GPU cloud: providers, prices and the power behind them

What a GPU cloud is

A GPU cloud rents graphics processors, and the servers, networks and storage around them, to customers who need AI compute but do not want to buy and run the hardware. The customer pays by the GPU-hour, or signs a contract for a fixed block of capacity, and gets remote access to machines in someone else's data center. Training a large model, fine-tuning a smaller one and serving answers to users (inference) all run on this kind of capacity.

The unit of sale is usually one GPU for one hour. In practice, AI GPUs are sold in groups of eight: a standard NVIDIA HGX server carries eight GPUs linked by NVLink, and most clouds bill the whole server. Larger jobs rent many such servers joined by a fast fabric such as InfiniBand or high-speed Ethernet. The newest rack-scale systems, NVIDIA's GB200 and GB300 NVL72, join 72 GPUs into a single NVLink domain inside one rack.

What separates a GPU cloud from a generic virtual machine is everything that keeps those GPUs busy: the interconnect between servers, storage that can feed data fast enough, schedulers such as Slurm or Kubernetes, and monitoring that catches a failing GPU before it stalls a training run that spans thousands of them. A cheap hourly rate on a cluster that loses a day to a bad node is not cheap.

This guide covers who sells GPU capacity, what it costs, how renting compares with owning, and where the power behind it comes from. For the buildings themselves, see our pillar guide to AI data centers.

Hyperscalers and neoclouds

Two kinds of seller dominate. The hyperscalers, Amazon Web Services, Microsoft Azure, Google Cloud and Oracle, sell GPUs as one product among hundreds, next to databases, storage and identity services that many enterprises already use. Their GPU prices tend to be high on demand and much lower under commitments, and the main draw is integration with an account the customer already runs.

The second group are specialists, often called neoclouds, that exist mainly to rent GPU capacity. They grew quickly because they could stand up NVIDIA's newest systems fast and sell them at lower list prices. The larger ones, as reported in their filings and announcements:

CoreWeave

The largest listed neocloud. Second-quarter 2026 revenue was about $2.58 billion, up from $1.21 billion a year earlier, with a revenue backlog of about $104 billion at June 30, 2026. Active power grew by nearly 500 MW in the quarter to 1.5 GW, with about 3.7 GW contracted.

Nebius

Listed on Nasdaq. Second-quarter 2026 revenue was $582.3 million, up 454% year on year, and annualised run-rate revenue reached $3.0 billion at the end of June 2026. The company expects to end 2026 with 5 GW of contracted power and has delivered all capacity tranches under its Microsoft contract.

Lambda

Founded in 2012 and privately held. It raised $1.5 billion in a Series E in November 2025, signed a multi-billion-dollar GPU contract with Microsoft in the fourth quarter of 2025 and, as of January 2026, was reported to be targeting an IPO in the second half of 2026. Its website listed 15 US data centers at the time.

Crusoe

Builds its own data centers and runs Crusoe Cloud on them. Its October 2025 Series E raised $1.375 billion at an expected valuation above $10 billion. It reported a power pipeline of more than 45 GW, a 1.2 GW campus in Abilene, Texas, and a 1.8 GW campus planned in Wyoming. Crusoe sold its Bitcoin mining business to NYDIG to focus on AI.

Together AI

Rents NVIDIA GPU clusters and hosts open-source models. It raised an $800 million Series C at an $8.3 billion valuation on July 1, 2026, led by Aramco Ventures, and reported annual bookings of more than $1.15 billion. Together AI is itself a Crusoe Cloud customer.

The line between the groups blurs. Hyperscalers buy capacity from neoclouds: Microsoft has contracts with Nebius, Lambda and IREN. And many neoclouds lease their buildings from third parties, including former Bitcoin miners, rather than owning them. Marketplaces and decentralised networks add a third tier that aggregates spare GPUs from smaller operators; for one model, read Flux's case for decentralised compute.

What GPUs rent for in 2025 and 2026

GPU rental prices are posted publicly by most neoclouds and all hyperscalers, but they are rarely comparable at a glance. Some quote per GPU, some per 8-GPU server, and some per rack-scale instance. Everything below has been converted to dollars per GPU-hour, on demand, before tax, from each provider's public price page as listed in October 2026. Prices change without notice; check the page before you plan around a figure.

