Free tools Windows power users keep installed
One-click scans. No signup required.
iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
There is no single cheapest or best GPU cloud for every large-model job. RunPod and DigitalOcean publish low single-GPU rates, while CoreWeave publishes attractive eight-GPU node pricing. AWS, Google Cloud, and Oracle Cloud Infrastructure can cost more or require a quote, but their high-speed fabrics, storage, identity, scheduling, and capacity options can shorten a distributed training run or simplify operating it. Compare the complete training run—not just the advertised GPU-hour.
What matters most when choosing a GPU cloud
Large-model training is a distributed-systems problem. The GPU is only one part of the bill and one part of performance. Once a job spans multiple GPUs, communication between accelerators, access to training data, checkpoint writes, scheduler behavior, and the likelihood of getting the requested capacity can matter as much as the nominal hourly rate.
- GPU memory and precision: H100, H200, B200, A100, and MI300X configurations differ in memory capacity and supported numerical formats. Check whether the model, optimizer states, activations, and batch size fit without excessive sharding or offload.
- Intra-node topology: NVLink and NVSwitch determine how quickly GPUs in one server exchange tensors. A nominally identical GPU can deliver a different training result when it is attached through a weaker host or PCIe path.
- Inter-node fabric: Ethernet bandwidth, RDMA, EFA, and the provider’s collective-communication stack affect all-reduce and parameter-sharding time. AWS documents up to 3,200 Gbps of EFA networking for P5, P5e, and P5en; Google documents up to 1,800 Gbps for A3 Mega.
- Capacity and placement: A low price is not useful if eight or 128 GPUs cannot be allocated together. Ask about guaranteed capacity, queue time, reservations, and the exact region.
- All-in cost: Include the host or node, storage, snapshots, data transfer, public IPs, orchestration, and idle time while a job is starting or checkpointing.
- Operations: IAM, private networking, logging, managed Kubernetes or Slurm paths, checkpoint storage, support response, and data-residency controls can outweigh a small rate difference.
Provider documentation describes capabilities and list prices; it does not constitute an independent performance benchmark. For a serious purchase, run the same short, reproducible training segment on the shortlisted configurations.
Published prices are not directly comparable
The table below uses US figures checked on October 13, 2026. Prices and availability change, and several entries are region-, machine-, commitment-, or capacity-dependent. A “GPU-hour” means one GPU running for one hour; a node-hour can contain several GPUs and must be divided before comparing it with a single-GPU offer.
#1 Best Overall
- AI Performance: 767 AI TOPS
- OC mode: 2632 MHz (OC mode)/ 2602 MHz (Default mode)
- Powered by the NVIDIA Blackwell architecture and DLSS 4
- Axial-tech fan design features a smaller fan hub that facilitates longer blades and a barrier ring that increases downward air pressure
- A 2.5-slot design maximizes compatibility and cooling efficiency for superior performance in small chassis
| Provider | Training hardware and scale | Published US price checked 2026-10-13 | What the figure means | Important qualification |
|---|---|---|---|---|
| AWS EC2 P5/P5e/P5en | Instances with one or eight H100/H200 GPUs; UltraClusters up to 20,000 H100/H200 GPUs | No single comparable hourly rate on the public product page | Quote requires the instance, region, and purchase model | On-demand, Capacity Block, and regional prices vary |
| Google Cloud A3/A2 | A3 High/Mega H100, A3 Ultra H200, A2 A100; A3 Mega has eight H100 GPUs | Region-specific GPU and commitment tables; Spot is dynamic | Machine, GPU, and commitment charges must be combined | Include regional availability and all machine charges |
| CoreWeave | Eight-GPU HGX H100 and H200 nodes | H100: $49.24/node-hour on demand; $19.71/node-hour Spot. H200: $50.44/node-hour on demand; $20.93/node-hour Spot | Derived GPU-hour: H100 about $6.16 on demand or $2.46 Spot; H200 about $6.31 on demand or $2.62 Spot | Derived figures assume all eight GPUs are used; storage, region, and capacity are separate considerations |
| Lambda | H100, H200, and B200; advertised interconnected clusters from 16 to more than 2,000 GPUs | B200 advertised from $6.69 | Headline starting price; exact SKU and billing unit must be confirmed | Verify region, reservation, availability, and whether the rate is for the required configuration |
| RunPod | H100 PCIe/SXM, A100, and more than 30 GPU models | H100 PCIe $2.89/GPU-hour; H100 SXM $3.99/GPU-hour | Self-serve Pod rates, billed by the second | Check multi-node fabric, sustained capacity, storage, and service-level terms |
| DigitalOcean | HGX H100/H200 droplets with one or eight GPUs | H100 $4.41/GPU-hour; H200 $4.47/GPU-hour; 12-month reserved H100 $3.26/GPU-hour | Published GPU-hour figures; reserved price applies to a 12-month H100 commitment | Validate transfer, storage, region, and scheduler fit for large distributed jobs |
| Oracle Cloud Infrastructure | H100, H200, B200, and MI300X bare metal and VMs; Superclusters up to 131,072 B200 GPUs | No current comparable H100 rate stated | Obtain a current regional quote | The product page’s example pricing is based on June 5, 2024, not a current universal price |
CoreWeave’s node figures illustrate the normalization problem: $49.24 is for eight H100 GPUs, not one. Conversely, a single-GPU Pod rate does not tell you what an eight-node, synchronized training job will cost when networking, storage, and idle allocation are included.
