Start with the billing clock
Serverless and edge platforms do not bill the same way, and the clock decides what a small job costs. Salad charges at per second, so a function that sits idle between events pays for nothing. SberCloud bills GPU capacity at per minute and lists hourly as its billing model, so a comparable workload carries a larger minimum charge. The gap is widest for bursty, event-driven work: webhook handlers, queue consumers, image transforms, anything that waits between requests.
Check that difference against the Serverless and edge compute rankings before you model a monthly bill. Then ask whether a provider's clock fits your traffic shape. Steady traffic suits a longer billing window, while spiky traffic pays off with fine granularity, and a plan that looks cheap per hour can lose once idle time is counted.
Separate the entry price from the running total
A headline price rarely covers the whole bill. SberCloud's entry plan is 1 vCPU, 1 GB RAM, 30 GB NVMe at RUB 23.75, and the managed Kubernetes control plane is billed on its own at RUB 2.98. A free control plane is not included, so a cluster carries that cost from its first hour of its life.
Network and storage sit outside that too. SberCloud lists RUB 1.15 per GB of outgoing traffic with 10 TB of outgoing traffic per month included, and RUB 814.68 for storage covering object storage that is S3-compatible. Add both lines to your own traffic estimate before you compare totals between providers.
Salad prices along a different axis: Flat monthly subscription or pay per token on its AI gateway. A flat monthly subscription suits predictable volume, while per-token billing suits traffic you cannot forecast. There is no free tier on Salad, while SberCloud has a free tier and grants bonus credit, with a larger amount for legal entities and sole proprietors than for individuals. Run both shapes through your own numbers rather than picking a winner by instinct.
Work the per-request maths
Function pricing is easiest to compare per request. SberCloud bills serverless calls at RUB 14.34 and RUB 0.0013 per GB second, with 1,000,000 free requests included. Multiply your monthly call volume by the request-block rate, then check whether the free allowance absorbs it. If the answer is yes, the effective rate drops sharply for a while, so the allowance is a real term of the comparison rather than a footnote.
Give the numbers scale before you decide. Across the listings we carry, free request allowances run 1,000 to 10,000,000, request-block prices run €0 to €4,384.81, and per-GB-second prices run €0. A provider sitting in the middle of those spans is often the sane starting point for a first production workload, while the ends of each range suit very different traffic shapes and commitments.
Let the GPU line pick the workload
Accelerated work narrows the field quickly. Salad's hourly GPU rate is $0.015, and the catalogue runs NVIDIA RTX 3060, NVIDIA RTX 3070 and NVIDIA RTX 3070 Ti, which points at inference, batch jobs and fine-tuning rather than tightly coupled training. SberCloud's GPU rate starts at RUB 240 with B300, H100 and A100 behind it, and its interconnect covers Infiniband (ML service, depending on configuration), NVLink, Direct Connect. On Salad the picture is None - single-GPU nodes, no multi-GPU NVLink interconnect, so a job that needs cards to talk to each other will not fit there at all.
Prices at the two ends are not comparable without that hardware context, and the currency difference makes the raw numbers misleading on their own. If distributed GPU capacity matters more than a single fast card, read the kluster.ai review and the Nosana review next to these two before settling.
Check where the work runs
Latency claims mean little without a location list. SberCloud reports RU as its region, 9 locations, and a headquarters in RU. It also offers virtual private servers and bare-metal machines through vps and bare metal, which matters when a function has to sit near a database rather than near a user. A region count is only useful if your traffic is in that region.
Data residency can settle the question before price does. SberCloud's certification list covers ISO/IEC 27001:2022, ISO/IEC 27701:2019 and ISO/IEC 27017:2015, and its legal entity is LLC Cloud Technologies. Salad is headquartered in US and founded in 2018, with a shorter certificate list at SOC 2 Type I and no bring-your-own-key option. For another angle on the same question, the PoolCompute review takes a different view of where workloads should sit. Confirm the paperwork your own compliance needs before you commit.
Reliability and support
Uptime is a contract question, not a marketing one. SberCloud publishes 99.9% as its availability commitment and adds DDoS protection, so the SberCloud review is the place to read how that commitment is written and what sits behind it. On the Salad side, treat the Salad review as where you confirm what is actually promised, because distributed consumer-class hardware and a single-region provider need different checks.
Support reach matters most for small teams. SberCloud offers email, phone and chat. Weigh that against your own working hours: an email-only path slows an incident down, and a phone and chat combination shortens it. Record which figures are contractual and which are best-effort, because that distinction is what you will be held to in the middle of an incident.
A short routine for the final decision
Score every candidate in the same order. Fix the billing granularity against your traffic shape, price the entry plan and the control plane separately, add egress and storage, then test the request-block rate against the free allowance. Only after that do you look at GPU models, interconnect, regions and certificates, because those answers narrow the list rather than rank it.
Finish with a like-for-like reliability check and a support check, and write down which numbers you would rely on in a contract. Start from the Serverless and edge compute rankings for the full field, then read the SberCloud and Salad pages before you put a shortlist in front of a team.