How much does an AI server cost? From purchase to retirement
 |
Simon Harrsen - EVP North America
Published October 28, 2026 · 8 min read
|
An AI server with eight NVIDIA Blackwell GPU chips costs about $360,000 to $470,000 to buy, as of October 2026. That price covers the server itself. Power, software licenses, and end-of-life costs come on top. Those costs apply whether a company buys or leases, which puts them at the center of AI infrastructure financing.
This article follows one AI server through its life. It starts with the purchase price. It then covers what the server costs to run. It ends with what the server costs to retire, and what it is worth when it leaves. Together, those three stages make up the total cost of ownership, sometimes called AI infrastructure TCO.
- An 8-GPU AI server built on NVIDIA's HGX B200 or B300 platform costs about $360,000 to $470,000 as of October 2026.
- NVIDIA AI Enterprise software lists at $4,500 per GPU per year, or $36,000 per 8-GPU server.
- One NVIDIA B300 server draws up to 14.5 kilowatts, nearly twice the power of an average data center rack.
- Power for one B300 server costs about $13,000 to $22,000 a year at US business electricity rates.
- Data wiping and testing at retirement typically cost a few hundred dollars per server, according to CHG-MERIDIAN's remarketing team.
- CHG-MERIDIAN estimates that a B300 server keeps about 27% of its purchase price after three years.
How much does an AI server cost to buy?
An AI server costs about $340,000 to $470,000 for a standard 8-GPU system, and about $5 million for a full rack-scale system. A GPU, or graphics processing unit, is the chip that does the heavy math in AI work. The price depends mostly on which GPU generation the server uses and how many GPU chips it holds.
Most enterprise AI servers follow an NVIDIA design called HGX. HGX is a board that holds eight GPU chips and connects them at high speed. Dell, HPE, Lenovo, Supermicro, and other manufacturers build servers around it. NVIDIA also sells its own complete systems under the DGX name. DGX systems typically cost more than comparable servers from other manufacturers.
Typical AI server prices by tier, as of October 2026:
| Tier | System | Typical price |
| Entry | 8x NVIDIA RTX PRO 6000 | €170,000 to €240,000 |
| Previous generation | NVIDIA HGX H200, 8 GPUs | About $340,000 |
| Current generation | NVIDIA HGX B200, 8 GPUs | $360,000 to $470,000 |
| Current generation | NVIDIA HGX B300, 8 GPUs | $400,000 to $450,000 |
| Rack scale | NVIDIA GB300 NVL72, 72 GPUs | About $5 million, or $5.7 million installed |
HGX prices are new-equipment benchmarks from American Compute’s October 2026 Rack Report. NVL72 figures come from purchase orders reviewed by Wing VC in August 2026. The RTX PRO 6000 range is from European reseller listings before VAT, since no reliable US price is published. Prices change quickly, so treat them as a starting point for a quote.
What the purchase price includes
The price of an HGX server covers the server itself. That means the GPU chips, the processors, the memory, the local drives, and the network cards inside the server. The price does not cover the network that links servers together, shared storage, racks, or installation. Those items are quoted separately for each project. Their cost depends on how many servers you buy and how your data center is set up.
Why AI server prices are rising in 2026
AI server prices are rising because memory costs have climbed sharply. In August 2026, Bloomberg reported that NVIDIA told some of its largest customers to expect price increases of more than 15% on servers with its AI chips. NVIDIA had not confirmed the increase at the time of reporting.
Demand is also outrunning supply. CHG-MERIDIAN’s remarketing team reports that buyers are not getting volume discounts on AI servers right now. Ordering ten servers instead of one does not lower the price per server.
What does an AI server cost to run?
Running one 8-GPU AI server costs about $49,000 to $58,000 a year in power and NVIDIA software alone. Power adds about $13,000 to $22,000. Software adds $36,000. Maintenance, support, and staff come on top, and those costs vary widely from one company to the next.
Power
An NVIDIA DGX B300 system is rated at 14.5 kilowatts. A kilowatt measures how much power a machine draws at one moment. Here is a worked example of what that means for the electricity bill.
