How does data center equipment financing work for GPU servers, cooling and power?
A GPU build is a lot more than the servers. Racks, switches, cooling and power gear go in with them, and all of it can be financed. Here’s how GPU and AI server financing works at $1 million and up: what the gear costs, who the programs are built for, how to set the terms and what the payments come to.
On this page
Private spending on data center construction in the U.S. ran at a seasonally adjusted annual rate of about $85 billion in August 2026, up about 73% from a year earlier, according to preliminary Census Bureau figures. That’s just the buildings. The servers inside can be out of date in a few years, while the UPS feeding them keeps running long after.
What data center equipment can you finance?
Nearly everything from the power feed to the GPUs, plus the labor to install it:
- GPU servers: 8-GPU HGX and DGX-class nodes, and rack-scale systems such as NVIDIA’s GB300 NVL72.
- Racks and storage for the compute.
- Networking: InfiniBand and Ethernet switches and the cabling between them.
- Cooling: air and liquid cooling, including coolant distribution units (CDUs).
- Power: UPS systems and switchgear.
- Installation and integration, as soft costs on qualifying deals.
New and used equipment can both qualify, depending on the program. The wider program is on our IT and data center hardware financing page. Pricing a few rack servers or a laptop fleet instead? Our guide to smaller server and laptop refreshes fits that better.
Why so much weight on power and cooling? NVIDIA lists a DGX B200, an 8-GPU system that fills 10 rack units, at about 14.3 kW max, and Supermicro puts a liquid-cooled GB300 NVL72 rack at 132 to 140 kW. Plan the financing around that load from the start.
How much of a deal can be soft costs depends on the program, so I won’t quote a percentage. Ask the integrator to split hardware, power and cooling, and labor on the quote.
Who qualifies for data center equipment financing?
Established companies with the revenue to carry a seven-figure equipment deal. The programs we place these deals with currently look for:
- Deal size: $1MM and up.
- Revenue: $25MM annual revenue required.
- Time in business: 3+ years preferred.
The $25MM figure is specific to these programs. Equipment finance has no universal revenue threshold. The fit here is an established company running AI or compute workloads, or building out data center capacity, with the revenue to back it up. If your project comes in under $1 million, the IT hardware page covers our regular programs, starting around $15,000.
Every one of these deals is above our $500,000 application-only limit, so plan on a full financial package. That typically means:
- CPA-prepared statements and tax returns, an interim P&L and balance sheet, a debt schedule, and AR and AP aging.
- An itemized quote from the OEM or integrator, with GPUs, networking, power and cooling, and installation on separate lines.
- The install schedule, since the in-service date decides the tax year.
Deals this size can come with covenants, too. The application starts with a soft credit inquiry, and you’ll typically hear back within 1 to 2 business hours, but a full package takes longer to underwrite than a small deal, so start well before the purchase order is due.
How much does a GPU server or AI server rack cost?
A complete new 8-GPU server lists for roughly $297,000 to $562,000, depending on the GPU generation and the builder, and a liquid-cooled rack-scale system runs about $5 million:
| Equipment | Price | Notes |
|---|---|---|
| 8-GPU HGX H100 server (GIGABYTE G593-SD0) | $273,000 | CPUs and memory not included |
| 8-GPU HGX H200 server (Supermicro SYS-821GE-TNHR) | $296,958 | Complete system, 2TB memory |
| Used 8-GPU H100 server | About $150,000 to $180,000 | Secondary-market estimate, April 2026, not a listing |
| 8-GPU HGX B200 server (Supermicro) | $394,406 to $414,599 | Air-cooled and liquid-cooled models |
| NVIDIA DGX B200 | About $562,154 | One reseller’s default configuration |
| GB300 NVL72 rack (72 GPUs, liquid cooled) | Just under $5.0 million; about $5.03 million with a 250 kW in-rack CDU | Reported purchase-order figure; media estimates run $6 to $6.5 million |
| 64-port 400G InfiniBand switch (NVIDIA Quantum-2) | About $44,544 | New |
| 150 to 200 kW three-phase UPS | About $59,317 to $74,971 | The 150 kW listing is a 400 V model, batteries extra |
Listed asking prices as of October 9, 2026, except the used H100 row (an April 2026 market estimate) and the GB300 NVL72 row (a reported purchase-order figure from August 2026). None are appraisals, offers or availability. The four new HGX server listings (H100, H200 and both B200s) and the 200 kW UPS were out of stock. Freight, installation and tax are usually extra.
So $1 million buys about two 8-GPU B200 servers plus the switches to connect them. Note the used number, too: 8-GPU H100 servers trading at about $150,000 to $180,000, well under what new Hopper-generation systems list for. The gap between new and used is why the term matters.
