The New GPU Asset Class

TL;DR

  • The Rise of Bare Metal: Bare metal is now the dominant model for renting GPU clusters, sitting directly between traditional colocation and public cloud by allowing customers to rent entire, provider-owned servers without an intervening virtualization layer.
  • A Distinct New Asset Class: GPUs have separated from data center real estate to become a standalone asset class because they possess all four necessary traits: identifiable assets, observable pricing, active resale markets, and lenders willing to accept them as collateral.
  • The Portability Advantage: Unlike traditional real estate, GPU servers are portable and serial-numbered, meaning lenders can physically repossess and re-market the hardware in the event of a loan default.

# # #

Bare metal has become the dominant way GPU clusters are rented. The customer’s software runs directly on the hardware, with no hypervisor, no virtualization layer, and no managed services in between.

Since 2023, it has pulled the server into a separate asset class of its own from Data Centers. GPU clusters are financed in the capital markets, where it is now priced, traded, and borrowed against.

Where bare metal sits between colocation and cloud

Bare metal sits between the two models the industry already knows. Colocation has the tenant owning the servers and renting space, power, and cooling. Public cloud has the provider owning everything and selling virtual slices with managed services on top.

Bare metal is in the middle. The provider owns the servers, and the customer rents them whole, by the hour or on multi-year contracts. Neoclouds such as CoreWeave, Lambda, and Crusoe, cloud providers built specifically around GPU capacity, run their entire business on this model.

And now, traditional colocation operators are looking to do it too. They’re adding bare metal and GPU-as-a-service lines to capture the same demand, taking on hardware ownership their business model never required before.

The hardware became an asset class

An asset class needs four things: identifiable assets, observable pricing, a resale market, and lenders willing to take the asset as collateral. GPU servers now have all four.

The GPU can leave the building. When a data center loan defaults, the lender ends up owning real estate in one location. But when a GPU loan defaults, the lender holds serial-numbered servers that can be repossessed and re-marketed. That portability is what allows the hardware to be financed separately from the building it sits in.

In March 2026, CoreWeave closed an $8.5 billion facility rated A3 by Moody’s, the first investment-grade rating ever assigned to GPU-backed debt, supported in part by Meta contracts worth at least $19 billion.

The pricing of that debt tells the story faster than the headlines. CoreWeave’s earliest GPU-backed borrowing reportedly priced in the low teens. The floating-rate portion of the March 2026 facility priced at 2.25 percentage points over SOFR, the overnight benchmark rate. Lenders re-rated the asset from exotic to ordinary in under three years.

The data center industry spent three decades learning to finance buildings. Now, racks inside carry capital markets of their own.

# # #

About the Author

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

The post The New GPU Asset Class appeared first on Data Center POST.

TL;DR The Rise of Bare Metal: Bare metal is now the dominant model for renting GPU clusters, sitting directly between traditional colocation and public cloud by allowing customers to rent entire, provider-owned servers without an intervening virtualization layer. A Distinct New Asset Class: GPUs have separated from data center real estate to become a standalone
The post The New GPU Asset Class appeared first on Data Center POST. Read More Data Center POST

Tags:

TL;DR

  • The Rise of Bare Metal: Bare metal is now the dominant model for renting GPU clusters, sitting directly between traditional colocation and public cloud by allowing customers to rent entire, provider-owned servers without an intervening virtualization layer.
  • A Distinct New Asset Class: GPUs have separated from data center real estate to become a standalone asset class because they possess all four necessary traits: identifiable assets, observable pricing, active resale markets, and lenders willing to accept them as collateral.
  • The Portability Advantage: Unlike traditional real estate, GPU servers are portable and serial-numbered, meaning lenders can physically repossess and re-market the hardware in the event of a loan default.

# # #

Bare metal has become the dominant way GPU clusters are rented. The customer’s software runs directly on the hardware, with no hypervisor, no virtualization layer, and no managed services in between.

Since 2023, it has pulled the server into a separate asset class of its own from Data Centers. GPU clusters are financed in the capital markets, where it is now priced, traded, and borrowed against.

Where bare metal sits between colocation and cloud

Bare metal sits between the two models the industry already knows. Colocation has the tenant owning the servers and renting space, power, and cooling. Public cloud has the provider owning everything and selling virtual slices with managed services on top.

Bare metal is in the middle. The provider owns the servers, and the customer rents them whole, by the hour or on multi-year contracts. Neoclouds such as CoreWeave, Lambda, and Crusoe, cloud providers built specifically around GPU capacity, run their entire business on this model.

And now, traditional colocation operators are looking to do it too. They’re adding bare metal and GPU-as-a-service lines to capture the same demand, taking on hardware ownership their business model never required before.

The hardware became an asset class

An asset class needs four things: identifiable assets, observable pricing, a resale market, and lenders willing to take the asset as collateral. GPU servers now have all four.

The GPU can leave the building. When a data center loan defaults, the lender ends up owning real estate in one location. But when a GPU loan defaults, the lender holds serial-numbered servers that can be repossessed and re-marketed. That portability is what allows the hardware to be financed separately from the building it sits in.

In March 2026, CoreWeave closed an $8.5 billion facility rated A3 by Moody’s, the first investment-grade rating ever assigned to GPU-backed debt, supported in part by Meta contracts worth at least $19 billion.

The pricing of that debt tells the story faster than the headlines. CoreWeave’s earliest GPU-backed borrowing reportedly priced in the low teens. The floating-rate portion of the March 2026 facility priced at 2.25 percentage points over SOFR, the overnight benchmark rate. Lenders re-rated the asset from exotic to ordinary in under three years.

The data center industry spent three decades learning to finance buildings. Now, racks inside carry capital markets of their own.

# # #

About the Author

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