The GPU generations on offer

NVIDIA A100

The Ampere-generation part, still rented widely for smaller training jobs and inference. 40 GB and 80 GB versions.

NVIDIA H100

Hopper, with 80 GB of HBM3 per GPU in its SXM form. The reference GPU for most rental indices.

NVIDIA H200

Hopper with 141 GB of HBM3e, which lets larger models fit on fewer GPUs.

NVIDIA B200

Blackwell, with 180 GB of HBM3e and configurable up to 1 kW per GPU, per NVIDIA's reference architecture. An HGX B200 server carries 1.4 TB of GPU memory.

NVIDIA B300

Blackwell Ultra. An HGX B300 server carries 2.1 TB of GPU memory and NVIDIA rates it at 108 petaflops of dense FP4, against 72 for HGX B200.

GB200 and GB300 NVL72

Rack-scale systems pairing Grace CPUs with 72 Blackwell GPUs in one liquid-cooled rack drawing about 120 kW, according to NVIDIA.

Next is Vera Rubin. NVIDIA said on May 31, 2026, that the platform was ramping into full production, with production shipments set to begin in the fall of 2026. How quickly Rubin capacity reaches rental price lists, and what it does to Blackwell and Hopper rates, is the next thing to watch in the indices below.

Neocloud list prices, October 2026

H100

Lambda $3.99 (H100 SXM instance). Nebius $4.50. CoreWeave about $6.16 ($49.24 per hour for an 8-GPU HGX H100 node).

H200

Nebius $5.40. CoreWeave about $6.31 ($50.44 per 8-GPU node).

B200

Lambda $6.69 (instance). Nebius $8.50. CoreWeave about $8.60 ($68.80 per 8-GPU node).

B300

Nebius $9.50. CoreWeave lists HGX B300 as contact-sales.

GB200 NVL72

CoreWeave about $10.50 ($42.00 per hour for a 4-GPU slice). Nebius lists it as contact-sales.

A100 80 GB

Lambda $2.79. CoreWeave about $2.70 ($21.60 per 8-GPU node).

L40S

Nebius from $1.55. CoreWeave about $2.25.

Lambda also sells 1-Click Clusters of HGX B200 for terms of two weeks to a year, at $9.86 per GPU-hour for 16 GPUs, falling to $8.87 for 256 or more.

Hyperscaler prices

Hyperscaler on-demand rates run much higher. Google Cloud's a3-highgpu-8g VM, eight H100s, lists at $88.49 an hour on demand, about $11.06 per GPU-hour; its H200-based a3-ultragpu-8g lists at about $10.60 per GPU-hour. AWS cut on-demand prices for its H100-based P5 instances by up to 44% from June 1, 2025, and for A100-based P4 instances by up to 33%. AWS Capacity Blocks, which reserve GPUs for a fixed window, were priced in October 2026 in US East at about $5.97 per GPU-hour for p5 (H100), $7.90 for p5en (H200), $14.21 for p6-b200 and $16.15 for p6-b300, with the GB200 UltraServer in the Dallas Local Zone at about $10.58. AWS says those rates move with supply and demand and are next scheduled to update in January 2027.

What the indices show

List prices are a ceiling. What GPUs actually clear at is tracked by Silicon Data, which in May 2025 launched a daily H100 rental index, published on Bloomberg under the ticker SDH100RT and built from millions of price points across neoclouds and marketplaces. Its readings tell a story of recovery:

• In August 2025 the H100 index held between $2.22 and $2.26 per GPU-hour, dipping to $2.13 in mid-September 2025.
• Through March 2026 it traded between about $2.43 and $2.63.
• On October 18, 2026, it stood at $2.83, up 7.6% over seven days.
• The B200 index opened 2026 at $4.40, jumped to a high of $6.11 on March 25, 2026, around NVIDIA's GTC conference, and closed March at $5.48. The index page showed $6.26 in mid-October 2026, with the H200 at $3.51 and the A100 at $1.53.

Silicon Data publishes a separate hyperscaler H100 reading. Between March 1 and April 20, 2026, it held between $7.40 and $7.52, roughly three times the neocloud level, and barely moved; the firm describes it as an administered ceiling above a large book of reserved capacity rather than a cleared market. On May 12, 2026, CME Group and Silicon Data announced plans to launch compute futures based on these indices later in 2026, subject to regulatory review. If they launch, buyers and sellers of GPU time will be able to hedge price risk much as power buyers hedge electricity.

