AI server cost in 2026: what the hardware costs to buy, build or rent

An AI server costs about $5,000 to $25,000 as a desk machine with one GPU, about $150,000 as a server with eight 96 GB cards, about $360,000 to $550,000 as an eight-GPU HGX Blackwell server, and $2.8 million to $7.8 million as a rack, on prices dated between January and November 2026. NVIDIA publishes prices for the DGX Spark and its workstation cards but not for servers, so nearly every server and card price below comes from one US dealer, Central Computer, as listed on 2026-11-12. The rest come from named reports, and where sources disagree I give the range. The invoice is not the whole bill: power, cooling, space, networking and spares come on top. In the worked example at the end, an eight-card server has to be busy about 54 hours a week for three years before buying it beats renting the same GPUs by the hour.

What AI server hardware costs, by tier#

A desk machine: $4,999 to about $25,000

The machines that sit on or under a desk run from a 1.2 kg box to a 1,600 W tower. These are the dated prices I found:

The GPU decides most of the price, and GPU prices moved a long way in 2026. The card is 59% of that RTX PRO 6000 workstation (our calculation: 14,999.99 / 25,292.90), and the card itself launched at $8,565 in April 2025 and was listed at $16,000 on NVIDIA's US store in August 2026, according to Tom's Hardware. The RTX 5090 launched at $1,999 in January 2025, and on 2026-09-05 Tom's Hardware reported it "regularly listed for above $5,000". The DGX Station is a different class of machine: 748 GB of memory in all, and a 1,600 W draw that NVIDIA says needs a 20 A circuit. The DGX Spark guide and the RTX PRO 6000 editions guide work out what each desk machine is worth in rented hours.

A server with four to eight 48 GB or 96 GB cards: about $150,000

The eight-card server I could price comes to $145,759.81 before storage. On 2026-11-12, Central Computer's configurator for an ASUS ESC8000A-E13P 4U server with eight RTX PRO 6000 Blackwell Server Edition cards started at that figure:

The cards are 82% of the price (our calculation: 119,999.92 / 145,759.81). The memory in that listing is the minimum, not a recommendation: eight 128 GB modules instead of eight 16 GB ones cost $43,999.92 on the same configurator, $36,240 more.

A server of 48 GB cards costs less because the cards do. On 2026-11-12 the same dealer listed the 48 GB RTX PRO 5000 Blackwell at $7,999.99, in stock, and the 48 GB RTX 6000 Ada at $8,999.99, unavailable online, so eight cards come to $63,999.92 to $71,999.92, against $119,999.92 for eight RTX PRO 6000s (our calculation). I did not find a listing for a configured eight-card 48 GB server to quote whole. The VRAM guide shows what fits in 48 and in 96 GB, which is the trade you are paying for.

An eight-GPU HGX server: about $360,000 to $550,000

HGX servers are where list prices stop. HGX is the GPU board, usually with eight GPUs, that server makers build their systems around, and NVIDIA publishes no price for those systems. On 2026-11-12 I could open only a few US dealer listings, so the table adds Reuters' figure for a B300 server:

The two ASRock listings are the chassis and the GPU board alone, so a working server costs more: the configured B200 above comes to $363,698.95, and Reuters put an eight-GPU B300 server at about $550,000. One dealer is not a market, so read its listings as the start of a quote, not the price you will pay. The only figure from NVIDIA for this generation is Jensen Huang's estimate in March 2024 that a Blackwell GPU would cost $30,000 to $40,000, which puts eight GPUs at $240,000 to $320,000 before the rest of the server (our calculation), and CNBC noted at the time that NVIDIA does not reveal list prices. For the H200, the PCIe route is the one with a public price: four H200 NVL cards at $31,999.99 each on 2026-11-12, when the NVIDIA-branded card showed an ETA of 30 weeks. The H200 price guide and the B200 page cover the cards, and the DGX B300 guide explains DGX against HGX.

A rack: $2.8 million to $7.8 million

Rack-scale systems have no list price either, and the two published estimates for a GB300 rack differ by $1.7 million to $2.2 million:

Tom's Hardware's source warned that many quotes online are a manufacturer's price without a proper warranty, and the same report says NVIDIA has never confirmed the list prices of its NVL72 products. Prices are still moving too. On 2026-08-23, Tom's Hardware reported Bloomberg's story that NVIDIA had told some of its largest customers that server prices will rise by more than 15% in many cases, for Grace Blackwell and Vera Rubin systems shipping early in 2027. The GB300 page and the GB200 page cover what each rack holds.

