Bare Metal: Understanding GPU Rental Economics

GPU rental economics depend on the price and duration of customer commitments, the capital needed to deliver capacity and the equipment's useful commercial life. This September 14 explainer uses a disclosed contract and clearly labeled arithmetic to explain those relationships; its sources were reviewed on September 27.

The key term is bare metal: renting actual, dedicated physical servers, rather than just a virtual machine running on a server. AWS itself offers bare metal through EC2, so bare metal can absolutely be a cloud service. AWS EC2 documentation

What bare metal actually means

Think of a server as a powerful computer sitting in a data center. In this context, it contains GPUs used to train or run AI models.

With bare metal, the customer gets direct access to the physical machine’s resources. Its operating system can run directly on the hardware, rather than inside a provider’s virtualization layer, which creates virtual computers on top of the physical machine. The customer is renting access to the equipment, not necessarily buying it. AWS EC2 documentation

A useful way to distinguish the offerings:

These are simplified commercial categories, not rigid technical boundaries. CoreWeave runs Kubernetes, software that coordinates workloads, directly on bare metal servers. Bare metal does not mean there is no software or management involved. CoreWeave bare metal

A bare metal agreement primarily grants access to dedicated infrastructure. The exact software and operational responsibilities still depend on the contract.

What are neoclouds?

Here, neoclouds means specialist cloud providers focused on GPU computing for AI. CoreWeave, Nebius and IREN illustrate businesses discussed in this context, with different asset and service models.

Their product is not simply electricity or data center space. It is computing capacity, potentially bundled with networking, storage and software. CoreWeave's bare-metal description illustrates how a provider can combine dedicated physical infrastructure with orchestration software. CoreWeave bare metal

What does $10–12/W mean?

Annual revenue per watt divides a computing contract's annualized value by the relevant IT-capacity denominator. It is not an electricity price. Always state the time period and distinguish IT power from total facility power.

IREN's November 3, 2025 announcement provides a dated example: its Microsoft contract had approximately $9.7 billion of value over five years and was tied to new facilities supporting 200 MW of critical IT load. The release expected phased GPU deployment through 2026 and included a 20% prepayment. That announcement describes contract terms and planned delivery, not proof that all capacity was already operating. IREN contract announcement

This is a simplified contract-value normalization. It is not a schedule of recognized revenue: capacity is delivered in phases, and actual accounting and cash receipts depend on contract terms.

The calculation is:

$9.7 billion ÷ 5 years ÷ 200 million watts = $9.70 per watt per year.

So, for an illustrative 100 MW deployment, assuming that capacity is fully contracted at the stated rate:

This is revenue from providing computing capacity, not profit, and not just revenue from supplying power.

The $10, $12, $15 and $20 values above are illustrative inputs, not verified market benchmarks or quotes. Comparing real contracts also requires GPU type, service scope, networking, utilization obligations, delivery dates, prepayments and termination terms. A rate summary without those details can create a misleading comparison.

Why would customers pay more, and why does GPU lifespan matter?

Higher pricing comes with tradeoffs. Earlier availability, newer equipment, shorter commitments and smaller prepayments can shift more risk to the provider. Whether a particular contract earns a premium has to be established from its terms.

The commercial logic is straightforward: a customer committing for five years and helping fund the equipment upfront gives the provider more certainty. A customer asking for capacity immediately, without a long commitment or much upfront cash, is asking the provider to take more financing and future rental risk. A higher headline rental rate does not automatically mean a better investment return.

Longer GPU life could improve economics, but physical survival is not the same as economic usefulness. A functioning GPU may become less attractive relative to new equipment. CoreWeave's 2025 Form 10-K identifies uncertainty in useful-life estimates and the ability to redeploy infrastructure after contracted use. CoreWeave 2025 Form 10-K, risk factors

There is also an accounting distinction. As a simplified illustration, depreciating $6 billion of equipment with no residual value over eight years instead of six lowers annual straight-line depreciation from $1 billion to $750 million. That changes reported expense, but changing the estimate does not itself generate cash. CoreWeave's 2025 Form 10-K states that its computing-equipment life estimate changed from five to six years effective January 1, 2023, reducing that year's expense by $20 million. This is a historical accounting estimate, not a guarantee of future rental income. CoreWeave property-and-equipment accounting policy

The genuine economic upside comes from customers continuing to pay profitable rental rates for those older GPUs, potentially delaying the need to replace them.

A GPU rental business earns its return through profitable customer use over the equipment's economic life. Evaluate the cash collected after equipment, financing, power, staffing, maintenance and replacement costs. Bare metal describes the dedicated physical infrastructure; it does not settle the investment return.

Sources and assumptions