Provider-by-provider fit
AWS: strongest when your data and controls are already there
AWS P5 documentation states that an instance can provide up to eight NVIDIA H100 GPUs and up to 640 GB of total HBM3 memory. P5, P5e, and P5en expose up to 3,200 Gbps of Elastic Fabric Adapter networking. AWS also describes UltraClusters scaling to as many as 20,000 interconnected H100/H200 GPUs.
This combination is aimed at large, tightly coupled jobs. FSx for Lustre can provide a high-throughput training filesystem, S3 is a natural object-storage source and checkpoint target, and IAM, private networking, and security controls are integrated with the rest of an AWS estate. SageMaker and HyperPod provide managed paths for teams that do not want to assemble every scheduler and recovery component themselves.
The trade-off is price complexity. The public product page does not present one hourly number that is comparable with a single-GPU Pod. You must specify P5 generation, region, on-demand versus Capacity Block or another commitment, storage, and data-transfer assumptions, then obtain a current quote.
Google Cloud: a clear path from one machine to cluster orchestration
Google’s A3 families target H100 and H200 training: A3 High and A3 Mega use H100, A3 Ultra uses H200, and A2 uses A100. A3 Mega provides eight H100 SXM GPUs and up to 1,800 Gbps of network bandwidth. Google documents Compute Engine, Google Kubernetes Engine, and Slurm deployment paths, plus regional zones, commitments, and Spot capacity.
Rank #2
- Powered by the NVIDIA Blackwell architecture and DLSS 4
- Powered by GeForce RTX 5060
- Integrated with 8GB GDDR7 128bit memory interface
- PCIe 5.0
- WINDFORCE cooling system
Google is a practical choice when the workload already uses its data, identity, or Kubernetes services, or when you need a documented path from a single eight-GPU machine to scheduled multi-node training. GPU, machine, regional, and commitment charges are separate, and Spot discounts are dynamic, so compare a complete machine configuration in the exact region rather than copying a table value from another location.
CoreWeave: transparent HGX node pricing for GPU-first clusters
CoreWeave publishes US eight-GPU HGX prices: H100 nodes at $49.24 per hour on demand or $19.71 per hour Spot, and H200 nodes at $50.44 per hour on demand or $20.93 per hour Spot. Dividing by eight gives approximately $6.16 and $2.46 per H100 GPU-hour, and $6.31 and $2.62 per H200 GPU-hour, respectively. Those derived numbers assume every GPU in the node is used continuously.
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 matchThe platform’s GPU-first design and Spot option can suit teams that need a full HGX server rather than an isolated accelerator. Confirm the region, queue behavior, storage, interruption policy, and the network topology available across nodes. Spot is cheaper because capacity and continuity are not equivalent to on-demand allocation; checkpoint frequently enough to tolerate an interruption.
Lambda: an advertised route to large interconnected reservations
Lambda advertises H100, H200, and B200 systems, with interconnected clusters ranging from 16 to more than 2,000 GPUs. Its page advertises B200 pricing from $6.69, but that is a starting figure rather than a universal apples-to-apples rate. Confirm the exact GPU, node size, interconnect, term, region, reservation method, and current availability before using it in a budget.
Its appeal is a GPU-specialist procurement path for teams that want a sizable interconnected reservation without designing a hyperscaler environment from scratch. Ask for the expected queue time and a written description of the fabric and support commitment for the planned topology.
Rank #3
- Powered by the NVIDIA Blackwell architecture and DLSS 4
- Powered by GeForce RTX 5070 Ti
- Integrated with 16GB GDDR7 256bit memory interface
- PCIe 5.0
- WINDFORCE cooling system
RunPod: low-friction self-service for experiments and smaller runs
RunPod lists H100 PCIe at $2.89 per GPU-hour and H100 SXM at $3.99 per GPU-hour. Its billing statement says charges begin when the Pod is running and stop when it is terminated, with billing by the second rather than by the hour. That granularity is useful for short experiments, development sessions, and bursty jobs.