The costs that are not on the invoice#

Power, at 14.47 cents per kWh

Power is the one hidden cost you can work out exactly. The EIA put the average US commercial electricity price at 14.47 cents per kWh in August 2026, and a year has 8,760 hours, so at full load around the clock (our calculation: kW x 8,760 x $0.1447):

The server row counts only the GPUs and CPUs at their rated limits, 8 x 600 W plus 2 x 280 W, so fans, memory and power-supply losses come on top, and a server that sits idle half the day draws less. The DGX B300 and GB300 figures are NVIDIA's own. At home the residential rate applies, 18.16 cents per kWh in the same EIA table, which puts a DGX Spark running all year at its full 240 W at $382.

Cooling

Cooling adds about half again to the power bill in an average facility. The Uptime Institute's 2026 survey found an industry-wide average power usage effectiveness of 1.52, meaning the building uses 1.52 kWh for every kWh the IT equipment uses, and 1.36 when larger facilities are weighted by their size. At 1.52 the eight-card server's $6,794 a year becomes $10,327, a DGX B300's $17,746 becomes $26,974, and a GB300 NVL72's $179,995 becomes $273,593 (our calculation: x 1.52).

The cooling hardware changes with the tier. Passive server cards depend on the chassis: NVIDIA's product brief asks for at least 43 CFM of ducted airflow through each RTX PRO 6000 Server Edition card at a 25 C inlet and 125 CFM at 45 C. A GB300 NVL72 is liquid-cooled, and Morgan Stanley put the value of its liquid-cooling components at $49,860 per rack, Tom's Hardware reported on 2025-11-05.

A place to put it

The rack's power limit decides what you can install before the price does. NVIDIA, citing a survey, says roughly 70% of enterprise racks are 20 kW and below and use air cooling. One DGX B300 at about 14 kW takes 70% of such a rack's power on its own (our calculation: 14 / 20), and a GB300 NVL72 needs up to 142 kW and liquid cooling, which means a facility built for it.

Colocation is sold by the kW of power you reserve, and in 2026 space was scarce. CBRE's North America Data Center Trends report for the first half of 2026 put vacancy across the primary markets at a record low of 1.4%, and colocation vacancy in Northern Virginia at 0.24%, with asking rates up 4.3% in six months for 250 to 500 kW requirements and 6.7% for requirements over 10 MW. In Northern Virginia, CBRE reports that all capacity delivering in 2026 has been committed, pushing preleasing into 2027 and beyond, so ask for the space quote at the same time as the server quote.

Networking

The network grows with the GPU count. A single server needs a port on your network and a fast way to move models in. A cluster needs a GPU fabric: NVIDIA's DGX B300 has eight 800 Gb/s ConnectX-8 ports for it, plus two dual-port BlueField-3 DPUs at up to 400 Gb/s, so every server in the cluster takes eight fabric switch ports before storage and management. Even four DGX Sparks need a 200 GbE switch with QSFP56-DD ports and RoCE v2, NVIDIA says. I did not find switch prices I could check, so they belong in the build quote. InfiniBand vs Ethernet and InfiniBand vs NVLink explain which link carries what.

Spares and lead times

A failed card costs you the weeks until its replacement arrives. On 2026-11-12, Central Computer showed an ETA of 30 weeks for NVIDIA's own H200 NVL card, CDW showed 4 to 6+ weeks for PNY's version, and CDW had all three RTX PRO 6000 editions on backorder. A spare RTX PRO 6000 for the eight-card server cost $14,999.99 at the same dealer that day, an eighth of its GPU spend (our calculation: 14,999.99 / 119,999.92 = 0.125). Warranties vary by seller: Central Computer's custom builds come with a one-year basic system warranty, and Exxact listed its DGX Station with a three-year warranty.

What the hardware is worth in three years

Nobody publishes a resale price you can plan with. I have not found a reputable, dated source for used data-center GPU prices, so the break-even below counts the hardware as worth nothing at the end, and whatever you sell it for comes off the purchase price. Two things are on record. Prices rose in 2026 rather than fell: the RTX PRO 6000 went from a launch price of $8,565 in April 2025 to $16,000 on NVIDIA's US store in August 2026, and the DGX Spark from $3,999 at launch in October 2025 to $6,950 in October 2026, both blamed on memory costs. And generations turn over fast: NVIDIA put the A100 into full production in May 2020, and One Stop Systems, a server maker, set April 26, 2024 as its last order date for the A100 80GB. The A100 80GB still rents on QuantaCloud from $1.48/GPU-hr, six years after it went into production, so an older GPU can keep doing paid work long after it stops being sold.

Three ways to get one#

Buy it

Buying gives you the parts you choose and every cost on this page: the purchase, the power and cooling, the space, the network, the spares and the resale risk. It fits when you already run racks with the power and airflow for 600 W cards, and your measured use keeps the GPUs busy past the break-even below. The on-premise LLM guide covers what running the hardware yourself involves.

Have it built as dedicated capacity

QuantaCloud builds dedicated GPU servers and clusters to order: we order and build the hardware to your spec, and the configuration, lead time and terms come back in writing before you commit. You choose the GPU, the count, the network and the term, and you skip the parts list and the dealer search. Dedicated GPU servers covers single servers, GPU clusters covers multi-node builds, and the reserved capacity page explains the process. Send a capacity brief with the GPU, the count and the start date.

