Colocation providers moving into bare metal and GPUaaS (GPU-as-a-Service) are taking on a risk their business model has never carried: technology obsolescence risk.

The AI buildout has shifted where costs sit. A decade ago, the expensive part of a data center was the building, the power, and the cooling. As of 2026, the hardware inside the racks costs more than the facility around it, and it loses value far faster.

Why colos are moving up the stack

The demand for compute is overwhelming. Everybody is looking for an energized rack of GPU servers today, not in five months.

The progression runs from powered shell, to colocation, to bare metal, to GPUaaS. Each hop carries both higher margin and higher operational risk than the one before it. Bare metal, the most popular way to buy enterprise compute, means renting out whole servers with nothing between the customer's workload and the hardware. GPUaaS adds orchestration, scheduling, and support on top.

Colos already own the hardest parts. Power allocations and interconnect are the biggest bottlenecks in AI infrastructure right now, and the colo has more access to both than most. From there, buying servers and renting them out looks like a small step for a large multiple on revenue per megawatt.

Enterprises and AI startups want GPU capacity without hyperscaler lock-in and without building their own clusters, and a colo with available power can stand up an offering faster than a new entrant can find a site.

GPUs became a financeable asset class

GPUs have become their own asset class, with financing pools separate from those for data centers. GPUs are bought with debt, pledged as collateral, securitized, and traded on a secondary market. CoreWeave raised $7.5 billion in debt from Blackstone and Magnetar in May 2024, secured largely by its GPU fleet and customer contracts. Lenders across private credit have followed with similar structures ever since.

The mechanics of a typical deal look like this. An SPV, a special purpose vehicle, buys the GPUs to separate the deal from the operator’s balance sheet. Debt is sized against the customer contracts the fleet is committed to, plus the assumed residual at maturity. Equity sits underneath and targets a levered IRR, the annualized return after debt.

Every one of those structures rests on a residual value assumption. Residual value is what the hardware is worth when the financing matures, typically three to six years out. The lender sizes the loan against contracted revenue plus that residual. The operator prices its leases assuming it can resell or redeploy the hardware at that residual. The equity investors need the residual to hit their ambitious return targets.

The overbooked residual

Residuals today look good. A three-year-old H100 GPU can still be resold for 60 percent of its original price. A large portion of investors in AI infrastructure believe high residuals will continue to hold.

But GPU residuals have almost nothing to do with physical wear. They will not follow a simple depreciation curve. They are driven by Nvidia's release cadence and pricing. If new competition to Nvidia appears, Nvidia can easily give up some of its 75 percent gross margin to win back market share. A price reduction on new GPUs would crush used GPU prices and push residuals far below projections.

The overbooked residual is a huge problem: the same optimistic residual gets counted three times. It is baked into the lease pricing, baked into the debt sizing, and baked into the equity return model. One asset's value underpins three different parties, none of whom can survive if the residual actually clears lower.

The data center industry has run this experiment before. In the late 1990s, carriers borrowed tens of billions to lay fiber on the thesis that internet traffic would double every hundred days. The capacity arrived faster than the demand, wholesale bandwidth prices collapsed, and the debt could not be serviced from the cash flows the fiber actually produced. Global Crossing filed for bankruptcy in January 2002 with over $12 billion in debt. WorldCom followed six months later in what was then the largest bankruptcy in US history.

The fiber itself was fine. It carried traffic for decades afterward, at prices that only worked for buyers who picked it up for pennies on the dollar. The thesis was directionally correct, and it was still fatal for many of the entrepreneurs who bet on it.

What discipline looks like

Competition for deals is pushing underwriting toward IRR targets that only pencil if three things hold at once: high utilization, stable GPU-hour pricing, and a strong residual. Those three are correlated. In a downturn, they soften together, because the same surplus of capacity that cuts utilization also cuts pricing and floods the secondary market.

The operators handling this well underwrite the hardware like technology and the building like real estate, and they refuse to let one set of assumptions bleed into the other.

Bare metal is attractive, and it makes sense why colos and ex-crypto miners are moving into it. The customer demand is very real. But the question still stands: which operators will survive the first residual repricing, and which won't?