Buying looks like a fifth of the rental bill. Then you count the hours the GPUs sit idle.
Price out eight NVIDIA B200 GPUs two ways and the answer looks obvious. A Supermicro HGX B200 server with eight GPUs was last listed at $394,406 in September 2026. Renting the same eight GPUs on CoreWeave at its published $68.80 per node-hour, around the clock for three years, comes to about $1.81 million. Buying is roughly a fifth of the rental bill. Case closed.
Except the comparison is wrong in two directions at once. The purchase price leaves out power, colocation, networking, storage, people, and what the hardware is worth when you are done with it. The rental figure assumes you use every hour you pay for. Most enterprises do not come close. A June 2026 VentureBeat Research survey of 573 technical leaders found that 86 percent of enterprises running their own GPUs report utilization of 50 percent or less.
This article builds the full three-year total cost of ownership (TCO) for an owned B200 server and a used H100 server, line by line, with every input sourced or labelled as an assumption. The output is a single number you can test against your own fleet: the utilization rate at which buying beats renting.
For most of 2024 and 2025 the default advice was simple: rent, because rental prices were falling and hardware bought today would be worth less tomorrow. Three things broke that assumption this year.
First, rental prices stopped falling. SemiAnalysis, which surveys more than 100 market participants for its rental index, reported that one-year H100 contract pricing rose almost 40 percent, from a low of $1.70 per GPU-hour in October 2025 to $2.35 by March 2026. List prices followed. Nebius raised its on-demand HGX H100 rate from $3.85 to $4.50 per GPU-hour effective October 1, 2026, according to an IntuitionLabs review of first-party price lists.
Second, compute became a priced commodity. CME Group and Silicon Data listed the first regulated H100 and B200 rental index futures on October 5, 2026, each contract representing a month of rental for one GPU. Renters can now hedge the price risk that used to be the main argument for owning.
Third, the cost of owning went up too. Colocation vacancy in primary North American markets fell to a record 1.4 percent at the end of 2025, according to CBRE. On episode 7 of the DataStorage.com Podcast, Sunny Smith, founder and CTO of Massed Compute, said NVMe prices have roughly tripled and that he expects scarcity until the end of 2027. Both sides of the ledger moved, which is exactly why the old rule of thumb no longer holds.
The sticker price is the largest line item and the least interesting one. Here is the full stack for a single eight-GPU B200 server over three years.
OEM HGX B200 servers cluster around $384,000 to $394,000 in public listings (Supermicro, via Newegg, September 2026). Enterprise channel pricing runs higher: Lenovo Press lists a usual customer price of $550,475 for an eight-GPU B200 ThinkSystem SR680a V3, and a default NVIDIA DGX B200 configuration was listed at $562,154 by Broadberry in September. Our model uses the $394,406 OEM figure, which favors buying.
The server is not the cluster. Add a share of the network fabric, cabling, and local NVMe storage. We assume $60,000 per node, a conservative figure given the NVMe price shock. Storage is no longer a casual add-on.
NVIDIA rates the DGX B200 at a maximum of 14.3 kW. CBRE puts the average asking rate for 250 to 500 kW deployments in primary North American markets at $196.25 per kW per month in H2 2025, and that rate covers space, not electricity. Power comes on top: the U.S. Energy Information Administration (EIA) forecasts an average industrial electricity price of 9.01 cents per kWh in 2026. At a power usage effectiveness (PUE) of 1.3, that adds about $85 per kW per month.
The all-in figure, roughly $282 per kW per month, is a useful sanity check. American Compute, which appraises used GPU servers, assumes $280 per kW per month all-in for its own hosting model. Two independent routes, the same answer.
Someone has to rack the server, patch firmware, replace failed parts, and keep the scheduler fed. We assume $1,500 per month per server for fractional engineering time, remote hands, and a support contract. Teams that do not already run GPU infrastructure will pay far more, because the skills are scarce. Smith estimates only tens of thousands of people worldwide can build AI infrastructure end to end.
Your accountants will depreciate the server; your buyers will care what it sells for. The two numbers are not the same. AWS, Microsoft, and Google Cloud book servers over six years, while Nebius uses four and Lambda Labs five, according to theCUBE Research. Resale tells a harsher story. American Compute forecasts an H100 will be worth 11 to 19 percent of its original list price in 2027 in its conservative case. Our model assumes the B200 is worth 10 percent of its purchase price after three years.
| Cost line | B200 (new) | H100 (used) | Basis |
|---|---|---|---|
| Server purchase | $394,406 | $195,000 | Listing / Rack Report |
| Fabric, cabling, NVMe | $60,000 | $30,000 | Assumption |
| Colocation, 36 mo | $105,975 | $77,715 | CBRE $196.25/kW |
| Power, 36 mo | $44,018 | $31,397 | EIA 9.01¢/kWh |
| Operations, 36 mo | $54,000 | $54,000 | $1,500/mo |
| Less residual value | -$39,441 | -$9,750 | 10% / 5% |
| Total 3-year TCO | $618,958 | $378,362 | 8 GPUs |
| Cost per GPU-hour at 100% use | $2.94 | $1.80 | 210,240 hrs |
Put the stack together and an owned B200 server costs about $619,000 over three years. That buys 210,240 GPU-hours of capacity (8 GPUs x 8,760 hours x 3 years). If you use every one of them, your cost is $2.94 per GPU-hour.
