Rent per rack or revenue per GPU hour? Here is the math for colo owners: three models compared with sourced assumptions, plus the case for staying a landlord.
Rent per rack or revenue per GPU hour? That is the choice facing colocation owners as AI demand pulls GPUs into their halls, and it decides how much gpu colocation revenue the move from colo to gpu cloud actually adds. This guide compares three models for the same megawatt: stay a landlord on a fixed lease, take a share of a partner's compute revenue, or buy GPUs and sell compute yourself. The inputs are public and dated, every assumption is shown, the example rates are arithmetic, not market data, a quote or a forecast, and the outputs compare business models under the same assumptions, so they are not profit forecasts. One disclosure: packet.ai runs a provider program, so we have an interest in colo owners listing capacity, and this guide includes the honest case for staying a landlord.
Key takeaways
Start with what rent looks like. Published 2026 colocation quotes run roughly $130 to $470 per kW per month depending on market, density and contract size, and CBRE's H2 2025 US wholesale average was about $196 per kW for 250 to 500 kW deployments, as reported. Quotes differ in what they include, since some bundle power, cooling and a network port while others bill metered electricity separately, so compare like with like. The large public AI leases are the wholesale data center model, where a tenant takes a block of megawatts on a long lease and the owner of a wholesale colocation data center delivers the power and cooling, and they land inside that range. TeraWulf's August 2025 Lake Mariner agreements with Fluidstack were worth about $3.7 billion over 10-year initial terms for more than 200 MW, which is at most about $154 per kW per month spread over 200 MW. The later 168 MW Abernathy joint venture, a 25-year lease worth about $9.5 billion, works out to about $188. Those two figures are our arithmetic from company disclosures and ignore escalators and other lease terms, and press coverage describes the Fluidstack agreements as HPC colocation deals. Most of these tenants are neoclouds, so a neocloud data center is in practice a landlord's building filled with a tenant's GPUs.
The margin is the other half. TeraWulf's investor presentation projected about an 85% net operating income margin on the Lake Mariner leases, roughly $315 million a year at the site, on modified gross leases with annual escalators, and described Google's backstop of Fluidstack's lease obligations as supporting project debt. Cipher, by one summary, signed a 15-year lease with Amazon Data Services for 216 MW worth about $5.5 billion, guaranteed by Amazon. In each case the owner delivers shell, power and cooling and the tenant brings the silicon and the credit.
The landlord figure is a projected site NOI margin, which is generally measured before depreciation and amortization. The compute model later in this guide deducts GPU depreciation and modeled operations but excludes several other costs. They are different measures and should not be read as equivalent profit margins.
Selling compute is a different calculation. Revenue is billable GPU-hours times the rate you realize, the costs include GPU depreciation, and utilization decides how many hours pay. The packet.ai guides to what an idle GPU costs and GPU utilization explained set out that math.
Our view
Do not compare the two headline margins directly. An 85% margin on a lease and a margin on GPU rental measure different things: the first sits on facility project costs of roughly $8 million to $12 million per megawatt (benchmarks cited by TeraWulf and CleanSpark, whose scope varies by site) with a creditworthy tenant behind it, and the second sits on an illustrative GPU hardware cost of about $30 million per modeled megawatt of IT load, with utilization risk. The right question is how much extra capital you are putting at risk for how much extra margin.
Data center GPU monetization takes three forms: lease the space, share the revenue, or sell the compute. It is colocation vs cloud in miniature. To sell compute from colocation space, you add five things. Capital: you buy the GPUs. IREN, which owns the chips and sells compute, signed a five-year contract with Microsoft worth about $9.7 billion in November 2025 alongside a separate $5.8 billion equipment purchase from Dell, so it carries the hardware cost as well as the building. Operations: scheduling, high-speed networking such as InfiniBand and monitoring become your job. Customers: sales, support and billing become yours too. Risk: you take on depreciation and the utilization gap between allocated and billable hours. Trust: buyers will ask about isolation, security and certification.
There are middle paths that avoid most of the capital. You can host a partner's GPUs on a revenue share, list capacity through a provider program or marketplace, or use a white-label platform to run a cloud under your own brand. Vast.ai's datacenter tier, for example, is aimed at hosts with five or more GPU servers in a professionally managed facility, and RunPod says its Secure Cloud draws on vetted data center partners. The packet.ai guide to the five ways to turn idle GPUs into revenue compares those routes, and the post on why neoclouds are adding a second demand channel covers the market logic.
A data center for AI is screened by GPU buyers and partner programs on a short list of things. The table summarizes what public sources say, and it is not packet.ai's requirement list.
High density colocation is the term operators such as Core Scientific use for halls built to carry GPU racks. One caution on tiers. The availability percentages often quoted with them, 99.982% for Tier III and 99.995% for Tier IV, are legacy figures: Uptime Institute removed expected-downtime numbers in 2009, and a tier describes design and resilience, not a guaranteed uptime. Treat those percentages as history, not as an SLA.