On-demand, reserved or owned

There are three ways to get GPU capacity, and most serious buyers use more than one.

On demand

Pay by the hour, stop when done. The most flexible and the most expensive, and availability is not guaranteed: on-demand inventory of the newest GPUs is often sold out. Spot or preemptible capacity is cheaper again but can be reclaimed; Google's spot price for an A3 High H100 VM was about $6.60 per GPU-hour in October 2026, against $11.06 on demand.

Reserved or committed

Commit to a block of GPUs for weeks, months or years in exchange for a lower rate and guaranteed access. CoreWeave advertises discounts of up to 60% off on-demand for committed usage; Nebius up to 35% for multi-month cluster reservations. Google's three-year committed price for the A3 High VM is about $4.86 per GPU-hour, and its one-year price about $7.67. Short fixed blocks, such as AWS Capacity Blocks or Lambda's two-week minimum clusters, sit in between.

Owning

Buy servers and place them in your own or a colocation data center. Lowest cost per hour at high utilisation, but the buyer carries the hardware, the power contract, the cooling and the risk that the next generation makes the fleet less valuable.

The ownership arithmetic

The decision turns on utilisation and useful life. An owned GPU costs the same whether it runs or sits idle, so its effective hourly cost rises sharply when utilisation falls. A rented GPU costs nothing when released. Owners also have to choose a depreciation period. CoreWeave lengthened the useful life of its technology equipment, GPUs included, from five years to six at the start of 2023; critics, including investor Michael Burry in November 2025, argue that AI GPUs lose economic value faster than that. The rental indices above are one way to test the assumption: a GPU's resale and rental value three years out is what determines whether owning paid off.

Power is the other half of ownership cost. NVIDIA put the running cost of a data center at $1.9 million to $2.8 million per megawatt per year in April 2025, including nearly $500,000 for cooling energy and water. A buyer comparing a three-year reservation with a purchase should price in power, colocation, networking, spares and staff, not just the server invoice. aPanel lists power and running costs for individual systems such as the NVIDIA DGX B200 and the AMD MI300X.

A rough rule follows from the prices above: steady, predictable workloads that will keep GPUs busy for years favour reservations or ownership; bursty work, experiments and anything needing the very newest generation favour on-demand or short blocks. Many teams reserve a base load and burst on demand above it.

The power behind the capacity

A GPU cloud is a power business wearing a software interface. The binding constraint on new capacity in 2025 and 2026 has been energized sites, not GPUs. The IEA reported in April 2026 that data center electricity demand grew 17% in 2025, that AI-focused sites grew faster, and that slow grid connections and tight supplies of gas turbines and transformers were delaying projects. It expects data center electricity use to roughly double by 2030 and AI-focused data center use to roughly triple.

The neoclouds now describe themselves in megawatts as much as in GPUs. CoreWeave reported 1.5 GW active and about 3.7 GW contracted at mid-2026. Nebius expects 5 GW of contracted power by the end of 2026 and plans to deploy more than 1 GW a year from 2027. Crusoe reports a pipeline above 45 GW. For a buyer, this matters because a provider's ability to deliver a large cluster on time depends on power it has already secured, not on GPUs it has ordered.

Density adds a second constraint. A GB200 NVL72 rack draws about 120 kW and needs liquid cooling to the chip, against roughly 20 kW for racks in earlier data centers, according to NVIDIA. Older halls cannot simply be refilled with Blackwell racks, which is one reason providers are building or leasing new campuses rather than retrofitting. The cooling side is covered in designing for the AI cooling crisis and modular approaches in Mod42's modular AI data centers.

Bitcoin miners as GPU cloud hosts

Bitcoin miners own what GPU clouds most need: large, energized, interconnected sites, often in West Texas and other places where power is available. Since 2024 a string of miners has turned that into AI revenue. Most act as landlords, leasing powered and cooled space to a GPU cloud that installs and runs its own hardware. A few run GPU clouds themselves.