How long do GPUs last in a data center, and how long should you finance them?
On paper, about five to six years. The biggest buyers depreciate their servers over that span, and lately they’ve moved in opposite directions. From their 2025 annual filings:
- CoreWeave depreciates the computing equipment in its data centers over six years (it moved from five to six in 2023), and its data center equipment and leasehold improvements over up to 12 years (CoreWeave 10-K).
- Amazon cut the useful life of some servers and networking equipment from six years to five starting in 2025, citing the pace of AI and machine learning.
- Meta went the other way, lengthening the useful life of most servers and network assets to 5.5 years in 2025.
NVIDIA, meanwhile, now ships new data center products every year, and in January it said Rubin-based products would come from partners in the second half of 2026. At that pace, a GPU bought today could be two or three generations old within three years.
Power and cooling last longer. CoreWeave gives its data center equipment up to twice the life of its computing gear, and a UPS or cooling loop can stay put while the servers on it change.
So match the term to the equipment. I’d keep GPU servers on 36 or 48 months, inside even the shortest of those estimates, and put the UPS, switchgear and cooling on an equipment finance agreement (EFA) of up to 84 months. Leases top out at 60 months, so anything longer has to be an EFA.
Should you lease or buy GPU servers and data center equipment?
Lease the GPUs if you expect to replace them when the term ends, and own the power and cooling. FMV terms on GPUs depend on the program, since their resale value is hard to predict. Here’s how the structures compare:
| FMV lease | EFA or $1 buyout lease | |
|---|---|---|
| Who owns it during the term | The funder | You, on an EFA; on a $1 buyout lease, the funder until you pay the $1 |
| Terms at Five West | 24 to 60 months | 24 to 84 months on an EFA; 24 to 60 on a $1 buyout lease |
| End of the term | Return it, upgrade, renew, or buy it at fair market value | You own it |
| Payment on $1 million over 36 months | Lower than the EFA payment; set by the residual the program uses | About $31,800 |
| Section 179 | Generally no. You deduct the payments as rent. | Generally yes |
| Fits | GPU servers and network gear you plan to replace | UPS, switchgear, cooling, CDUs and racks you’ll keep |
Payments at an assumed 9%, in arrears, with no money down, for established businesses. The FMV payment depends on the residual each program sets, and GPU residuals are hard to predict, so it’s quoted deal by deal. Illustrative, not an offer.
Leasing AI servers or other data center equipment works best when you plan the refresh up front. Decide whether you’ll return, keep or upgrade the servers, and find the notice window for returns. If you’ll add nodes as demand grows, qualifying leases have add-on and upgrade paths mid-term. Not sure which structure a quote describes? Our FMV lease vs. EFA guide shows how to tell.
What are the monthly payments on $1 million or more of data center equipment?
Figure about $31,800 a month for every $1 million over 36 months, or about $20,758 over 60, at an assumed 9%:
| Amount financed | 36 months | 48 months | 60 months |
|---|---|---|---|
| $1,000,000 | About $31,800 | About $24,885 | About $20,758 |
| $2,500,000 | About $79,499 | About $62,213 | About $51,896 |
| $5,000,000 | About $158,999 | About $124,425 | About $103,792 |
Splitting the terms changes the math. Say a $2.5 million project is $1.8 million of GPU servers and networking plus $700,000 of UPS, switchgear and cooling:
| Piece | Amount | Term | Monthly payment |
|---|---|---|---|
| GPU servers and networking | $1,800,000 | 36 months | About $57,240 |
| UPS, switchgear and cooling | $700,000 | 84 months (EFA) | About $11,262 |
| Both, split as above | $2,500,000 | 36 and 84 months | About $68,502 for 36 months, then $11,262 |
| Everything on one term | $2,500,000 | 60 months | About $51,896 |
Example payments at an assumed 9%, in arrears, with no money down, on an EFA or $1 buyout lease, for established businesses only. Project sizes and splits are examples, not quotes, and the 84-month term needs an EFA. Illustrative, not an offer.
The split costs about $16,600 more a month for three years. Then the GPUs are paid off, right around when newer generations are shipping, the payment drops to $11,262, and total interest ends up about $107,000 lower than on the single 60-month term. The single term is easier on cash for the first three years. If the business can carry the higher payment, I’d take the split.
Does Section 179 apply to GPU servers and data center equipment in 2026?
Usually, yes, on equipment you own through an EFA or a $1 buyout lease, as long as it’s placed in service by the end of your tax year. For 2026 the Section 179 limit is $2,560,000, reduced dollar for dollar once you place more than $4,090,000 of qualifying property in service during the year (IRS Rev. Proc. 2025-32).