Certification is how a buyer who cannot visit your hall decides whether to trust it. Uptime Institute issues tier certifications in stages: design documents, the constructed facility, and operational sustainability. A facility advertised as "Tier III design certified" has certified paperwork, not necessarily a certified building, and the constructed-facility stage is the stronger claim. SOC 2 Type II is an audit of a provider's controls over a period by an independent firm, and ISO 27001 certifies an information security management system. Neither is a substitute for the other, and neither says anything about power density.
Partner programs already ask. RunPod's documentation says Secure Cloud uses vetted infrastructure partners meeting standards such as SOC 2, ISO 27001 and PCI DSS. Vast.ai's datacenter documentation lists an active third-party certificate such as ISO/IEC 27001, a registered business, a signed hosting agreement, identity verification and at least five GPU servers, while its application page describes certifications as encouraged but not strictly required, so check the current wording before you apply.
The model puts one megawatt of IT load through three options. The inputs reuse the cost model from what an idle GPU costs. All figures are illustrative, exclude facility operating costs, networking, financing, sales and support, and use hypothetical rates and shares that are not market data, packet.ai terms or a forecast.
How the options are defined. In option A the owner receives fixed rent per kW and the tenant brings and finances the GPUs. In option B a partner owns and operates the GPUs and pays the owner a percentage of gross compute revenue, which is billable GPU-hours times the realized rate; the partner bears GPU depreciation, operations and platform costs, and the owner's facility power and cooling costs are outside this model. In option C the owner buys the GPUs and keeps the revenue, bearing the GPU depreciation and operations shown, with facility costs, networking, financing, sales and support outside the model. So A and B are modeled revenue to the owner, C is modeled contribution after GPU depreciation and operations, and none of them is net profit or a fully loaded comparison.
Option A: stay a landlord. One megawatt of fixed rent is $130,000, $200,000 or $470,000 a month at the three rent levels, with no GPU capital at risk.
Options B and C at three utilization levels (revenue is billable GPU-hours times $2.00; B takes a hypothetical 20% or 30% of it; C keeps it and pays about $594,000 a month of GPU depreciation and operations):
Compare against $200,000 of rent revenue. At 60% billable and $2.00, running GPUs leaves a modeled contribution of about $184,000, slightly less than the rent, for $31.7 million of extra capital. At 40% the contribution is negative. A 30% revenue share beats $200,000 from 60% billable upward, while a 20% share needs 80%. Only at 80% does running GPUs clearly come out ahead: a modeled contribution of about $444,000 a month, about $244,000 more than the rent, or about 9% a year on the $31.7 million of GPUs before financing, sales, support and networking. These are scenario outputs under stated assumptions, not forecasts of what a colo owner will earn, and the comparison is not like for like: rent is revenue before the landlord's facility costs, while C is a contribution after GPU depreciation and operations only.
The break-even rate for the GPU side alone, before any facility cost, is about $2.29 per billable GPU-hour at 40% billable, $1.53 at 60% and $1.15 at 80%. It is the same fixed-cost break-even logic as in GPU utilization explained.
Duos Technologies is a dated example, and a more nuanced one than a simple exit. On October 5, 2026 it announced that it had completed the sale of its GPU-as-a-service entity to Axe Compute, with the transaction closing on September 30. Axe took over the GPU cluster and satisfied the roughly $98.1 million equipment financing, Duos keeps its Columbus customer under a revised five-year agreement, and Duos now leases the underlying GPU capacity from Axe instead of owning and financing the equipment. Duos describes the move as sharpening it into a pure-play colocation operator. Core Scientific, TeraWulf, Hut 8 and Cipher have chosen the landlord version at larger scale, delivering shell, power and cooling on long leases while tenants bring or finance the silicon.
Our view
The decision is about risk-adjusted return on capital, not the highest possible revenue. Stay a landlord if you can sign a creditworthy anchor tenant, if you do not have tens of millions of dollars per megawatt for GPUs, or if you cannot staff sales, support and operations for a cloud. Consider running GPUs only if you have the capital, a credible plan for 80%-type billable utilization, and buyers lined up, because the example shows the extra capital earns its keep only at high utilization. The middle paths, a revenue share or a provider-program listing, let you test compute demand without buying the hardware. This is a framework, not a forecast, and your own rent, power cost and tenant mix will change the numbers.
See how the provider program works or apply to list capacity. For the terms used here, the GPU provider glossary defines committed power, colocation, SLA and tier.
Retrieved October 14, 2026. Status of figures: announced contract terms and filings (TeraWulf, Duos), management projections (TeraWulf's site NOI margin), market estimates (rent ranges, GPU prices, CBRE), secondary-source summaries (Cipher, IREN, CleanSpark), and our own illustrative assumptions (rates, shares, utilization, the operations allowance). Secondary summaries are noted as such.
Last reviewed: October 14, 2026. This is an illustrative model, not financial, tax or engineering advice, and it contains no market rates, no earnings figures and no guarantee of income. Company projections are the companies' own. For packet.ai's current provider requirements and terms, see the provider page.
Same models. Same API. Fraction of the cost. Start free — no credit card required.
Start Building →