On-premise AI vs cloud cost: where the crossover happens

Simon Harrsen - EVP North America Published October 15, 2026  ·  11 min read

Most enterprises running AI are not AI companies. They are manufacturers, banks, insurers, and health systems. Sooner or later, each one asks the same question: is it cheaper to run AI on our own hardware, or to rent it in the cloud?

The honest answer is that it depends. This article shows what it depends on, and how to run the numbers yourself. It is part of our guide to AI infrastructure financing.

One note on where we stand. CHG-MERIDIAN leases hardware. We do not sell cloud compute. That gives us a view of one input most cost comparisons skip: what AI servers are worth at the end of their use. We share our estimates below.

What is the cost crossover between on-premise AI and cloud?

The cost crossover is the point where the total cost of running AI on your own hardware falls below the cost of renting the same capacity in the cloud. Below the crossover, the cloud costs less. Above it, on-premise AI costs less.

Three inputs decide where the crossover falls. The first is utilization: how many hours your GPUs spend doing useful work. The second is the cloud rate you would actually pay. The third is end value: what your hardware is still worth when you stop using it.

Most comparisons treat the first two carefully. The third is usually set to zero.

What does on-premise AI cost, and what does cloud AI cost?

On-premise AI and cloud AI put costs in different places. Some costs you pay up front, some you pay every hour, and one works in your favor at the end.

On-premise costs

On-premise AI means you run the hardware in your own data center, or in colocation space, which is rented space in someone else's facility. The costs fall into two groups.

The first group is the hardware itself. That includes the GPU servers, plus the networking and storage that connect them.

The second group is the cost of running it. That includes power, cooling, space, maintenance, and the staff who operate the systems.

One item works in your favor: end value. When you finish with the hardware, it can be resold or returned. That value offsets part of what you paid.

Cloud costs

Cloud AI means you rent GPU capacity from a provider. The main cost is the hourly rate. There are two ways to pay it. On demand, you pay for each hour you use, with no commitment. With reserved capacity, you commit to a term of one to three years or more, and you pay a lower rate.

Other costs sit on top of the hourly rate. You pay for storage. Some providers also charge egress fees, which are fees to move data out of their cloud.

Why most break-even calculations get it wrong

The most cited study on this question is Lenovo’s on-premise vs cloud cost paper. Its math is transparent, which makes it useful. Its 2026 edition finds that owning an 8-GPU H200 server breaks even against on-demand cloud in about five months. Against a three-year cloud reservation, it breaks even in just over a year.

Those results rest on three assumptions worth checking.

The first is the cloud price. The paper compares its B300 server against AWS at $142.75 an hour for eight GPUs. That is about $17.84 per GPU-hour. The wider market charges far less. Silicon Data’s B300 rental price index tracks rental prices across hyperscalers, specialized GPU clouds known as neoclouds, and other providers. It stood at $6.95 per GPU-hour in October 2026.

That one input changes the answer. Using the paper's own B300 costs, break-even against AWS on demand comes after about 8.5 months of round-the-clock use. Against $6.95, it comes after about 28 months. That is our calculation, using the paper's formula.

The second assumption is end value. The paper spreads the server's cost over five years and assumes it is worth nothing at the end, while noting that a real system would have some value left. Leaving that value out makes owning look more expensive than it is.

The third is utilization. The break-even cases assume the server runs around the clock. Few enterprise clusters do.

Purchase prices vary too. Lenovo prices its 8-GPU B300 system at about $786,000. A US reseller listed a comparable Supermicro system at about $420,000 in July 2026. The right number is the quote in front of you.

What is AI hardware worth at the end of a lease?

End value, also called residual value, is what hardware is still worth when you finish using it. For GPU servers, it is hard to predict. Each new GPU generation lowers the value of the one before it, and the used market has changed quickly.

We asked CHG-MERIDIAN's international asset risk management team for estimates. We gave them two common configurations, purchased new for a lease starting July 1, 2027. The team values equipment per GPU. We multiplied by eight to get the value of a full server. We then compared each value with a public reference price to get a percentage. Those percentages are our calculation, not the team's.

Estimated end-of-lease values, as of October 2026, for a lease starting July 2027

Estimates provided by CHG-MERIDIAN’s international asset risk management team, October 2026. Values converted from euros at about $1.15 per euro. Percentages compare those values with public reference prices: about $420,000 for the B300 server, and €170,000 to €240,000 before VAT for the RTX PRO 6000 server, which explains its range.

These are estimates, not guarantees. GPU servers have a short resale history, and the used market has changed fast. Purchase prices were lower a few years ago, and used prices are high today. Actual values will depend on purchase price, configuration, and market conditions at the end of the lease. We will update these figures as the market develops.

Two points from the team are worth noting. First, the team values hardware by its age on a specific end date, not by the length of the lease. A 24-month lease that starts in 2027 is a different case from one that starts today. Second, the brand matters less than you might think. What drives value is the components inside: the GPUs, the processors, and the memory.

Worked example: what does a B300 server cost per GPU-hour?

A GPU-hour is one GPU running for one hour. Pricing everything per GPU-hour lets you compare your own hardware with a cloud rate directly. Here is the math for one 8-GPU B300 server used for 36 months.

