There is no single cheapest cloud for large-model training. RunPod and DigitalOcean show the lowest published per-GPU rates in this comparison, while CoreWeave publishes attractive eight-GPU node pricing and AWS, Google Cloud and Oracle Cloud Infrastructure offer the networking, identity, storage and scheduling integrations needed for very large jobs. Once a training run spans multiple machines, interconnect speed, guaranteed capacity, storage movement and interruption risk can matter more than the hourly GPU figure.
The comparison below normalizes the published figures available on October 13, 2026, explains what each provider actually sells, and gives a practical method for choosing a platform without mistaking a single-GPU price for the cost of a distributed training run.
As an Amazon Associate I earn from qualifying purchases.
What matters more than the headline GPU price
GPU memory and numerical support
Start with the model, not the provider. Large language models may require more than the memory on one accelerator, and optimizer states, gradients, activations and checkpoint buffers add to the requirement. AWS P5 instances, for example, can provide eight H100 GPUs with up to 640 GB of aggregate HBM3 memory. H100, H200, B200, A100 and AMD MI300X systems also differ in memory capacity, bandwidth and support for FP8, BF16, FP16 and other precisions. A lower hourly rate is irrelevant if the chosen GPU forces smaller batches, excessive sharding or host-memory offload.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Intra-node and inter-node fabric
Distributed training repeatedly exchanges gradients and activations. NVLink or NVSwitch handles traffic inside a server; the inter-node fabric determines how well a cluster scales. AWS P5, P5e and P5en instances provide up to 3,200 Gbps of Elastic Fabric Adapter networking, and AWS describes UltraClusters scaling to as many as 20,000 H100 or H200 GPUs. Google Cloud A3 Mega machines provide eight H100 SXM GPUs and up to 1,800 Gbps of network bandwidth. Those figures are provider specifications, not an independent benchmark, but they show why a cheap isolated GPU can lose to a more expensive tightly connected node.
Capacity, billing and total run cost
Compare the complete run: GPU or node rental, host charges, boot disks, high-performance file storage, object-storage requests, data transfer, checkpoint retention, support and any software license. Identify whether a quote is on-demand, Spot, reserved, a capacity block, a committed-use discount or a promotional figure. Spot capacity can be interrupted, and a discount is only valuable if checkpoints let the job resume without losing more time than the saving is worth.
Published provider comparison
| Provider | Training hardware and scale | Published US pricing checked October 13, 2026 | Best fit | Important watch-outs |
|---|---|---|---|---|
| AWS EC2 | P5, P5e and P5en with one or eight H100/H200 GPUs; up to 3,200 Gbps EFA; UltraClusters up to 20,000 GPUs | No single comparable hourly rate on the reviewed product page; regional on-demand, Capacity Block and other quotes vary | Teams already using S3, FSx for Lustre, IAM, EFA, SageMaker or HyperPod; very large distributed jobs | Obtain a current regional quote and confirm capacity before committing |
| Google Cloud | A3 High and Mega with H100, A3 Ultra with H200, and A2 with A100; A3 Mega has eight H100 SXM GPUs and up to 1,800 Gbps | GPU, machine and commitment prices are region-specific; Spot pricing is dynamic | GKE or Slurm workflows, regional placement, committed use and Spot options | Include the machine charge, region, storage and actual GPU availability |
| CoreWeave | Eight-GPU HGX H100 and H200 nodes | H100: $49.24 per node-hour on demand or $19.71 Spot. H200: $50.44 on demand or $20.93 Spot | Transparent GPU-first pricing and full HGX nodes | These are eight-GPU node prices; normalize them before comparing with a one-GPU Pod. Check region, storage and capacity |
| Lambda | H100, H200 and B200 systems; interconnected clusters advertised from 16 to more than 2,000 GPUs | B200 advertised from $6.69; the page does not establish one universally comparable SKU or availability | Fast GPU-focused provisioning and large interconnected clusters | Verify the exact GPU, reservation terms, region and queue time |
