Why colos are moving up the stack into bare metal and GPUaaS

Power is the binding constraint on new GPU deployments. Operators who already hold energized space are looking to convert compute demand into revenue themselves. Colocation facilities and former bitcoin miners are moving up the stack into bare metal and GPU-as-a-Service. And colocation facilities can finance the climb on better terms than the new neoclouds they compete with.

Energized space is the scarce asset

North America’s data center vacancy ended 2025 at 1.4%, a record low, with three quarters of capacity under construction already preleased. New supply is gated by the grid. Transformers take two and a half years, high-voltage substations three to five, and only 13% of what entered US interconnection queues between 2000 and 2019 ever got built.

“Time to power” is a defining success metric for many GPU operators. This makes it natural for colocation facilities to consider providing bare metal themselves. If you have space, it seems feasible to buy and rack GPU servers, and rent them out whole.

The catch is CapEx. GPU servers are more expensive than the infrastructure that house them. IREN needs to buy $5.8 billion of hardware from Dell to fill its 200 megawatts. The payback on typical GPU projects is around 24 to 30 months. If the AI bubble pops at 30 months, or rental rates drop, you might never turn a profit on the equipment.

Operators are already choosing sides

Operators are choosing between staying as landlords, or competing with their tenants. It’s uncomfortable for the neoclouds: their landlords are becoming competitors with more control over the biggest bottleneck, energized space.

Lenders split an AI data center into two assets on different clocks. The shell gets real estate debt over 15 to 20 years. The chips get their own loans over three to four years.

GPU-backed debt is expensive. CoreWeave’s first facility was priced around 15% in 2023, its second near 11%, and no GPU loan reached investment grade until March 2026. Data center real estate has been investment-grade collateral since 2018 and produced $25 billion of securitizations in 2025.

A colo that climbs the stack needs to borrow against both pools. Luckily for colos, years of audited rent are exactly the credit file that equipment financing companies want to see. Compared to another “wannabe neocloud”, a credible colo brings years of operating history that impresses lenders, plus real estate collateral that improves financing terms.

The move up the stack is risky

The neocloud business is not for the faint of heart. While used H100s still fetch 60% of their original price today, that number will fall as new generations ship, and whoever owns the chips owns that risk. Colos can avoid some of the AI bubble effects, but once they do bare metal, they’re fully exposed to a future AI market crash. A colo that buys GPUs on spec owns depreciating silicon, at an expensive interest rate.

As of mid-2026, compute demand is insatiable. And a new megawatt takes years. Operators who already control the megawatts can easily enter bare metal, but at what cost?

Author

Bernie Margulies is the founder and CEO of American Compute (amcompute.com), which helps GPU buyers get access to better GPU financing, and structures residual value insurance for lenders/lessors.

If you haven't already, please take our Reader Survey! Just 3 questions to help us better understand who is reading Telecom Ramblings so we can serve you better!