this case, the tenant is the real story. A warehouse full of last-generation processors has limited alternative use once the original lessee walks away.
Circular Guarantees And Concentrated Exposure
The structures holding these deals together rely on residual value guarantees. One major chip supplier has extended guarantees that can reach well over one hundred billion dollars across various commitments. Those guarantees become payable to lessors if certain tenants default. The same tenants often agree to reimburse the chip supplier for any payout. In other words, the protection only holds if the very parties who might struggle are still able to perform. That circularity is elegant on paper and fragile in stress.
What appears to be diversified asset-backed credit is, in practice, highly concentrated exposure. A handful of tenants and one dominant product cycle drive the economics. The packaging makes it look like infrastructure. The underlying cash flows look more like a leveraged bet on continued aggressive capital spending by a few large technology companies.
Early Signs Of Strain In The Market
None of the current stress amounts to outright default. Yet the market is already repricing risk. One large project saw a force majeure notice related to power delays. If accepted, the tenant could defer rent for an extended period. Trading levels on the related financing debt have weakened noticeably. Bank groups that arranged large construction facilities have struggled to distribute the paper as quickly as hoped.
Issuance patterns have shifted as well. A growing share of data center bonds has moved into the below-investment-grade category. Volume in that segment has risen sharply compared with the prior year. Several planned public offerings by infrastructure providers were delayed or shelved within a short window. Credit default swap spreads for some major technology companies have widened to multi-year highs. Behind these moves sits a large pipeline of lease commitments that have not yet commenced. The absolute numbers run into the trillions of dollars on an undiscounted basis.
In my view, these are not isolated events. They are early signals that the market is beginning to distinguish between the shiny narrative of AI infrastructure and the harder reality of short-lived collateral and concentrated tenant risk.
Why Washington May Ultimately Hold The Risk
Nothing in the public record currently shows a formal plan for a federal backstop of data center debt. This remains a prediction based on incentives and precedent. The private residual value guarantees already in place create a template. Once those guarantees start getting tested, pressure for broader support can build quickly.
Policy has already begun treating large-scale compute as strategic. Federal sites have been opened for data center development. Official statements link data center infrastructure to economic and national security goals. Regulatory discussions around large loads and critical infrastructure continue. Once an asset class is framed as essential to national security, allowing facilities to go dark stops looking like a pure credit event and starts looking like a policy failure.
Industry voices have floated the idea of government support in the past. Comments about backstops that could lower financing costs and raise loan-to-value ratios appeared briefly before being walked back. Formal requests have focused more on grants, loans, and tax credits for related industrial capacity rather than direct guarantees on data center debt. Officials have publicly stated there will be no bailout. Those statements came before financing conditions tightened. History suggests the conversation can change when stress becomes visible.
The risk is already present inside the regulated system. Bank commitments to private credit funds run into the hundreds of billions of dollars. Insurers play a major role in the private placement market. Projections suggest data center securitization could grow into a meaningful share of overall asset-backed and commercial mortgage-backed issuance. That is the profile of an asset class that can become systemic.
The mortgage giants of the previous era funded themselves on the assumption that the government stood behind them. When the crisis arrived, that assumption became explicit. Similar dynamics can develop here. Paper that currently prices off corporate residual guarantees could, under stress, begin to price off an implicit sovereign backstop. It will not be labeled a bailout. It may arrive as a continuity facility, an expanded loan program, or a conservatorship justified on security grounds.
Who Is The Real Mark In This Trade
A well-known technology executive recently used a poker analogy when discussing AI liability. The person who thinks someone else is the mark often turns out to be the mark. That line applies cleanly to the credit market around data centers.
Many allocators believe they purchased senior, diversified, asset-backed exposure. What they actually own is correlated risk tied to a single product roadmap and the capital expenditure budgets of a few large tenants. The equity holders in the originating vehicles have already seen significant share price pressure. The taxpayer sits at the end of the chain, often without having been asked.
Perhaps the most interesting aspect is how quietly this concentration has built. The language of infrastructure and national priority makes the risk feel more diffuse than it is. In reality, the cash flows depend on a narrow set of decisions by a small group of companies and one dominant supplier.
What To Watch Next
Several developments would clarify the direction of travel. The first hyperscaler that declines to renew at the end of an initial short lease term would send a strong signal. The first residual value guarantee that is actually called would matter even more. Resolution of existing force majeure situations will show whether tenants gain flexibility that weakens lender protections. Further sales of construction debt at discounts or significant risk transfer trades would indicate banks are reducing exposure. Any formal policy language that extends loan guarantees from semiconductor plants and power infrastructure into the data halls themselves would confirm the path toward public support.
Rating agencies have already flagged the possibility of adjusting debt metrics for expected lease cash outflows. Follow-through on that front would force greater transparency around the true leverage in these structures.
The logic that leads some technology leaders to conclude that certain liabilities are simply too large for private markets applies with equal force one layer down the stack. The lenders who financed the physical buildout may eventually find themselves on the same road, requesting support under a different name.
The Broader Implications For Investors And Policymakers
For investors, the lesson is straightforward yet easy to overlook. Asset-backed labels do not automatically create diversification when the underlying economics depend on a concentrated set of counterparties and a rapid technology cycle. Understanding the residual value assumptions and the strength of the ultimate obligors is more important than the glossy description of the collateral.
For policymakers, the challenge is timing. Waiting until facilities are stranded creates pressure for emergency measures. Addressing the concentration and the short collateral lives earlier allows for cleaner risk-sharing arrangements. Treating compute capacity as strategic infrastructure without also considering the financing structures that support it leaves an incomplete picture.
I do not claim certainty about the exact path. Markets can remain liquid longer than many expect, and technology cycles can extend useful lives through software optimization or secondary markets. Still, the combination of short-lived assets, concentrated tenants, large off-balance-sheet commitments, and growing strategic importance creates a setup that deserves closer scrutiny than it currently receives.
The last time private gains were paired with an eventual public backstop on this scale, the assets were houses. This time the assets are processors and the buildings that house them. The underlying dynamic, however, looks remarkably similar. Private credit is financing a boom against collateral with a short clock. When the first large tenant walks, the question of who owns the stranded capacity will no longer be theoretical. I expect the answer will ultimately involve Washington, whether anyone wants to admit it today or not.
The AI buildout continues at remarkable speed. The financing structures supporting it have grown complex and opaque in places. Clarity about where the residual risk truly sits will matter more with each passing quarter. For now, the market still prices much of this exposure as infrastructure. The test will come when the first major renewal decisions arrive and the residual value assumptions face real stress. Until then, the quiet accumulation of risk continues, one GPU-backed facility at a time.