Nobody uses every one of them. Ownership costs are fixed; rental costs scale with use. So the real cost of an owned GPU-hour is the fixed cost divided by the hours you actually consume. At 70 percent utilization the owned B200 costs $4.21 per useful GPU-hour. At 50 percent it costs $5.89. At 30 percent it costs $9.81, more than the most expensive rental on the market.
That gives a clean breakeven test. Divide the three-year owned cost by the cost of renting every available hour, and the result is the utilization you must sustain for buying to win:
Against the Silicon Data B200 rental index of $5.45 per GPU-hour, you need 54 percent utilization for the owned B200 to break even. Against Lambda Labs at $6.69, you need 44 percent. Against CoreWeave at $8.60, you need 34 percent.
The H100 math is tighter. A used eight-GPU HGX H100 node was valued at $195,000 in American Compute's October 2026 Rack Report. Run the same stack with a 10.2 kW draw, $30,000 of networking and storage, and a 5 percent residual, and the three-year cost is about $378,000, or $1.80 per GPU-hour at full use. Against the median on-demand H100 rate of $3.49, breakeven is 52 percent. Against the $2.35 one-year contract rate SemiAnalysis tracked in March, breakeven is 77 percent. Committed rental pricing is the real competitor to ownership, and it is a strong one.
| Rental benchmark | Rate / GPU-hr | Owned server | Breakeven |
|---|---|---|---|
| Silicon Data B200 index | $5.45 | B200 new | 54% |
| Lambda Labs B200, 8x | $6.69 | B200 new | 44% |
| CoreWeave B200, 8x node | $8.60 | B200 new | 34% |
| H100 1-year contract (SemiAnalysis, Mar 2026) | $2.35 | H100 used | 77% |
| H100 median on-demand (Oct 2026) | $3.49 | H100 used | 52% |
| CoreWeave H100, 8x node | $6.16 | H100 used | 29% |
Now compare those thresholds with reality. The VentureBeat finding that 86 percent of enterprises running their own GPUs sit at or below 50 percent utilization puts most of them under the B200 breakeven against index pricing. Cast AI's 2026 State of Kubernetes Optimization Report, based on telemetry from roughly 23,000 clusters, measured average GPU utilization at just 5 percent. Treat that figure with care: Cast AI sells optimization software, and Kubernetes clusters are not the whole market. But the independent survey points the same way. The hardware is not the expensive part. The idle hours are.
The flip side is equally clear. Teams running steady inference at 70 percent or higher, around the clock, are paying a large premium to rent. At 70 percent utilization the owned B200 costs $4.21 per useful hour against $6.69 at Lambda Labs and $8.60 at CoreWeave, a saving of roughly $366,000 to $647,000 per server over three years.
The breakeven number is the start of the decision, not the end. Several costs sit outside both the purchase order and the rental rate card.
Data movement. GPU capacity is scarce and shifting, so renters often move between AWS, a neocloud, and back again as availability changes. Every move drags training data and checkpoints along, and hyperscaler egress fees on that data can erase the savings that justified the switch. If your strategy is to rent, keep your datasets in zero-egress, S3-compatible storage such as Backblaze B2, Wasabi, or Cloudflare R2, so the data can follow the GPUs without a toll.
Who actually owns the hardware. Many GPU cloud brands resell capacity they do not own. Smith's first vetting question for any provider is simply whether it owns the GPU, because support tickets that pass between intermediaries stall. A cheap rate from a broker can cost more in lost engineering days than it saves.
Minimum purchase sizes. CoreWeave sells HGX nodes, not single GPUs, so a team that needs three GPUs pays for eight.
Capacity you cannot get. With colocation vacancy at 1.4 percent, the binding constraint is often power delivery dates, not server lead times. A server that arrives before its rack power does earns nothing.
The next generation. Blackwell Ultra is already listed: Lenovo prices an eight-GPU B300 ThinkSystem at $785,606. Owners carry the risk that a newer chip resets rental prices and resale values. Renters do not.
Cost of capital. Our model leaves out financing. At any positive interest rate, the breakeven utilization rises, because you pay for all of the capacity on day one.
The long tail works in the owner's favor, though. Smith notes that V100s are still rented at near capacity years after launch. A GPU that has finished training work can carry inference or batch jobs for years after the three-year window in this model.
Most teams should not choose one answer for the whole fleet. The strongest pattern in 2026 is to own the baseload and rent the peaks: buy enough capacity to cover the floor of demand you can prove, and rent everything above it.
Map your workloads first. Steady production inference with predictable traffic is the best candidate for ownership, because it runs around the clock. Training runs, experiments, and evaluation sweeps are bursty and belong on rented capacity, ideally on commitment terms that match the length of the project. Regulated data with residency requirements may push toward owned or dedicated hardware regardless of the math.
There is also a middle path. Some operators, Massed Compute among them, offer what Smith calls local metal: an on-premises GPU stack paid as operating expense, with hardware upgrades available on a 12-month cycle. It shifts obsolescence risk back to the provider while keeping data on your own floor.
Whichever mix you choose, reprice it regularly. Rental rates moved by double digits in both directions within 2026 alone. Track current rates across providers on the GPU Price Explorer, and if you rent at scale, the new CME futures give your finance team a way to lock in a price for planning.
Five steps turn this article into a decision for your team.
The GPU is not the expensive part of owning AI infrastructure. The hours it spends doing nothing are.
Compare AWS, Google Cloud, Azure, and alternatives like Backblaze B2 Discover how much you could save in seconds