| RunPod | H100 PCIe and SXM, A100 and more than 30 GPU models; self-serve Pods | H100 PCIe $2.89 per GPU-hour; H100 SXM $3.99 per GPU-hour; billed by the second | Short experiments, development and jobs that benefit from no-contract, per-second billing | Confirm multi-node fabric, sustained capacity, service-level terms and checkpoint behavior for long runs |
| DigitalOcean | HGX H100 and H200 droplets with one or eight GPUs, NVMe scratch and a transfer allowance | H100 $4.41 per GPU-hour; H200 $4.47 per GPU-hour; 12-month reserved H100 $3.26 per GPU-hour | Simple VM operations and predictable single-node workflows | Evidence for hyperscale training fabric and scheduler integration is less extensive; verify cluster topology |
| Oracle Cloud Infrastructure | H100, H200, B200 and AMD MI300X bare metal and VM options; RDMA networking; Superclusters up to 131,072 B200 GPUs | The product page gives historical pricing context based on June 5, 2024, rather than a current comparable H100 rate | Enterprise bare metal, RDMA, very large Superclusters and an AMD alternative | Request a current, region-specific quote and validate availability |
How to interpret each provider
AWS: strongest ecosystem and extreme scale
AWS is compelling when training data already sits in S3 or when the organization depends on IAM, FSx for Lustre, EFA, NVSwitch, SageMaker or HyperPod. P5-family instances can be configured with one or eight H100/H200 GPUs, and the documented EFA bandwidth reaches 3,200 Gbps. UltraCluster designs are intended for jobs that need thousands of accelerators rather than a handful of rentable GPUs.
The trade-off is price complexity. On-demand, Capacity Block and regional capacity are separate purchasing decisions, so a public product page does not provide a fair universal hourly number. Ask for the exact region, instance shape, commitment period, storage path and expected allocation before comparing AWS with a specialist cloud.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Clear out junk files and repair common Windows errors3Scan for outdated or missing drivers - takes under a minuteGoogle Cloud: strong networking with regional and orchestration choices
Google’s A3 family covers H100 and H200 configurations, while A2 provides A100 options. A3 Mega combines eight H100 SXM GPUs with up to 1,800 Gbps of network bandwidth and is positioned for large-scale training and serving. Compute Engine, GKE and Slurm paths let teams choose between VM-oriented operations, Kubernetes and batch scheduling.
Google pricing is split among GPU, machine, region and purchasing model. Committed-use discounts can improve steady workloads, while Spot rates change with supply. Include all attached machine and storage costs, and confirm that the selected zone can deliver the required GPU count at the time the run is scheduled.
Rank #2
- Part number 900-53651-2500-000 and model: P3651
- This is the 2 slot version for when there is no empty slots between 2 slot cards. If you have one or more empty slots between the cards or the cards are 3 slot this NVLink will not work. See the attached images showing the card layout.
- NVLink 3.0 for any brand of RTX Ampere model graphics cards: 3090, A30, A40, A100 / H100 (Requires three NVLinks), A800, A4500, A5000, A5500, A6000
- This is the same as PNY part number: NVLAMP-2SLOT-BSP and RTXA6000NVLINK-KIT
- This is the same as Dell part number: 0RWJ7Y
CoreWeave: transparent HGX node economics
CoreWeave’s published H100 and H200 figures are for complete eight-GPU HGX nodes. Dividing the on-demand prices by eight gives approximately $6.16 per H100 GPU-hour and $6.31 per H200 GPU-hour; the corresponding Spot figures are about $2.46 and $2.62. Those normalized values exclude storage, networking and any other charge, and they do not turn the node into an independently schedulable single GPU.
HGX topology is useful when the workload needs all eight accelerators and fast intra-node communication. For multi-node training, ask which fabric is available between nodes, how many nodes can be reserved together, and what happens when Spot capacity is reclaimed.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallOutdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchLambda: cluster-focused specialist access
Lambda advertises interconnected H100, H200 and B200 clusters from 16 to more than 2,000 GPUs. Its page also advertises B200 capacity from $6.69, but the exact SKU, region, billing model and availability must be confirmed before using that figure in a budget. The main attraction is a GPU-first procurement path when a team does not need a broad hyperscaler platform.
