Start with the accelerator list
Hardware is the first filter, because a provider is only useful if it offers the card your workload needs. The lists differ sharply. Amazon covers a short, mainstream set with NVIDIA A10G, NVIDIA T4 and NVIDIA H100. Aethir offers B200, H200 and H100, a data centre line built around current generation parts and sold as bare metal, no virtualization overhead.
Baseten spans NVIDIA T4, NVIDIA L4 and NVIDIA A10G, which includes a fractional H100 option, so a full card and a slice of one can sit in the same comparison. Cerebrium reaches further still: NVIDIA B200, NVIDIA H200 and NVIDIA H100 goes beyond NVIDIA cards to AWS accelerators, which matters if you want to test an alternative backend before committing.
Some providers deliberately step outside the NVIDIA default. Cirrascale covers AMD Instinct MI325X, AMD Instinct MI300X and AMD Instinct MI250, mixing AMD Instinct parts, Qualcomm Cloud AI hardware and systems from other specialist vendors. AutoDL runs the opposite way, with a long tail of consumer and regional cards in Tesla P40, TITAN Xp and GTX 1080 Ti, while Biren Technology offers its own silicon in BR100, BR104 and Bili 166M. The Cloud GPUs ranking places these lists next to each other, which is the fastest way to see who covers what.
Read the entry plan before the price list
The smallest plan shows how a provider prices the basics, and it often reveals more than a headline rate. AceCloud's entry plan is 1 vCPU, 4 GB RAM (S3a.medium), a CPU and memory configuration rather than a GPU machine, which suits orchestration and pipeline work that does not need a card.
Bitdeer's entry plan is 1x NVIDIA H100 GPU virtual machine, on-demand, placing one current generation card on on-demand billing, so you can start a single machine without a long commitment. Cirrascale's entry plan is 1x Qualcomm AI 100 Pro, 12 vCPU, 48 GB RAM, 1 TB NVMe, pairing an accelerator with vCPU, memory and NVMe storage in one listing, which makes it easy to see whether storage is included in the machine you are quoted for.
Two habits pay off here. First, check whether the entry plan includes storage and how much memory comes with the card, since memory often decides whether a model fits at all. Second, keep the entry plan separate from the per accelerator rate: the AceCloud review and the Bitdeer review show how plans are described in practice, and the differences between them are easier to judge side by side than one at a time.
Regions and location counts
Footprint is the second filter, and it splits into two questions: is your region on offer, and how wide is the spread. AceCloud runs 3 locations with regions covering IN and US. Bitdeer has 7 locations across NO and US. Boost Run reports 3 locations in US, and Cassava Technologies (Kasi Cloud) lists 5 locations in NG.
Broader providers work differently. Cerebrium supports multi-region deployment across 12 locations, and Amazon's infrastructure spans 39 regions and availability zones. Akash Network describes a decentralised network of from 85 locations rather than fixed data centre regions, so capacity appears where providers offer it.
Read a location count as a footprint, not a capacity guarantee. The region you need for data residency or low latency may not be the one with the newest cards, so check both before you commit. The Amazon review is a useful contrast with the smaller regional providers, and the Akash Network review covers the decentralised model in more depth.
Interconnect decides whether multi-node training is practical
For any job spanning several machines, the interconnect matters more than the per card price. Where numbers are published, you can compare them directly. Bitdeer documents InfiniBand (ConnectX-7), inter-instance connectivity up to 800 Gbps. Cirrascale offers InfiniBand up to 3200Gb on multi-node clusters, and CINECA runs Dual-rail Mellanox EDR InfiniBand across 6 locations. Baseten uses High-speed InfiniBand between nodes in multi-node deployments. Cassava Technologies (Kasi Cloud) works with carrier-neutral interconnected data centres.
Not every provider publishes a figure. Akash Network points to training clusters that span several nodes, with Ray clusters built in. AutoDL points to NVLink inside its A100 SXM4 and H20-NVLink zones, and to parallel work spread over several nodes and cards. Anyscale takes a different route and spreads work over several clouds at once, naming AWS, GCP, Azure, Nebius or CoreWeave as possible targets.
The practical rule: if your job is single-machine, interconnect barely matters and you can optimise on price and region. If it is distributed, treat the interconnect description as a requirement to confirm before you sign up, and ask the provider what you actually need. The Baseten review and the AutoDL review cover two very different answers to that question.
A repeatable comparison routine
Run the comparison in a fixed order so that one factor does not quietly decide the others. Name the workload first: training, inference or a scheduled job. Training pulls you toward high-memory cards and multi-node interconnect, inference often fits a smaller card or a fractional slice, and scheduled work suits on-demand entry plans.
Then check the accelerator list against that workload, the region list against residency and latency, and the interconnect description if more than one machine is involved. Only after those three pass should you read prices, because a low rate on the wrong card or in the wrong region is not a saving.
Keep notes in one place. The Aethir review, the Beam Cloud review and the Anyscale review show three different answers to the same hardware and footprint questions, and the Cloud GPUs ranking lets you weigh price against hardware in a single pass.