Step 1: hardware value used. The server costs about $420,000. Our estimate says it will be worth about $114,000 after 36 months. So you use up about $306,000 of value.

Step 2: GPU-hours available. Thirty-six months is 26,280 hours. With eight GPUs, that is 210,240 GPU-hours.

Step 3: hardware cost per GPU-hour. Divide $306,000 by 210,240. That comes to about $1.46 per GPU-hour.

Step 4: running costs. We used Lenovo's assumptions here. Those are yearly maintenance at 12% of the system price, plus power, cooling, and colocation. For this server, that comes to about $11.43 an hour, or about $1.43 per GPU-hour.

Total: about $2.89 per GPU-hour, if the GPUs are busy every hour.

They won't be. Idle hours still cost money, so the real cost of each useful hour is higher. To get it, divide $2.89 by your utilization. At 70% utilization, each useful GPU-hour costs about $4.13. At 50%, it costs about $5.78.

That gives you the crossover. Divide $2.89 by a cloud rate, and you get the utilization you need for your own hardware to cost less.

Cloud rates as of October 2026. AWS lists its 8-GPU B300 instance at about $142.42 an hour, or about $17.80 per GPU-hour. The index rate is from Silicon Data. The three-year rate is an assumption for illustration, since multi-year prices vary by provider and contract.

The last column shows why end value matters. If you can commit to a cloud provider for three years at about $5 per GPU-hour, counting end value lowers the bar from about 69% utilization to about 58%.

These figures leave out staff and financing costs, which would raise the on-premise number. They also assume power costs stay the same when GPUs sit idle, which overstates them slightly. Treat the crossover as a starting range, not a final answer.

How does leasing change the crossover?

Leasing does not change the math above. You still use up about $306,000 of value, whether you buy the server or lease it. What leasing changes is when you pay, and who carries the risk.

In a fair market value lease, or FMV lease, the lessor sets the expected end value before the lease starts. Your payments cover the value you use, plus a financing cost. In our example, $306,000 spread over 36 months is about $8,500 a month before financing. Actual payments depend on your credit and on interest rates.

Three things change.

First, cash. You avoid paying $420,000 up front.

Second, risk. Suppose the server is worth only $80,000 at the end instead of $114,000. If you bought the server, that $34,000 loss is yours. In an FMV lease, the loss belongs to the lessor.

Third, refresh. At the end of the lease, you can return the hardware and move to a newer generation. You do not have to find a buyer for the old servers yourself.

One thing does not change. A lease is a fixed cost. Idle GPUs cost the same under a lease as they do when you own them. Utilization still decides whether on-premise beats the cloud.

CHG-MERIDIAN works this way when leasing servers and data center equipment, including AI clusters that mix vendors under one contract.

Why are enterprises moving AI workloads back on premise?

Cloud repatriation means moving workloads out of the public cloud and back onto hardware a company controls. For AI, it is common. In a 2026 Cloudera survey, two-thirds of respondents said their organizations had moved AI workloads out of public cloud in the past year.

Three reasons come up most. The first is steady demand. Once a model is in production, it often runs all day. That pushes utilization up, which is exactly where owned or leased hardware wins.

The second is data. Regulation, contracts, or internal policy may require sensitive data to stay on premises.

The third is predictability. A fixed monthly cost is easier to budget than a usage bill that grows with demand.

Few companies move everything. A common pattern is to keep experiments in the cloud and move steady production work on premise.

When does cloud still win?

The cloud wins when utilization is low or hard to predict. In our example, a server busy less than about 40% of the time costs more per useful hour than the market-wide index rate.

It also wins when you are still experimenting. If you don't yet know which models you will run, or for how long, a long commitment is a risk.

It wins when your facility can't host the hardware. Current GPU servers draw far more power than traditional servers, and many need liquid cooling.

And it wins when you need the newest GPUs right away for a short project. Our guides to GPU as a service vs leasing and to neoclouds for enterprise AI cover those options in more detail.

How to run the numbers for your own workload

Use these seven steps to find your own crossover.

1. Get a real quote. Price the servers, plus the networking and storage they need.
2. Pick the cloud rate you would actually pay. Compare on-demand, neocloud, and reserved rates, not just the list price.
3. Measure utilization honestly. Base it on the workloads you run today, not the ones you hope to run.
4. Add running costs. Include power, cooling, space, maintenance, and staff.
5. Set an end value, with a date. Ask a lessor or remarketer what the hardware will be worth on the day you stop using it.
6. Do the division. Take the total cost, subtract the end value, and divide by useful GPU-hours. Compare the result with the cloud rate.
7. Test the answer. Rerun it with lower utilization and a lower end value. If the answer flips, your decision is closer than it looks.

GPU hardware ages faster than the servers most IT teams are used to. So the cost question is also a lifecycle question. Decide how long you will use the hardware, what it will be worth at the end, and who carries the risk if that number comes in lower.

About the author Simon Harrsen EVP North America, CHG-MERIDIAN Simon Harrsen is EVP North America at CHG-MERIDIAN, with nearly 15 years of experience in the equipment leasing and technology lifecycle industry. He leads the company's growth and technology lifecycle strategy across the North American market, and has been quoted in and contributed bylines to publications including Monitor Daily, Channel Insider, Channel Dive, and Procurement Magazine, among many others. Connect on LinkedIn › | Email Simon ›