Neocloud Rental vs Buying GPU Servers in 2026: Which Costs Less?

Renting from a neocloud wins when demand is short, uneven or uncertain, and buying wins when the GPUs will stay busy for years and you have somewhere to host them. Against on-demand rates, buying a B200 server pays back above roughly one-third utilization on the published model used below. Against a multi-year reserved contract the gap is much smaller, and the decision turns on contract risk, hosting and financing more than on the hourly rate. This guide gives current list rates, shows the break-even method, and lists the contract terms to check. For Hopper GPUs and a full worked example, see our guide to H100 and H200 prices. All prices are reported or listed third-party figures, not Nexus Compute prices.

What is a neocloud?

A neocloud is a specialist cloud provider that mainly rents GPU capacity for AI work, rather than offering the full range of a hyperscaler such as AWS, Azure or Google Cloud. CoreWeave, Lambda and Nebius are well-known examples. The Uvik GPU Price Index, which sampled 19 providers on October 7, 2026, found hyperscalers charging 1.7 to 2.7 times the GPU-cloud price per GPU-hour, median against median.

How much does it cost to rent GPUs from a neocloud in 2026?

These are first-party list prices per GPU-hour, read from each provider's pricing page on November 6, 2026. They exclude tax, can change at any time, and may not be available in the quantity you want.

• B200 on demand: $7.69 on an 8-GPU instance at Lambda, $8.50 at Nebius, and $8.60 at CoreWeave ($68.80 for an 8-GPU node).
• H100 on demand: $4.50 at Nebius, $4.59 on an 8-GPU instance at Lambda, and $6.16 at CoreWeave.
• Short clusters: Lambda lists B200 1-Click Clusters of 2 weeks to 1 year at $9.86 for 16 GPUs, falling to $8.87 at 256 or more.
• Hyperscaler comparison: AWS Capacity Blocks list at $14.208 for B200 and $5.970 for H100 in most regions.
• Spot or preemptible: Nebius lists B200 from $0.99, with spot prices that can change as often as every 15 minutes. CoreWeave listed B200 spot at about $34 to $35 for an 8-GPU node, roughly $4.30 per GPU.
• Index median, October 7, 2026: Uvik puts H100 SXM at $4.20 and B200 at $8.19 on demand.

Rates are rising, not falling. Uvik Software's index records Nebius moving its H100 rate from $3.85 to $4.50 on October 1, 2026. SemiAnalysis reported that H100 one-year contracts rose from $1.70 in October 2025 to $2.35 in March 2026, with on-demand capacity sold out across GPU types at that time (SemiAnalysis).

How much does it cost to buy a B200 server instead?

Reported prices for an 8-GPU HGX B200 server vary widely, and many listings are not real stock. American Compute, which publishes GPU server values and runs a residual-value insurance program, gives $360,000 to $470,000 per new node in its October 2026 Rack Report. GPU Per Hour recorded Supermicro systems at about $384,000 to $394,000 on Newegg on September 21, both out of stock, and a Lenovo usual customer sale price of $550,475 as of June 15. Each system also needs a real power budget: NVIDIA lists the DGX B200 at about 14.3kW maximum. Our HGX B200 buyer's guide covers configuration, and AI server pricing covers budget ranges.

At what utilization does buying beat renting?

The simplest test is to divide your owned cost per GPU-hour by the rental rate. The result is the share of hours the GPUs must be busy to break even. American Compute models a $415,000 B200 node at $2.77 per GPU-hour over three years, including colocation at $280 per kW per month for 14kW, management fees and a 31% resale value. It excludes financing, insurance, support contracts and sales tax, and a single node needs no cluster network. On that model, break-even against on-demand rates of $7.69 to $8.60 is 32% to 36% utilization (our arithmetic). Because that firm forecasts three-year resale at 16% to 46% of the new price, owned cost could run from about $2.47 to $3.07, or 29% to 40% against those same rates.

Utilization stops mattering against a reserved contract, because you pay for every hour under either option. There, compare owned cost directly with the contract rate. CoreWeave advertises up to 60% off on-demand for committed usage, and 60% off its $8.60 B200 rate would be about $3.44, though that is a maximum and the terms are not public. American Compute's own three-year contract case uses $4.00. Against those figures, owning saves roughly $0.67 to $1.23 per GPU-hour, before the financing, insurance, support and networking that the model leaves out. Run the same test with your own numbers, using our AI cluster TCO model.

When does renting from a neocloud win?

• The work is short: a training run, a pilot or a burst of a few weeks.
• Demand is uncertain, and you cannot yet show steady utilization above break-even.
• You have nowhere to host 14kW per server. Our guide to the colocation shortage explains why that may take time to fix.
• You need capacity sooner than you could receive and install hardware, assuming the provider truly has it available. Our GPU shortage guide covers both routes.
• You want the newest GPUs without carrying residual value risk. American Compute forecasts that a B200 bought in 2026 could keep only 16% to 46% of its price after three years.

When does owning GPU servers win?

• Utilization stays above break-even for most of the hardware's life, typically for production inference or constant internal use.
• Data location or regulation rules out shared infrastructure.
• You need price certainty. A purchase fixes the cost, while rental renewals follow the market.
• You can keep the hardware past three years. See our guide to GPU depreciation and service life.
• You already have power, cooling and staff, which removes the largest barriers.

Many teams do both, owning enough for steady baseload and renting for peaks. That usually gives the best utilization on owned hardware, and it keeps a second supply route open.

What contract terms should you check with a neocloud?

• Prepayment: SemiAnalysis said in April 2026 that providers can demand higher prepay and longer terms. Ask how prepaid funds are treated if the provider fails or is sold.
• Term and renewal: SemiAnalysis reported H100 contracts being renewed at the same rate as 2 to 3 years earlier, some for 4 years through 2028. Fix the renewal price or get an exit right.
• Start date: the same report says providers may pick start and end dates to suit their inventory. Add a remedy for late delivery.
• Minimum size: SemiAnalysis said some large neoclouds no longer sell single nodes, and Lambda's cluster pricing starts at 16 GPUs.
• Take-or-pay: confirm whether you pay for every hour, and what happens during outages or GPU failures.
• Price changes: find out whether the rate can move during the term. On-demand rates can move without a contract.
• Extras: price storage, networking and data transfer. CoreWeave advertises no ingress, egress or transfer fees on its storage, while Nebius lists $0.015 per GiB for standard object storage egress.
• Exit: confirm data return, and what happens on assignment or a change of control.

What happens if a neocloud goes out of business?

Provider risk belongs in the model. Gartner predicted on October 14, 2026 that more than 80% of neocloud providers will cease independent operation by 2030, citing cost pressure and a lack of differentiated business models. That is a forecast, and it does not mean they all fail: CRN Asia notes Gartner did not say how many it expects to be acquired or merged and how many to shut down. For a buyer, the practical risk is a change of owner, terms or service in the middle of a contract. Keep data portable, avoid large prepayments to thinly capitalized providers, and keep a second source.

Nexus Compute supplies new B200, H200 and H100 servers through authorized distribution channels with the full manufacturer warranty, and offers purchase orders and net terms for qualified organizations. Send your expected utilization, term and hosting plan through our quote form, or email sales@nexus-compute.com, and we will reply within 48 business hours. If renting is the better answer for your numbers, we will say so.