Miners as landlords

• Core Scientific contracted about 590 MW of critical IT load across six sites to CoreWeave on 12-year terms, worth about $10.2 billion in total, after adding a 70 MW expansion at Denton, Texas, in February 2025. Read how Core Scientific and CoreWeave built physical AI infrastructure.
• Galaxy contracted all 526 MW of critical IT load planned at its Helios campus in West Texas, a former Bitcoin mine, to CoreWeave on 15-year leases, and delivered the first 133 MW in July 2026. The leases are expected to generate more than $1 billion a year in revenue on average.
• Hut 8 signed a 15-year, $7 billion lease with Fluidstack in December 2025 for 245 MW of IT capacity at River Bend, Louisiana, serving Anthropic, with Google backstopping the lease payments. Renewal options could lift the total to about $17.7 billion. Completion is scheduled for the second quarter of 2027.
• Cipher Mining signed a 15-year, roughly $5.5 billion lease with AWS for 300 MW in November 2025, with rent starting in August 2026, after a $3 billion, 168 MW lease with Fluidstack, backed by Google, at Barber Lake, Texas.
• TeraWulf signed two 10-year agreements with Fluidstack in August 2025 for more than 200 MW of critical IT load at Lake Mariner, New York, worth about $3.7 billion, with Google backstopping $1.8 billion of the lease obligations and taking warrants for about 8% of the company.

Miners as cloud operators

IREN went further and became a GPU cloud. In November 2025 it signed a five-year, $9.7 billion cloud services contract with Microsoft for NVIDIA GB300 NVL72 capacity at Childress, Texas. On August 13, 2026, it reported that Horizon 1, the first of four 50 MW liquid-cooled deployments scheduled for 2026, had been delivered and accepted, and that NVIDIA had designated it an Exemplar Cloud on GB300 NVL72. IREN targets 480 MW of AI cloud capacity in 2026 and 1.2 GW in 2027. Crusoe took the opposite path, selling its mining business to concentrate on its cloud.

The landlord and operator models carry different risks. A landlord collects rent on a long lease, often with a large-company backstop, and leaves GPU pricing risk with the tenant. An operator owns the GPUs and keeps the margin, but also the depreciation and the exposure to rental prices. For the strategic debate inside mining companies, read Fog Hashing on the miner pivot to AI and HIVE's dual-engine strategy.

How to choose a GPU cloud provider

No provider is best for every job, and this guide does not recommend one. The checklist below is what experienced buyers compare before they sign.

The exact GPU and configuration

SXM or PCIe, memory per GPU, and whether the price is per GPU, per server or per rack-scale slice. An H100 PCIe is not an H100 SXM, and a 4-GPU GB200 slice is not an 8-GPU HGX node.

All-in price

Compare per GPU-hour after converting units. Add storage, egress, public IPs, CPU and memory if billed separately, support tiers and tax. Ask what discount a commitment of your size earns.

Availability and lead time

Can the provider deliver the number of GPUs you need on the date you need them, and is that capacity already powered? Ask for site, megawatts energized and delivery schedule in writing.

Interconnect and storage

For multi-node training, confirm the fabric (InfiniBand or high-speed Ethernet), bandwidth per GPU and the storage throughput behind the cluster.

Reliability terms

Uptime commitment, how fast a failed GPU or node is replaced, whether you get credits, and whether health checks run before a job lands on a node.

Contract flexibility

Minimum term, notice periods, the right to swap to a newer generation, and what happens if capacity is delivered late.

Software and access

Bare metal or virtual machines, managed Kubernetes or Slurm, images and drivers, and whether your team can bring its own tooling.

Location and compliance

Data residency, certifications your customers require, and latency to users if you are serving inference.

Counterparty strength

Who owns the building and the power, how the provider is financed, and whether a larger company stands behind the lease. Public filings answer much of this for listed providers.

Price benchmarks

Check quoted rates against a public index such as Silicon Data's before committing, and recheck at renewal.

The pattern across 2025 and 2026 is that power, not chips, sets who can deliver. Providers that secured energized sites early, often from miners, are the ones signing the largest contracts. That convergence of power owners and compute buyers is the subject of Intelligence Disrupt 2027, March 22-24, 2027, at the Irving Convention Center at Las Colinas in Dallas-Fort Worth, co-located with Mining Disrupt, the Bitcoin mining expo. See passes or exhibit.

Questions

What is a GPU cloud?

A service that rents GPU servers by the hour or under a term contract, so a customer can train or run AI models without buying the hardware. Hyperscalers such as AWS and Google Cloud sell GPUs alongside many other services; neoclouds such as CoreWeave, Lambda and Nebius specialise in GPU compute.