RunPod: low-friction Pods and fine-grained billing
RunPod lists H100 PCIe at $2.89 per GPU-hour and H100 SXM at $3.99, with billing by the second: the meter starts when a Pod is running and stops when it is terminated. That model is efficient for bursty development, preprocessing and short training jobs because idle minutes do not round up to a full hour.
For a serious distributed run, the inexpensive single-Pod rate is only the first check. Confirm the inter-node network, placement of all workers, maximum run duration, image and driver compatibility, persistent checkpoint storage and the provider’s capacity guarantee.
Rank #3
- Video/Sound Cards
- Passive Cooling
DigitalOcean: straightforward droplets with published GPU rates
DigitalOcean publishes H100 at $4.41 per GPU-hour and H200 at $4.47, with a 12-month reserved H100 figure of $3.26 per GPU-hour. Its HGX droplets are offered with one or eight GPUs and include NVMe scratch space plus a transfer allowance, which can simplify a conventional VM workflow.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →The published figures should not be treated as evidence of hyperscale collective-communication performance. If the model requires many nodes, obtain the topology and scheduler details and price the data path between the training set, checkpoints and workers.
Oracle Cloud Infrastructure: enterprise bare metal and AMD choice
Oracle documents bare-metal and VM options for H100, H200, B200 and AMD MI300X accelerators, with RDMA networking and Supercluster designs advertised up to 131,072 B200 GPUs. This makes it a distinct option for organizations that require bare-metal control, Oracle enterprise integration or an AMD deployment path.
The cited product page uses a June 5, 2024 pricing basis, so it cannot serve as a current apples-to-apples H100 quote. Request a fresh regional proposal that states the GPU count, host shape, RDMA configuration, storage and commitment terms.
Normalize prices before choosing
Use one unit throughout the comparison: GPU-hour for isolated accelerators, node-hour for complete servers, or the total cost of a completed training run. Never divide a node price by eight and then assume the resulting number buys one freely deployable GPU; it is only a normalization aid.
Rank #4
- CUDA Cores: 4608 / NVIDIA Tensor Cores: 576 / NVIDIA RT Cores: 72
- GPU Memory: 24 GB GDDR6 with ECC / Bandwidth: 624 GB/Sec
- System Interface: PCI Express 3.0 x16
- Four DisplayPort 1.4 Connectors
- 3D Stereo Support with Stereo Connector
| Published offer | Original unit | Approximate normalized GPU-hour | What the number does not include |
|---|---|---|---|
| CoreWeave H100 on demand | $49.24 per eight-GPU node-hour | $6.16 | Storage, networking, host overhead and the requirement to rent the full HGX node |
| CoreWeave H100 Spot | $19.71 per eight-GPU node-hour | $2.46 | Interruption risk and non-compute charges |
| CoreWeave H200 on demand | $50.44 per eight-GPU node-hour | $6.31 | Storage, networking and full-node scheduling constraints |
| CoreWeave H200 Spot | $20.93 per eight-GPU node-hour | $2.62 | Interruption risk and non-compute charges |
| RunPod H100 PCIe | $2.89 per GPU-hour | $2.89 | Multi-node fabric, persistent storage and any additional Pod resources |
| RunPod H100 SXM | $3.99 per GPU-hour | $3.99 | Multi-node fabric, persistent storage and any additional Pod resources |
| DigitalOcean H100 | $4.41 per GPU-hour | $4.41 | Other droplet, storage, transfer and orchestration charges |
| DigitalOcean H200 | $4.47 per GPU-hour | $4.47 | Other droplet, storage, transfer and orchestration charges |
For a simple budget, multiply the number of GPUs by wall-clock hours and the normalized rate, then add storage, transfer and failed-run overhead. For example, 100 hours on eight RunPod H100 SXM GPUs at the published rate is 8 × 100 × $3.99, or $3,192 before other charges. An eight-GPU CoreWeave H100 node for 100 on-demand hours is 100 × $49.24, or $4,924, while its published Spot figure would be $1,971 for the same nominal time if capacity remained available. These are arithmetic illustrations, not performance predictions.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.A repeatable selection process
- Describe the workload. Record model parameter count, sequence length, target precision, memory estimate, expected GPU count, checkpoint interval and maximum acceptable interruption.
