Top cloud GPUs compared

About Top cloud GPUs compared

Cloud GPU comparison is awkward because the headline rate hides the parts that decide the bill: how the hour is sliced, which accelerator the rate buys, and how far the fleet is spread. This ranking covers 280 brands that rent GPU capacity, and the central trade-off is entry price against hardware and reach. In this category the lowest published GPU hourly rates start at €0.01 and reach €349.91, the median entry rate is €0.61 per hour, and a quarter of the field sits at €0.39. Two habits pay off. First, separate the introductory offer from the rate that follows: bonus credits, sign-up allowances and promo codes soften the opening bill, while the hourly or per-second figure is what keeps applying once they are spent. Second, read the rate against the hardware. A low number on a consumer card is a different job from the same number on a datacenter part with an interconnect behind it. Granularity is the third lever. Hourly billing covers 50.9% of this ranking, with per-second at 26.7% and per-minute at 22.4%, so short test runs cost less where the clock is finer. Region count matters next: the median brand here runs 6 locations, and where your data may sit is a procurement question as much as a latency one.

How we picked

Scores here rest on three weighted inputs, and the weights are published rather than impressionistic. Cost carries 40%, offer carries 40% and trust carries 20%. In practice the entry rate and what a plan includes decide most of the order, while certifications, location spread, support channels and the entity behind the contract settle the remaining ties. Cost is measured as the lowest GPU hourly rate a brand publishes, with billing granularity recorded alongside it, so a per-second rate is not read as an hourly one. Offer covers the accelerator list, the interconnect between GPUs, free credits and free tiers. Trust covers certifications, the contracting entity, the home country and the support channels offered. Figures come from each brand's own public pages: pricing pages for rates and granularity, product pages for the accelerator list and interconnect, security pages for certifications, and terms pages for the entity and support. Category medians and quartiles are calculated across the 280 brands in this ranking, not a handpicked subset. Partners are listed first where a ranking has them, then every other brand follows in score order, and the label on a listing reflects placement only.

Our top picks

1. 4. SberCloud: best for top-pick

   SberCloud holds the top pick in this ranking because it pairs the widest support spread here with datacenter hardware and a certification page that goes deep. GPU hours start at RUB 240 on hourly billing, and the fleet covers B300, H100 and A100, from the B300 down to the V100. Support runs on email, phone and chat, a set no other member here matches, and multi-GPU work is backed by Infiniband on the machine learning service depending on configuration, NVLink and Direct Connect. The compliance page is long (ISO/IEC 27001:2022, ISO/IEC 27701:2019 and ISO/IEC 27017:2015) and the fleet spans 9 locations. Two things to weigh: the operation is based in RU, and sign-up credits are paid in roubles. The member ranked next, ComputePrices.com, quotes a lower headline rate but tracks other providers rather than running the machines.

   • Price per GB per month: RUB 814.68
   • Egress price per GB: RUB 1.15
   • Free egress: 10 TB of outgoing traffic per month

   Pros

   • Outgoing traffic includes 10 TB of outgoing traffic per month each month, so routine transfers cost nothing before RUB 1.15 per GB applies.
   • A free tier plus bonus credits for individuals, legal entities and sole proprietors keeps early testing cheap.
   • Evolution compute starts at RUB 23.75 with 1 vCPU, 1 GB RAM, 30 GB NVMe, billed hourly.
   • GPU capacity covers B300, H100, A100, V100 and A40 from RUB 240 an hour, billed per minute.

   Cons

   • Kubernetes has no free control plane, so every cluster adds RUB 2.98 on top of the node price.
   • Infrastructure sits in RU across 9 locations, so there is no reach outside that market.
   • Storage covers object only, so block and file tiers are not part of the same plan.

2. 5. TensorDock

   TensorDock closes this ranking on price rather than capability. GPU hours start at $0.10 on per second billing, at the high end of the entry rates in this group, and that buys a datacenter list: H100 SXM5 80GB, A100 SXM4 80GB and A100 PCIe 80GB, including the H100 and the A100 in SXM and PCIe form. Reach is the argument for it, with a network counted at from 100 locations, and the egress fee line reads as no ingress or egress charges, which matters when a job writes checkpoints all night. Connectivity is modest next to UCloud, since 1 Gbps minimum port included, while the certification set (SOC 2, PCI DSS and ISO 27001) is the broadest among the marketplaces here and support runs on email, help centre and community.

   • GPU models: H100 SXM5 80GB, A100 SXM4 80GB, A100 PCIe 80GB, and 7 more
   • GPU hour from: $0.10
   • Billing granularity: Per second

   Pros

   • per second billing means an idle or unfinished test costs only what it used
   • No ingress/egress fees, so pulling checkpoints and datasets out does not add a surprise line
   • GPU hours start at $0.10, below the market median of €0.63
   • SOC 2, PCI DSS and ISO 27001 cover regulated workloads without a separate audit
   • from 100 with a 1 Gbps minimum port included attached

   Cons

   • Per-second billing suits bursts, but a long steady training run bills on the same meter with no committed-use rate listed under hourly
   • Support runs on email, help centre and community only, with no live phone channel for incidents
   • from 100 is broad but plan names are not given per region, so pinning a workload to a single market needs care

3. 7. Nosana

   Nosana is fifth in this ranking for the widest spread of consumer and datacenter cards at a middle price. The trade-off against the two marketplaces above it is the hourly rate, which sits above $0.015 and $0.02 per hour. UCloud, ranked next, prices in a similar band and adds a published interconnect specification, while Nosana lists the longer card catalogue.