Assume the server runs at 80% of its rated power, around the clock. That uses about 101,600 kilowatt-hours a year. A kilowatt-hour is one kilowatt used for one hour, and it is the unit on an electricity bill. Data centers also use power for cooling and other overhead. The Uptime Institute measures that overhead with a ratio called power usage effectiveness, or PUE. The Uptime Institute’s 2026 PUE findings put the average between 1.36 and 1.52. That adds 36% to 52% on top of the server’s own use.
US electricity for businesses averaged 9.64 to 14.47 cents per kilowatt-hour in August 2026, according to the US Energy Information Administration. The industrial rate is the low end, and the commercial rate is the high end. Put together, one B300 server costs about $13,000 to $22,000 a year to power.
Software
NVIDIA AI Enterprise is the software suite NVIDIA sells for running AI workloads in production. It lists at $4,500 per GPU per year, according to NVIDIA’s published license pricing. For an 8-GPU server, that is $36,000 a year. Over three years, it adds $108,000, about a quarter of the server’s purchase price. Few cost guides mention this line.
Facility readiness
Many enterprise data centers were not built for AI servers. The Uptime Institute’s 2026 survey found that the average data center rack is designed for about 7.8 kilowatts. A single B300 server draws 14.5 kilowatts, nearly twice that. A full GB300 NVL72 rack draws about 135 kilowatts and needs liquid cooling.
Liquid cooling raises costs at both ends of a server’s life. CHG-MERIDIAN’s remarketing team reports that liquid-cooled systems typically cost more upfront. They also sell for less on the used market, because fewer buyers can run them.
Maintenance, support, and staff
Support contracts, repairs, and the engineers who run the servers are real costs. They vary too much to give a reliable range. The main drivers are the length of the manufacturer’s warranty, the support level you choose, and how much of the work your own team handles. Ask for these costs in writing before you buy. The right budget also depends on how steadily you will use the hardware, which is the core question in the on-prem vs cloud cost crossover.
What does it cost to retire an AI server?
Retiring an AI server costs far less than buying or running one, and the server’s resale value usually covers the cost. Retirement is handled through IT asset disposition, or ITAD. ITAD is the process of removing old IT equipment, erasing its data, and then reselling or recycling it.
The steps in retiring an AI server
An ITAD project for AI servers usually covers five steps.
- Removal. Technicians disconnect the servers and take them out of the racks. Liquid-cooled systems take extra work, because the cooling lines must be drained and disconnected.
- Packing and shipping. The servers are packed and moved to a processing site. Cost depends on your location and how many servers are moving.
- Data wiping and testing. Every drive is erased, and the server is tested so it can be resold. An NVIDIA DGX B300 has 10 drives: two for the operating system and eight for local data. Servers from other manufacturers vary. CHG-MERIDIAN’s remarketing team reports that wiping and testing typically cost a few hundred dollars per server.
- Certificates and reporting. You receive proof that the data was erased and a record of where each server went.
- Resale or recycling. Working servers are resold. Parts that cannot be reused are recycled.
Removal, packing, and shipping are quoted per project. Ask for those costs before the end of the server’s life, not after.
What an AI server is worth at retirement
AI servers lose value fast, because each new GPU generation makes the last one less attractive. CHG-MERIDIAN’s international asset risk management team estimates that an NVIDIA B300 server keeps about a third of its purchase price after two years. It keeps about a quarter after three years, and about a fifth after four years. These are market estimates as of October 2026, and they will change with the market. Our analysis of how fast GPU servers depreciate covers the full numbers.
Even at those rates, resale value is far higher than the cost of wiping and testing. A B300 server worth about a quarter of its purchase price is worth over $100,000. Wiping and testing cost a few hundred dollars.
What does an AI server cost over three years?
One NVIDIA B300 server costs about $454,000 to $481,000 over three years, after resale value comes back. That total covers purchase, software, and power. It does not include networking, storage, installation, support, staff, or removal and shipping, which are quoted per project.