- Choose the topology. Decide whether one GPU, one eight-GPU node or multiple nodes is required. Ask for NVLink or NVSwitch details inside a node and the inter-node fabric between nodes.
- Normalize the quote. Convert node, VM, reservation and Spot offers to GPU-hour or completed-run cost. Keep the original unit beside the converted figure.
- Add non-GPU costs. Price the input data path, high-performance filesystem, object storage, checkpoint retention, boot volumes, egress and support.
- Validate capacity. Get written confirmation of region, GPU model, quantity, start date, reservation duration and replacement policy if capacity disappears.
- Test the actual software stack. Verify CUDA or ROCm versions, NCCL or RCCL behavior, container images, drivers, Slurm or Kubernetes integration and checkpoint restore.
- Run a short acceptance job. Measure tokens per second, scaling efficiency, data-loader throughput and checkpoint time on the exact topology. Treat provider specifications as capabilities, not as a substitute for your workload’s measurements.
Which provider fits common situations?
- Data already in AWS and a multi-thousand-GPU target: begin with P5-family capacity and EFA-based UltraCluster options, then compare the complete regional quote with alternatives.
- Data already in Google Cloud or a GKE/Slurm operation: evaluate A3 Mega or A3 Ultra and include the machine, zone and commitment terms.
- One to eight GPUs for experiments or bursty development: compare RunPod’s per-second Pods with DigitalOcean’s straightforward droplets.
- Eight-GPU HGX training nodes: CoreWeave’s node pricing is easy to model, provided the job can use a full node and the required capacity is reserved.
- More than 16 GPUs without building a hyperscaler platform: ask Lambda or another specialist provider for an interconnected-cluster quote that states fabric, placement and support.
- Bare metal, RDMA or AMD MI300X requirements: include Oracle Cloud Infrastructure in the request-for-quote process and insist on current regional pricing.
Common mistakes and how to avoid them
Comparing unlike units
A full eight-GPU HGX node, an isolated GPU-hour and a VM with attached host resources are not interchangeable. Preserve the original unit and topology in every spreadsheet row.
Ignoring failed or interrupted work
Spot savings disappear when a job restarts from an old checkpoint. Set checkpoint frequency from the cost of recomputation, test resume behavior and keep a reliable copy outside an ephemeral boot disk.
Pricing only the accelerator
Large datasets and frequent checkpoints can make filesystem and transfer charges material. Place data near the GPUs where policy permits, and include egress when results must leave the region.
Assuming advertised scale is guaranteed
A documented maximum cluster size is not a promise that the exact count is available in your chosen region on your start date. Treat capacity, lead time and reservation terms as procurement requirements.
Choosing a GPU without testing the model
H100, H200, B200, A100 and MI300X differ in memory and software behavior. Use a reproducible pilot on the intended container and precision, then record throughput and scaling before signing a long commitment.
Bottom line
Use RunPod or DigitalOcean as starting points for small, price-sensitive jobs, CoreWeave when an eight-GPU HGX node and transparent specialist pricing match the workload, and Lambda when you need a GPU-focused interconnected cluster. Choose AWS or Google Cloud when ecosystem integration, data locality, orchestration and very large fabrics outweigh a lower headline rate. Include Oracle Cloud Infrastructure when bare metal, RDMA, AMD accelerators or Supercluster scale are strategic requirements. In every case, make the final decision from a normalized completed-run cost and a measured pilot on the exact topology—not from a single advertised GPU-hour.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