On projects this size, the phase-out can come into play. A company that puts $5 million of qualifying equipment in service in 2026 still has about $1.65 million of Section 179, and it’s gone at $6.65 million. The limit counts everything placed in service that year, and the deduction can’t exceed the business’s taxable income for the year (any excess carries forward). Bonus depreciation, now permanently 100% for qualified property acquired after January 19, 2025, can cover the rest (IRS guidance in Notice 2026-11). On an FMV lease, you deduct the payments as rent instead.
Timing matters, too. Placed in service means installed and ready to use, and all four HGX server listings in the price table were out of stock in October. For a calendar-year business, a server that arrives in late December but isn’t ready to use by December 31 counts next year. Your tax advisor should confirm the treatment.
Can a newer AI company or GPU cloud finance data center equipment?
Sometimes. It depends on the project’s size and how far along the company is. The programs above are built around established revenue and time in business, so a newer company often starts with a smaller project.
A few things make a newer company’s file stronger:
- Revenue history, even a short one.
- A down payment, which shrinks the amount financed.
- Cash left over after the purchase.
- Strong personal credit and industry experience. A business under a year old is underwritten as a startup, which typically means 700+ personal credit.
For a GPU cloud, signed customer contracts and a billing history make the file stronger. For what startup programs look for, see our guide to startup equipment financing.
Frequently asked questions
Can you finance GPU servers and AI infrastructure?
Yes. GPU servers, including 8-GPU HGX and DGX-class systems and rack-scale systems, can be financed along with racks, networking, cooling, coolant distribution units, UPS systems and switchgear. Installation and integration can go on the same agreement on qualifying deals. The programs we place these deals with currently look for deals of $1MM and up.
What do you need to qualify for data center equipment financing?
The programs we place these deals with currently look for deals of $1MM and up, with $25MM annual revenue required and 3+ years in business preferred. Every deal this size is above the $500,000 application-only limit, so expect to send a full financial package, including financial statements, tax returns, a debt schedule and interim figures.
How much does an 8-GPU server cost?
As of October 9, 2026, new 8-GPU HGX H100 and H200 servers listed for about $273,000 (without CPUs or memory) to $297,000, 8-GPU HGX B200 servers for about $394,000 to $415,000, and an NVIDIA DGX B200 for about $562,000 in one reseller’s default configuration. A GB300 NVL72 rack with 72 GPUs runs just under $5 million on reported purchase orders. These are listed or reported prices, not offers.
How long should you finance GPUs?
Usually for less time than the power and cooling. Large buyers depreciate servers over about five to six years, and NVIDIA releases new data center products every year, so a 36- or 48-month term keeps the payments inside the GPUs’ useful life. UPS systems, switchgear and cooling can go longer on an equipment finance agreement, up to 84 months.
Can you lease AI servers instead of buying them?
Yes. A fair market value lease runs 24 to 60 months and has a lower payment than an equipment finance agreement, with the option to return the servers, upgrade, renew or buy them at the end of the term. It fits GPU servers you expect to replace, though FMV terms on GPUs vary by program because their resale value is hard to predict. Equipment you’ll keep, like UPS systems and cooling, is usually financed and owned.
Can installation and integration be financed with data center equipment?
Yes, on qualifying transactions. Installation, integration and configuration can go on the same agreement as the hardware, so the project is one payment. Programs set their own limits on soft costs, so ask the integrator for a quote that lists hardware, power and cooling, and labor separately.
Does Section 179 apply to GPU servers in 2026?
Usually, yes, on equipment financed with an equipment finance agreement or $1 buyout lease and placed in service by the end of the tax year. The 2026 limit is $2,560,000, reduced dollar for dollar once more than $4,090,000 of qualifying property is placed in service in the year, so it phases out completely at $6,650,000. 100% bonus depreciation, permanent for qualified property acquired after January 19, 2025, can cover the rest. On a fair market value lease you deduct the payments as rent. Confirm with your tax advisor.
Can a startup or new GPU cloud get data center equipment financing?
It depends on the size of the project and how established the company is. The programs for $1MM-and-up data center deals are built around established revenue and time in business. Newer companies make a file stronger with revenue history, a down payment, cash left after the purchase, and, for a business under a year old, 700+ personal credit and industry experience. For a GPU cloud, signed customer contracts and a billing history help, too.
Price your data center project.
Enter the project total, term and structure in the Quote Builder for an estimated payment, with no credit pull. It runs up to $2 million, and bigger projects are priced by hand once you apply. Either way, one application matches you with a dedicated funding professional who handles the deal through funding.
This article is general information about financing data center and AI infrastructure equipment. It is not tax, legal, or accounting advice, or a commitment to finance. Prices shown are listed asking prices and reported figures, not appraisals or offers. Tax and accounting treatment depends on your contract and your business; confirm with your CPA. All financing is subject to credit approval and underwriting.