I’ve been watching the AI buildout for a while now, and every few months something new lands that forces everyone to pause. This week it was the financing arrangements around Nvidia’s computing capacity. Big names in finance stepped forward and framed those GPUs and data-center assets as something close to a new investable class. Larry Fink even reached back to the early days of mortgage-backed securities for a comparison. The language was confident. The market reaction was more cautious. Nvidia shares dipped right after the news broke, then recovered a bit, but the conversation shifted quickly from celebration to hard questions about valuation.
That’s the part that stays with me. Turning computing power into collateral sounds neat on a slide deck. In practice it runs into a wall of unknowns. Auto loans and residential mortgages have decades of data behind them. Data-center loans backed by GPUs do not. The hardware burns through its useful life in just a handful of years. New capacity keeps arriving, some of it from places that play by different cost structures. And the financing itself already carries the faint odor of circular arrangements that make investors nervous.
Why Valuing These Assets Feels So Different
Most of us who look at credit markets for a living are used to assets that age in predictable ways. A car loses value, yes, but the curves are well mapped. Houses sit on land that tends to hold or gain value over long stretches. Data-center hardware is a different animal. The chips themselves can become less competitive in three to five years, sometimes faster when a new architecture lands. Operators talk about five- or six-year depreciation schedules as if that is conservative. In my experience that range already feels optimistic once you factor in the speed of the next product cycle.
Dan Alpert, who has spent years looking at complex credit structures, put it plainly: the metrics for underwriting data-center-backed loans are still hard to establish because the experience is so short. He is right. We simply do not have a long enough track record of how these assets behave through a full economic cycle, let alone through a period when supply of computing capacity suddenly jumps.
Wells Fargo traders noted that the involvement of heavyweight private-equity and asset-management firms acts almost like a form of insurance for investors who remain less comfortable with GPU collateral. That comfort is still thin. One open question they raised is whether these AI-factory loans eventually get sliced up and sold the way auto loans or commercial real-estate debt already are. If that happens, the market will need clearer rules for residual value, replacement cost, and technological obsolescence. Right now those rules feel more like educated guesses than settled practice.
The Speed of Depreciation Changes Everything
Paul Meeks, who follows technology closely, reminded everyone that these GPUs depreciate on a five- or six-year schedule. Even inside the tech world that is an emerging market where the final scorecard arrives years later. I keep coming back to that point. If you underwrite a ten-year loan against hardware that may be largely obsolete in year six, the recovery value in a stress scenario starts looking thin. Secondary markets for used high-end GPUs exist, but they are not deep or liquid in the way the used-car market is. Pricing can gap lower quickly when a new generation arrives.
There is also the simple physical reality of data centers themselves. Power, cooling, and networking gear age too. Some of that infrastructure lasts longer than the chips, but the whole facility is only as valuable as the revenue it can generate. If utilization drops or pricing for compute softens, the cash-flow coverage of the loans comes under pressure. That is not a theoretical worry. It is the same dynamic that hit earlier technology infrastructure booms.
Echoes of Past Infrastructure Races
Alpert reached for the fiber-optic cable example from the late 1990s and early 2000s. Global Crossing and others tried to turn vast networks of dark fiber into investable assets. The capacity was real. The demand forecasts were optimistic. When the cycle turned, a lot of that fiber sat unused and the debt structures behind it collapsed. The comparison is imperfect, of course. Today’s AI demand is more immediate and the cash flows from cloud and training workloads are already visible. Still, the structural risk feels familiar: build capacity first, assume utilization later, and hope the financing markets stay open.
Michael Burry went further and saw shades of Enron’s effort to make wholesale power an investable class. He pointed to the structuring of unnatural credits late in a bull phase as the place where worry sets in. I do not think the current arrangements are as extreme, yet the parallel is worth keeping in mind. When momentum is strong, the temptation grows to stretch definitions of collateral and to keep the financing machine running. That is usually when the fine print starts to matter most.
Circular Financing and the Optics Problem
The announcement itself immediately triggered talk of circular financing. Nvidia helps finance or structure deals that ultimately support demand for its own chips. The big asset managers and private-equity firms gain exposure to what they hope will be a growing asset class. On paper everyone can point to independent credit analysis and arm’s-length terms. In practice the optics are sticky. Investors who already feel the AI trade is crowded see another layer of interdependence and wonder how clean the risk transfer really is.
I’ve found that markets will often tolerate a degree of circularity when growth is strong and cash flows are rising. They become far less forgiving when growth slows or when residual values start missing forecasts. That is the scenario that keeps credit desks up at night. If GPU prices in the secondary market soften faster than modeled, or if a wave of new capacity from other regions undercuts utilization, the collateral that looked solid at origination can look thinner in a hurry.
What Underwriters Still Struggle to Quantify
Several practical questions keep coming up in conversations with people who actually price these risks.
- How should residual value be modeled when the next architecture can cut performance-per-watt by half in a single generation?