   • GPU models: NVIDIA 3060, NVIDIA 4060, NVIDIA 3070, and 18 more
   • GPU hour from: $0.048
   • Free credits: Free credits when signing up with email

   Pros

   • The card list in NVIDIA 3060, NVIDIA 4060 and NVIDIA 3070 covers consumer, workstation and datacenter parts, including the A100 and the H100.
   • A GPU hour from $0.048 puts entry well under the ranked median of €0.61.
   • Signing up with an email address brings free credits, which lowers the cost of a first test run.

   Cons

   • Billing is limited to hourly, so there is no reserved or monthly rate to lock in for steady training.
   • Support runs on email, help centre and community only, with no phone or live chat path listed.

4. 8. Salad

   Salad is third in this ranking for one reason: short jobs. Billing is per second, the rate starts at $0.015 per hour, and the machines sit in a network counted at from 190 locations, so a run that ends early costs minutes rather than a whole hour. The catalogue, NVIDIA RTX 3060, NVIDIA RTX 3070 and NVIDIA RTX 3070 Ti, is built from RTX consumer and professional parts, which suits inference, rendering and single-card fine-tuning. The ceiling is scale: interconnect is limited to single-GPU nodes, so multi-GPU NVLink training is not part of what this service offers, and the compliance page carries a single certification entry (SOC 2 Type I). PoolCompute, ranked next, charges a little more per hour but restricts its fleet to datacenter parts, listed as NVIDIA Tensor Core H100 SXM, NVIDIA Tensor Core H100 NVL and NVIDIA Tensor Core H100 PCIe.

   • GPU models: NVIDIA RTX 3060, NVIDIA RTX 3070, NVIDIA RTX 3070 Ti, and 6 more
   • GPU hour from: $0.015
   • Billing granularity: Per second

   Pros

   • Billing is per second against an hourly model, so short runs and interrupted jobs pay only for the time they hold a card.
   • Workloads sit in from 190 locations, well above the 6 median in the GPU ranking.
   • The catalogue in NVIDIA RTX 3060, NVIDIA RTX 3070 and NVIDIA RTX 3070 Ti spans several RTX generations, the RTX PRO workstation class and selected AMD classes.
   • The AI gateway is OpenAI-compatible, so clients already written against that API can be repointed at Salad.
   • The operator is Salad Technologies, Inc., a US company founded in 2018, with SOC 2 Type I listed for security.

   Cons

   • One GPU per node: None - single-GPU nodes, no multi-GPU NVLink interconnect, so any plan that needs several accelerators inside a single machine has to look elsewhere.
   • Bring your own key is not offered, and the gateway is billed as Flat monthly subscription or pay per token, so you cannot pay through your own provider account.
   • Support runs on email, help centre and community only, with no live chat or phone line listed for urgent issues.
   • A free tier is not offered, and only 27% of the GPU ranking sits in the same paid-only position.

What to look for

The entry rate per GPU hour

The lowest published rate is the number most comparisons lead with. The median in this ranking is €0.61 per hour, the lower quartile €0.39 and the upper quartile €1.67, so a rate well under the median is a genuine differentiator. The full spread runs from €0.01 to €349.91.

How finely the hour is sliced

Granularity decides what a short job costs. Hourly billing covers 50.9% of this ranking, per-second billing 26.7% and per-minute billing 22.4%. Where the clock is finer, a run that ends early costs proportionally less.

What the rate actually buys

A consumer card, a workstation part and a datacenter accelerator are not the same purchase. Hourly plans cover 88.5% of this ranking and a free tier 27.9%, so the cheapest figure may describe a trial shape rather than a production one. Read the accelerator list line by line before comparing rates.

Where the fleet runs

Region count drives latency and where data can sit. The median brand here runs 6 locations, the lower quartile 3 and the upper quartile 14, with the widest fleets reaching 3,000.

Support and the other charges

A stalled job costs more than any rate difference. Help-centre coverage stands at 76.9% of this ranking, email at 78.9%, community forums at 55.4% and phone at 36.3%. Add egress, where the median is €0.01 per gigabyte and the top of the range €0.22, plus multi-GPU nodes where the median starting price is €9.15.

Frequently asked questions

Top cloud GPUs compared by country

• Brazil
• Canada
• China
• France
• Germany
• India
• Indonesia
• Italy
• Japan
• Mexico
• South Korea
• Spain
• United Kingdom
• United States
• Vietnam

Providers compared head to head

• Green Mountain / Norway sovereign hosting (SovereignSky) vs HIVE Digital Technologies (BUZZ HPC)
• Green Mountain / Norway sovereign hosting (SovereignSky) vs Neural Internet (Nodexo)
• Green Mountain / Norway sovereign hosting (SovereignSky) vs NHN Cloud
• HIVE Digital Technologies (BUZZ HPC) vs Neural Internet (Nodexo)
• HIVE Digital Technologies (BUZZ HPC) vs NHN Cloud
• Neural Internet (Nodexo) vs NHN Cloud