How much does it cost to rent an H100 per hour?

It depends heavily on the seller. In October 2026 Lambda listed an H100 SXM instance at $3.99 per GPU-hour, Nebius at $4.50 and CoreWeave about $6.16, while Google Cloud's A3 High VM worked out to about $11.06 per GPU-hour on demand. Silicon Data's neocloud H100 index read $2.83 on October 18, 2026.

How much does it cost to rent a B200?

Listed on-demand prices in October 2026 ranged from $6.69 per GPU-hour at Lambda to $8.50 at Nebius and about $8.60 at CoreWeave. AWS Capacity Blocks for an 8-GPU p6-b200 instance in US East came to about $14.21 per GPU-hour.

Is reserved GPU capacity cheaper than on-demand?

Usually. CoreWeave advertises discounts of up to 60% for committed usage and Nebius up to 35% for multi-month cluster reservations. Google Cloud's three-year committed price for an A3 High H100 VM is about 56% below its on-demand rate.

What is a neocloud?

A cloud provider built mainly to rent GPU capacity for AI, rather than a general-purpose cloud. The largest, CoreWeave, reported second-quarter 2026 revenue of about $2.6 billion and 1.5 GW of active power.

Why do Bitcoin miners host GPU clouds?

They already control powered sites, which are the scarcest input. Core Scientific, Galaxy, Hut 8, Cipher and TeraWulf have leased capacity to CoreWeave, Fluidstack or AWS, and IREN runs its own GPU cloud under a five-year, $9.7 billion contract with Microsoft.

Can GPU rental prices be hedged?

CME Group and Silicon Data announced on May 12, 2026, that they plan to launch compute futures based on Silicon Data's daily GPU rental indices, subject to regulatory review.

More guides

• AI data centers: power, cooling, sites and the build-out
• Liquid cooling for data centers
• Data center power: how much it takes and how to get it

Sources

1. Lambda, GPU cloud pricing · 2026-10
2. CoreWeave, cloud pricing · 2026-10
3. Nebius, NVIDIA GPU pricing · 2026-10
4. Google Cloud, accelerator-optimized VM pricing · 2026-10
5. AWS, EC2 Capacity Blocks for ML pricing · 2026-10
6. AWS, up to 45% price reduction for EC2 NVIDIA GPU instances · 2025-06-05
7. Silicon Data, H100 Silicon Index · 2026-10-18
8. Silicon Data, H100 rental market update, September 2025 · 2025-10-20
9. Silicon Data, B200 rental price, March 2026 update · 2026-03-30
10. Silicon Data, H100 Hyperscaler Index, April 2026 · 2026-04-27
11. Silicon Data launches the H100 Rental Index · 2025-05-20
12. CME Group and Silicon Data to launch compute futures · 2026-05-12
13. CoreWeave, second-quarter 2026 results · 2026-08-11
14. Nebius Group, Q2 2026 shareholder letter (Form 6-K) · 2026-08-12
15. DCD, Lambda in talks to raise $350m in pre-IPO funding · 2026-01-09
16. Crusoe, Series E funding announcement · 2025-10-24
17. TechCrunch, Together AI raises $800M · 2026-07-01
18. NVIDIA, HGX platform · 2026
19. NVIDIA, HGX AI Factory reference architecture: components · 2026
20. NVIDIA, Blackwell platform and liquid cooling · 2025-04-22
21. NVIDIA, Vera Rubin ramps into full production · 2026-05-31
22. IEA, Data centre electricity use surged in 2025 · 2026-04-16
23. Nasdaq / Motley Fool, depreciation and CoreWeave · 2025-11-11
24. Core Scientific and CoreWeave, Denton expansion (SEC exhibit) · 2025-02-26
25. IREN delivers Horizon 1 to Microsoft · 2026-08-13
26. The Block, Galaxy delivers 133 MW to CoreWeave · 2026-07-07
27. The Block, Hut 8 signs $7 billion lease with Fluidstack · 2025-12-17
28. DCD, Hut 8 signs 245 MW deal with Fluidstack · 2025-12-17
29. DCD, AWS signs 300 MW deal with Cipher Mining · 2025-11-03
30. TeraWulf, Fluidstack agreements (Form 8-K exhibit) · 2025-08-14