GPU cloud providers: how to choose the best one for your workload

The checklist at a glance

1. Define the workload
2. Pick the GPU and the form factor
3. Check the interconnect
4. Check storage and networking
5. Confirm availability, not list price
6. Read the commitment and contract terms
7. Match the orchestration to your team
8. Look at reliability signals
9. Test support before you need it
10. Check security and compliance
11. Plan the exit and data egress

Facts and prices below were checked on 2026-11-09 and link to the provider's own page. Everything marked as our view is editorial. For definitions, keep the glossary open.

1. Define the workload

Write down four numbers before you look at any provider: GPU count, hours per week, how long the project runs, and how much data moves in and out. Then put the job in one of three buckets.

• Single-GPU or single-node inference and experiments. Any reputable on-demand cloud works. Optimize for price, per-second billing and quick start.
• Single-node fine-tuning. You need 8 GPUs with NVLink in one server and fast local or shared storage for checkpoints.
• Multi-node training. You need a cluster: SXM GPUs, an RDMA fabric between nodes, a parallel file system, a scheduler and a provider that replaces failed nodes fast. This bucket is where providers differ the most.

2. Pick the GPU and the form factor

Our H100 vs H200 vs B200 comparison covers which generation to rent. Form factor matters as much as the model number.

• SXM modules sit on an HGX board with NVLink between all 8 GPUs. NVIDIA lists the H100 SXM at up to 700 W with 900 GB/s of NVLink.
• PCIe cards (sold as "H100 PCIe" or "H100 NVL") run at 350 to 400 W with 600 GB/s of NVLink through a bridge and less compute: 3,341 TFLOPS of sparse FP8 against 3,958 for SXM (NVIDIA H100).

PCIe is cheaper. On 2026-11-09 Hyperstack listed H100 SXM at $3.20 and PCIe at $2.50 per GPU-hour, and Runpod listed $3.99 and $2.89 (Hyperstack, Runpod). For inference on one GPU, PCIe is often the better deal. For training across GPUs, buy SXM. When a listing says only "H100" or "H100 80GB", ask which one it is. Massed Compute, for example, lists an "H100 80GB" at $2.73 without stating the form factor.

3. Check the interconnect

Inside a node, GPUs talk over NVLink. Between nodes, they need an RDMA network: InfiniBand, or RoCE (RDMA over Converged Ethernet) (NVIDIA InfiniBand, IBTA). For multi-node training, ask three things.

• Which fabric and how fast per node? Examples of what providers state: CoreWeave runs NVIDIA Quantum InfiniBand with SHARP and more than 3.2 Tbps per HGX H100 node. Runpod Instant Clusters use InfiniBand or RoCE v2 at 1,600 to 3,200 Gbps per node. Latitude.sh B300 servers have 800 Gbps dual-plane RoCE (CoreWeave, Runpod, Latitude.sh).
• Is the fabric on this product? Hyperstack's on-demand VMs use SR-IOV networking up to 350 Gbps, while its Secure Private Cloud uses InfiniBand (Hyperstack). The same provider can sell both, at different prices.
• Is it non-blocking and rail-optimized? Ask for the topology diagram. Then measure it yourself with nccl-tests (all_reduce_perf) across all nodes.

4. Check storage and networking

Slow storage leaves expensive GPUs idle while they wait for data or checkpoints. Ask what file system backs the cluster, its measured read and write throughput, and whether locking works across nodes. Third-party tests find real differences here. SemiAnalysis measured Vultr's Lustre tier at 91.7 GB/s, and found that cross-node file locks were not enforced on Verda's shared file system (ClusterMAX 3.0).

Ask how storage is billed. GMI Cloud says networking and storage are billed separately from GPUs. Firmus's SMC lists file storage at $0.065 per GB-month and object storage at $0.025 (GMI Cloud, SMC). Ask too whether stopped instances still bill. DigitalOcean's page says powered-off GPU Droplets are still billed (DigitalOcean).

5. Confirm availability, not list price

A price without capacity is not an offer. When we checked on 2026-11-09:

• Shadeform's public feed listed 304 instance types across 18 clouds, and only 65 showed current availability. Its cheapest H100 SXM, H200 and B200 listings had none (Shadeform).
• Most of Vultr's top-end GPU plans showed zero deployable locations in its public plans API.
• "From" and "starting at" H100 claims of $1.50 (Hydra Host), $1.68 (Latitude.sh) and $1.99 (Voltage Park) were not shown as available offers, so we leave them out of our price tables.

Our view: ask sales for a dated, written quote that names the region, the start date and the number of GPUs they can deliver on that date. Compare quotes on our GPU cloud pricing page, where every price carries the date we saw it.

6. Read the commitment and contract terms

On-demand rarely has a minimum. Reserved capacity always does. Published terms on 2026-11-09:

Our view on what to read in the contract: when the term starts (delivery or signature), what happens if nodes fail during the term, whether unused time can be paused or resold, and the payment schedule. Prepaid terms move risk to you. The GPU cluster rental guide goes deeper on quotes and negotiation.