- What recovery assumptions make sense for specialized data-center real estate that may need expensive retrofits to stay competitive?
- How much weight should be given to offtake contracts that themselves depend on the continued growth of AI training and inference budgets?
- Where does geographic concentration risk sit when a large share of capacity sits in a handful of power-constrained regions?
None of these have clean answers yet. The short history of the asset class means models lean heavily on assumptions about technology curves and demand elasticity. Those assumptions can look reasonable in a rising market and fragile when conditions change.
International Capacity Adds Another Variable
Investors are also watching the buildout outside the United States. China in particular continues to add significant computing capacity even under export controls. The hardware may not always match the latest generation, but the volume is large and the cost structure is different. If that capacity finds its way into global cloud or AI service markets at lower prices, utilization and pricing for Western data-center operators could feel pressure. That is another input that traditional ABS models have little experience pricing.
Perhaps the most interesting aspect is how quickly the conversation has moved from pure enthusiasm to risk management. A year ago the dominant story was simply the scale of the opportunity. Now the same investors who want exposure are asking sharper questions about collateral quality, depreciation, and structural circularity. That shift feels healthy. It does not mean the financing will stop. It does mean the terms and the underwriting standards will face more scrutiny.
Could These Loans Become a New ABS Category?
The Wells Fargo note floated the possibility that AI-factory loans eventually get repackaged into something resembling the collateralized loan market. That would require a deeper secondary market, more standardized documentation, and a clearer consensus on residual values. Right now those pieces are still forming. Some managers are comfortable taking the risk on their balance sheets or in closed-end vehicles. Broader public ABS markets tend to demand more history and more transparency before they open the gates fully.
In my view the path toward a true asset-backed market for computing capacity will be longer and bumpier than the optimistic comparisons suggest. Mortgage-backed securities grew out of decades of housing data and government involvement. Auto ABS benefited from deep secondary markets and relatively standardized vehicles. Computing hardware sits closer to the technology equipment leasing world, where residual-value risk has always been a central concern. That history should temper expectations.
What Investors Should Watch Next
Several practical signals will matter more than the next press release. First, the actual performance of early data-center loan portfolios: delinquency rates, recovery values on any early terminations, and how residual values track against original appraisals. Second, the pace of new capacity additions, both domestically and abroad, and any signs that pricing power for compute is softening. Third, the willingness of the large financing partners to keep expanding their exposure if secondary-market prices for used GPUs start to slide.
I also pay attention to how the equity market treats the pure-play data-center operators and the chipmakers themselves. When share prices start to price in slower growth or higher capital intensity, the credit markets usually follow with tighter terms. The reverse is also true: strong equity performance can paper over a lot of residual-value uncertainty for a while.
None of this means the AI infrastructure buildout is somehow illegitimate. Demand for training and inference capacity is real and large. Power constraints and chip supply still limit how fast the industry can grow. The financing is simply running ahead of the historical data in some important places. That gap is what investors are probing right now.
Balancing Opportunity Against Structural Risk
The comparison to the mortgage market of the 1970s is seductive. It suggests a long runway of financial innovation and a new asset class that will eventually become ordinary. Maybe that is the path. Or maybe the closer analogy remains the earlier technology infrastructure cycles that produced both genuine capacity and painful write-downs. The difference will come down to how carefully the residual-value risk is priced and how transparent the structures remain when the cycle inevitably slows.
For now the big firms have put their names and capital behind the idea that computing capacity can sit alongside more traditional collateral. That vote of confidence carries weight. It does not answer the harder questions about depreciation schedules, secondary-market liquidity, or the impact of sudden new supply. Those answers will arrive the old-fashioned way: through the actual performance of the loans over the next several years.
I’ve found that the most useful stance is neither full embrace nor outright dismissal. Treat the financing as a live experiment in credit markets. Watch the early data carefully. Stay alert to any signs that residual values or utilization rates are diverging from the original models. And remember that technology assets have a long history of aging faster than the debt attached to them. That history is worth keeping in the foreground even while the current opportunity looks large.
The next phase of this story will not be written in press releases. It will show up in the fine print of loan documents, in the recovery rates on any early exits, and in the willingness of a broader set of investors to accept GPU collateral without the comfort of the largest names standing nearby. Until those signals become clearer, a healthy dose of skepticism about precise valuations remains the rational position.
Markets have a way of testing new asset classes under stress. The current cycle has not yet delivered that test. When it does, the difference between marketing language and underwriting reality will become much sharper. That is the moment when the true investable quality of data-center loans will be measured, not in the optimistic forecasts of today, but in the cold numbers of recovery and residual value that arrive later.
Until then, the questions investors are asking feel both timely and necessary. The hardware is impressive. The demand is real. The financing structures still need to prove they can survive a full cycle with the kind of predictability that more traditional asset-backed markets take for granted. That proof will take time, and the early chapters are only now being written.
