Price: start with the hourly floor, then check the same card
GPU capacity is quoted per accelerator per hour, and the floors sit far apart. Across the listings in our Cloud GPUs rankings, GPU hours run €0.01 to €349.91. ComputePrices.com advertises hardware from $0.015 on an hourly model, while SberCloud lists accelerators from RUB 240 with a per hour billing model.
A wide gap between floors tells you the market is segmented, not that one provider is a bargain. The rate that matters is the one for the exact card your job needs, so put the same model on both sides before drawing a conclusion. If a provider is cheaper at the floor and lacks the part you need, the floor tells you nothing about your bill.
Billing granularity decides what a short job costs
Rate cards mean little until you know how the clock runs. Salad bills per second, so an interrupted job stops accumulating cost almost immediately. SberCloud bills compute per minute on its GPU capacity, and its Evolution compute pricing is hourly. ComputePrices.com bills hourly and also lists a free tier model alongside, while SberCloud's listed model is hourly.
For a run that occupies a machine overnight, granularity changes little. For bursty work such as evaluations, rendering passes or many short agents, per second billing can matter more than a lower headline rate. Pick the billing model that matches the shape of your usage, then compare rates on that basis.
Hardware lists show whether the cheap tier is your tier
Read the catalogue rather than the price. ComputePrices.com publishes a wide range, spanning datacenter parts such as H100 SXM, A100 SXM, A100 PCIE, H200, B200, HGX B300, GB200, GB300 and MI300X, through L4, L40, L40S, A10, RTX PRO 6000, RTX 6000 Ada, RTX A6000, RTX 5090, RTX 4090 and RTX A4000, down to a GTX 1650. SberCloud's GPU line is narrower: B300, H100, A100, V100 and A40.
A consumer card and a datacenter card are different products, not the same product at two prices. The bottom of a long list usually serves single-user inference and hobby workloads, while multi-user training needs the top. Match the class first, then compare per hour within that class.
Free tier, credits and quotas
A free tier lowers the cost of trying, and quotas decide whether it helps. SberCloud has a free tier and offers free credits of 4,000 bonus RUB for individuals, 20,000 bonus RUB for legal entities and sole proprietors, with 10 TB of outgoing traffic per month of outgoing traffic included and 1,000,000 free function requests. Its control plane is not free, so budget RUB 2.98 for managed Kubernetes. The entry plan is 1 vCPU, 1 GB RAM, 30 GB NVMe at RUB 23.75, storage is priced at RUB 814.68 and outgoing traffic at RUB 1.15. ComputePrices.com structures its free tier around API access, with 750 requests/day available.
Read allowances next to rates. A low rate paired with a small request allowance can cost more than a higher rate with room to spare, and included traffic or requests only help if your workload stays inside them.
Networking and adjacent features that decide practicality
Distributed training is a networking problem before it is a price problem. SberCloud lists Infiniband for its ML service depending on configuration, plus NVLink and Direct Connect, so a multi node job has a documented path between cards and between machines. Check the interconnect first when your framework spans several nodes.
Supporting services matter too. SberCloud offers virtual private server and bare-metal server types, object storage with S3-compatible access, and DDoS protection, which lets existing code and tooling run without rework. Salad does not support bring your own key, so workloads that depend on customer managed encryption need another arrangement.
Reliability, regions and the counterparty
Before you benchmark, check the promise, the footprint and the contract. SberCloud commits to 99.9% uptime, operates across 9 locations, lists RU as its region and is headquartered in RU, with support through email, phone and chat. ComputePrices.com is run by Lansky Tech and offers email and help centre.
Certification lists are the next filter. SberCloud publishes a broad set of certifications, ISO/IEC 27001:2022, ISO/IEC 27701:2019 and ISO/IEC 27017:2015, spanning information security, cloud security, personal data and payment card handling along with data centre design, facility and operation. Salad publishes SOC 2 Type I. Region choice is a hard filter for some teams, and a single market narrows the shortlist further.
Work through the provider pages side by side
Take each shortlisted provider to its write-up before committing: SberCloud, ComputePrices.com and Salad. Check whether the pricing model, hardware list and uptime commitment there match what you priced.
Return to the Cloud GPUs rankings to widen the set whenever a requirement filters everyone out, be it a specific card, a region or a billing granularity. The right provider is the one that clears your filters on price and hardware, then holds up on quota and reliability.