7. Match the orchestration to your team

Slurm is an open source cluster manager and job scheduler for Linux clusters. Kubernetes is an open source platform for containerized workloads and services. Our view: Slurm fits batch training jobs and researchers who think in jobs and nodes, and Kubernetes fits teams that already run services on it. Several neoclouds run Slurm on top of Kubernetes, so you get both: CoreWeave sells SUNK, Nebius sells Soperator and Verda uses Slinky. Others offer only one. DigitalOcean's managed option is Kubernetes (DOKS), and marketplaces like Vast.ai give you a CLI and API. Denvr's pricing page says Kubernetes and Slurm are self-managed on CPU VMs.

Our view: if your team has never run Slurm, a managed Slurm with health checks built in is worth more than a lower hourly rate. Ask whether the provider's Slurm supports container tooling your jobs use. SemiAnalysis noted Verda's Slinky lacked Pyxis and Enroot when tested (ClusterMAX 3.0).

8. Look at reliability signals

GPUs fail. What matters is how fast the provider detects it and swaps the node. Three sources help.

• Third-party testing. SemiAnalysis's ClusterMAX 3.0 (September 2026) rates GPU clouds from Platinum to Bronze, plus Participation and Not recommended categories (ratings). The write-up explains each grade. It moved Crusoe from Gold to Bronze over XID errors on an H100 cluster, an old kernel with a CVE on a B300 cluster, and slow remediation. It downgraded Together AI to Bronze after health checks drained nodes mid-run. On DigitalOcean, its first injected XID error went undetected (ClusterMAX 3.0). These are SemiAnalysis's ratings, not ours.
• Incidents and status. Check the provider's status page history and recent news. When we checked Genesis Cloud on 2026-11-09, its site returned no response or HTTP 522.
• Your own acceptance test. Our view: before you accept a cluster, run nccl-tests across all nodes, a GPU burn test, and a short run of your real job. Watch the logs for Xid errors. Agree in the contract on what counts as a failed node and how fast it is replaced.

9. Test support before you need it

Our view: open a ticket during the trial with a real technical question, at a bad time of day for the provider's time zone. Note the response time and whether an engineer or a script answers. Ask whether you get a shared Slack channel, who is on call, and whether support covers the fabric and storage or only the VM. Write the response times you were promised into the order form.

10. Check security and compliance

Ask for the reports, not just the logos. A SOC 2 report is a CPA examination of a service organization's controls (AICPA). Check whether it is Type I or Type II, the period it covers, and whether the GPU data centers are in scope. Examples of what providers state: Nebius lists SOC 2 Type II, ISO 27001 and HIPAA among others. Runpod lists SOC 2 Type II, SOC 3 and ISO 27001:2022. Crusoe lists ISO 27001 and ISO 42001. Vast.ai states SOC 2 Type I (Nebius, Runpod, Crusoe, Vast.ai).

On a marketplace, your workload may run on a third-party host's hardware. Find out who the host is and which certifications apply to that host, not to the marketplace.

11. Plan the exit and data egress

Moving checkpoints and datasets out can cost real money and time. Some providers make it explicit. Cirrascale states no ingress or egress fees. Latitude.sh includes 20 TB of free egress per server. CoreWeave has a "zero-egress migration program" page aimed at customers switching in (Cirrascale, Latitude.sh). Where egress isn't stated, ask.

Our view: also weigh counterparty risk on long terms. CoreWeave's 10-Q, for example, shows $35.551 billion of debt principal and one customer at 36% of quarterly revenue (10-Q). That does not make it a bad choice. It means your exit plan should exist on day one: images in a registry you control, checkpoints synced to your own object storage, and infrastructure as code. Our neocloud market overview covers financing across the sector.

What to ask sales

1. Which exact GPU SKU and form factor, and how many can you deliver, in which region, on which date?
2. What is the inter-node fabric, speed per node and topology? Can we run nccl-tests during a trial?
3. What storage backs the cluster, and what read and write throughput has it measured?
4. Is the price per GPU-hour all-in? What do storage, egress, public IPs and support cost?
5. What is the minimum term, the payment schedule and the start trigger?
6. How do you detect and replace failed nodes, how fast, and do we get credits?
7. Which Slurm or Kubernetes setup is managed by you, and which by us?
8. Can we see your SOC 2 Type II report and data center certifications?
9. Who owns the hardware and the data center? Is any of this capacity resold from another provider?

Red flags

• A price with no inventory. See step 5.
• No published price and no written quote. Several large providers sell through sales only, which is normal. Refusing to put a quote in writing is not.
• Resale stacked on resale. SemiAnalysis described Sesterce as "a broker on top of a broker" and said Mithril is a marketplace wrapping three Nebius availability zones (ClusterMAX 3.0). Resale is not wrong in itself, but you should know whose hardware you are on and who answers when it breaks.
• "Not recommended: Unavailable" in ClusterMAX. SemiAnalysis uses it for providers that claimed zero spare capacity or would not be tested (ClusterMAX 3.0).
• Lookalike sites. Hyperstack's pricing page warned on 2026-11-09 that a fraudulent site, hyperstack.my, was impersonating it. Check the domain before you pay.
• No acceptance period. If you cannot test the cluster before the term starts, price that risk in.

Mid-size GPU clouds

These providers sell self-serve instances and smaller clusters, and most publish their prices.

Next steps: compare hardware in H100 vs H200 vs B200, check today's rates on the pricing page, or read how to rent a GPU cluster if you need more than a few nodes. Profiles: Nebius, CoreWeave, Hyperstack, Runpod.