Three years is used here because it is a common lease term, and because CHG-MERIDIAN’s resale estimates are strongest in that range. Here is the worked example for one server.
| Cost | Amount | Basis |
| Purchase | $420,000 | Within the $400,000 to $450,000 benchmark |
| NVIDIA AI Enterprise | $108,000 | $4,500 per GPU per year, 8 GPUs, 3 years |
| Power | $40,000 to $67,000 | $13,000 to $22,000 a year |
| Data wiping and testing | A few hundred dollars | CHG-MERIDIAN remarketing team |
| Removal and shipping | Quoted per project | Varies by location and volume |
| Resale value | About $114,000 back | About 27% after three years |
| Net three-year cost | About $454,000 to $481,000 | Before items quoted per project |
The example shows where the money goes. The purchase price is the largest single cost. Software and power add more than a third on top of it over three years. Resale value then returns about a quarter of the purchase price. Building these numbers into a budget early is the core of AI capital expenditure planning.
Owning vs leasing an AI server at end of life
The biggest difference between owning and leasing an AI server shows up at the end of its life. An owner must resell the server and carries the risk of what it will be worth. A lessee returns the server and moves on to the next generation.
Under a typical lease with CHG-MERIDIAN, some end-of-life tasks stay with the customer. The customer usually handles removal, return shipping, and data wiping. CHG-MERIDIAN handles reporting, resale, and recycling.
| Task | Owned server | Leased server (typical) |
| Removal and return shipping | Owner | Customer |
| Data wiping | Owner | Customer |
| Storing old equipment | Owner | Not needed |
| Resale and recycling | Owner | CHG-MERIDIAN |
| Reporting | Owner | CHG-MERIDIAN |
| Risk on resale value | Owner | CHG-MERIDIAN |
| Upgrade timing | Often delayed by depreciation | Set by the lease term |
That last row matters for AI hardware. Companies that own their servers often depreciate them over five or six years. That can push them to keep a server long after a newer generation arrives. A fair market value lease sets payments based on the value the server is expected to lose during the term. At the end, the customer returns the server and can upgrade on schedule. CHG-MERIDIAN offers this structure through server and data center equipment leasing.
Talk to us and discover how our expertise can benefit your business.
Frequently asked questions
An AI server with eight current-generation NVIDIA GPUs costs about $360,000 to $470,000 as of October 2026. Previous-generation H200 servers cost about $340,000. A full NVIDIA GB300 NVL72 rack with 72 GPUs costs about $5 million for the hardware. Prices are rising, so treat any figure as a starting point for a quote.
An AI server is a computer built to train and run artificial intelligence models. It holds several GPUs, usually eight, which handle the large volume of math that AI work requires. AI servers also need more power and cooling than standard servers.
An 8-GPU NVIDIA HGX H200 server costs about $340,000 new. An 8-GPU HGX B200 server costs about $360,000 to $470,000. These are October 2026 benchmarks from American Compute's Rack Report, and actual quotes vary by manufacturer and configuration.
An NVIDIA DGX B300 system is rated at 14.5 kilowatts, and a DGX B200 system at up to 14.3 kilowatts. That is nearly twice the 7.8 kilowatt design of an average data center rack. A full GB300 NVL72 rack draws about 135 kilowatts.
Power and NVIDIA software for one 8-GPU B300 server cost about $49,000 to $58,000 a year. Power is about $13,000 to $22,000 at US business electricity rates. NVIDIA AI Enterprise software is $36,000. Maintenance, support, and staff come on top.
AI server prices are rising mainly because memory costs have climbed sharply. In August 2026, Bloomberg reported that NVIDIA told major customers to expect price increases of more than 15%. Demand also exceeds supply, so volume discounts are rare.
ITAD stands for IT asset disposition. It is the process of retiring IT equipment: removing it, erasing its data, documenting the process, and reselling or recycling it.
Data wiping and testing typically cost a few hundred dollars per server, according to CHG-MERIDIAN's remarketing team. Removal, packing, and shipping are quoted per project. Resale value is usually far higher than these costs.
CHG-MERIDIAN's international asset risk management team estimates that an NVIDIA B300 server keeps about 27% of its purchase price after three years. These are market estimates as of October 2026 and will change with the market.
Buying suits companies that will run the same hardware for its full five-to-six-year life. Leasing suits companies that want to upgrade as new GPU generations arrive. With a fair market value lease, the leasing company takes on the resale risk, and the customer returns the server at the end of the term.