Rent by the hour

Renting moves every cost above into one hourly price. These are the live prices for the GPUs QuantaCloud rents:

Prices checked 14 Nov 2026, 09:38 UTC

Each instance is a VM with Ubuntu 22.04, the NVIDIA driver and Docker, in US regions only. The first hour is charged at launch, and the unused seconds of the current hour are refunded when you stop. Stopping terminates the VM and deletes its disk, and there are no volumes or snapshots, so a rented GPU suits work you can restart from files you copied off. The catalog moves with capacity: on 2026-11-12 it listed no H100 or H200, so for those, check the table above or have the server built. The pricing page and the billing docs have the rules, and reserved vs on-demand GPUs compares renting with a reserved term.

Break-even: buying against renting#

The worked example is the eight-card RTX PRO 6000 server at $149,738.80, against renting eight RTX PRO 6000 GPUs at $2.39 each, the 1-GPU price on 2026-11-12 (today $2.16/GPU-hr). That day the catalog listed the card one per VM, so eight at once means eight 1-GPU VMs, or a larger VM when the catalog lists one. The arithmetic, our calculation:

1. Renting eight GPUs costs 8 x $2.39 = $19.12 an hour.
2. On hardware alone, the purchase price equals $149,738.80 / $19.12 = 7,832 hours of rent.
3. Owning costs power and cooling for every busy hour: 5.36 kW x 1.52 x $0.1447 = $1.18. So each busy hour on your own server saves $19.12 - $1.18 = $17.94.
4. The break-even is $149,738.80 / $17.94 = 8,346 busy hours.
5. That is 0.95 years running around the clock, 1.9 years at 50% use and 3.8 years at 25%. To break even inside three years, the server must be busy 8,346 / 156 weeks = 53.5 hours a week, and inside five years 32.1 hours a week.

The sum leaves out the rack space, the network, spares, idle power and staff time, and every one of those moves the break-even later. It also counts the server as worth nothing at the end, and resale is the one item that moves it earlier. Renting includes more host per GPU than this configuration, too: each 1-GPU VM came with 16 vCPUs, 144 GB of RAM and 725 GB of disk.

For a single card, divide its price by the hourly price:

Those are card prices, before the host server, power and space, so the real break-even comes later than the table says. The GPU cost calculator runs the same sum for your own hours at live prices, and the H100 price guide, A100 price guide and H200 price guide work through each card in detail.

FAQ#

How much does an AI server cost?

On the dated prices above: about $5,000 to $25,000 for a desk machine with one GPU, about $146,000 to $150,000 for a server with eight 96 GB RTX PRO 6000 cards, about $360,000 to $550,000 for an eight-GPU HGX Blackwell server, and $2.8 million to $7.8 million for a rack. Most of the server prices come from one dealer's listings on 2026-11-12, so treat them as a starting point for a quote. A DGX Station sits between the first two tiers at $94,930 to $123,569.99.

What does a local AI server cost to run?

Mostly power and cooling. The eight-card server above uses about $566 of electricity a month at full load at the US commercial average of 14.47 cents per kWh, and about $861 with an average facility's cooling overhead (our calculation: $6,794 and $10,327 a year, divided by 12). A DGX Spark at home uses at most about $32 a month (our calculation: $382 / 12).

Should I buy or rent a GPU server?

Rent until you have measured how busy the GPUs will be. In the worked example, an eight-card server pays for itself against renting only if it is busy more than about 54 hours a week for three years, or 32 hours a week for five, before space, network and staff. If your measured hours clear that line, buy it or have it built.

Why do HGX servers and NVL72 racks have no list price?

NVIDIA does not publish one. Server makers build HGX systems with the CPUs, memory, storage and network cards each buyer picks, and CNBC reported in 2024 that what a buyer pays depends on volume and on whether the GPUs come in a complete system. For NVL72 racks, Tom's Hardware reports that NVIDIA has never confirmed a list price.

Can I have a server built instead of buying one?

Yes, as reserved capacity built to order. We order and build the hardware to your spec, from one server to InfiniBand clusters, and the configuration, lead time and terms come back in writing. Send a capacity brief, or read how dedicated GPU servers work first.

What happens to my data on a rented GPU?

It is deleted when you stop. Stopping an instance terminates it and deletes its disk, and there are no volumes or snapshots, so copy checkpoints and results off before you stop. The unused seconds of the current hour go back to your balance.

My rule is to price the busy hours before the hardware. An AI server is cheap per hour only when it is busy, and in 2026 the hardware itself got more expensive, not less. Rent the GPU you would buy for a month, count the hours it is really working, and put that number into the break-even above. If it says to own, send a brief and compare the written quote with the dealer's before you sign either.