PCIe and SXM are not interchangeable choices for a distributed workload. Before committing, verify whether the required GPUs can be placed in one node or across a supported multi-node fabric, what persistent storage costs, how long capacity remains available, and what service-level or support terms apply. A low isolated-GPU rate does not predict the cost or speed of a large synchronized run.
DigitalOcean: straightforward droplets with published GPU-hour rates
DigitalOcean lists HGX H100 at $4.41 per GPU-hour and H200 at $4.47 per GPU-hour. A 12-month reserved H100 price is listed at $3.26 per GPU-hour. The offering is structured as droplets with one or eight GPUs, NVMe scratch storage, and a transfer allowance, which can make the VM workflow familiar to teams already using the platform.
Use the published rate as a starting point, then add persistent storage, data movement, and the cost of keeping eight GPUs allocated during setup or checkpointing. The available evidence is less detailed for hyperscale training fabrics and scheduler integration than for the major hyperscalers, so validate inter-node bandwidth and orchestration with a representative job.
Oracle Cloud Infrastructure: bare metal, RDMA, and an AMD alternative
Oracle documents H100, H200, B200, and AMD MI300X support across bare-metal and VM options, with RDMA networking and Superclusters scaling to 131,072 B200 GPUs. This makes OCI a distinct enterprise option when bare-metal control, RDMA, or AMD capacity is strategically important.
Rank #4
- Powered by the NVIDIA Blackwell architecture and DLSS 4 OC mode: 2640MHz/Default mode: 2610MHz (Boost Clock)
- Military-grade components deliver rock-solid power and longer lifespan for ultimate durability
- Protective PCB coating helps protect against short circuits caused by moisture, dust, or debris
- 3.125-slot design with massive fin array optimized for airflow from three Axial-tech fans
- Phase-change GPU thermal pad helps ensure optimal thermal performance and longevity, outlasting traditional thermal paste for graphics cards under heavy loads
OCI’s GPU page uses an example pricing basis of June 5, 2024, so it should not be treated as a current comparable H100 price. Request a quote for the region, machine type, commitment, storage, and network design you actually need. Confirm software compatibility and porting effort when selecting MI300X instead of NVIDIA hardware.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How to compare two providers fairly
- Write the workload specification. Record model size, sequence length, global batch, precision, optimizer, expected GPU count, checkpoint frequency, training duration, and whether pre-emption is acceptable.
- Normalize the hardware unit. Convert node-hour, VM-hour, and isolated-GPU rates to GPU-hour. Keep a separate column for host and node charges so a full eight-GPU machine is not confused with one accelerator.
- Price a completed run. Multiply the normalized rate by the number of GPUs and wall-clock hours, then add storage, snapshots, data transfer, orchestration, and expected restart or idle time. For Spot, model interruption and checkpoint recovery rather than assuming uninterrupted execution.
- Check topology. Verify NVLink or NVSwitch within a node, RDMA or EFA between nodes, bandwidth, latency, collective-communication support, and the largest cluster size that can be allocated together.
- Check capacity and geography. Ask whether the quoted GPUs are available in the required region, whether capacity is guaranteed, how reservations work, and where data and checkpoints reside.
- Run a reproducible pilot. Use the same container, dataset shard, batch, precision, and number of steps on each candidate. Record tokens per second, scaling efficiency, checkpoint time, failure recovery, and total billed cost. Treat provider specifications as capabilities to verify, not as independent benchmark results.
- Choose the operating model. Select self-service Pods for rapid, smaller experiments; a specialist node or reservation for predictable GPU-first capacity; or a hyperscaler/OCI environment when integrated storage, identity, scheduling, support, or very large clusters justify the added complexity.
Common pricing and deployment traps
- Comparing a node with a GPU: divide CoreWeave’s eight-GPU node price by eight only when all eight GPUs are required and occupied.
- Ignoring the host: some offers bundle CPUs, memory, local NVMe, or networking while others expose separate charges. Include them in the run estimate.
- Assuming Spot is a reservation: dynamic Spot capacity can disappear or be unavailable in the desired region. Keep checkpoints durable and budget recovery time.
- Using a single-region price as a global quote: AWS, Google Cloud, DigitalOcean, and OCI pricing and availability vary by region; recheck immediately before purchase.
- Choosing by GPU name alone: H100 PCIe, H100 SXM, H200, and B200 differ in memory, topology, and software availability. Match the complete machine and interconnect to the parallelism strategy.
- Forgetting data movement: repeated downloads, checkpoint uploads, and cross-region transfers can erase a low compute rate. Place data near the training cluster and measure transfer volume.
- Overlooking scheduler fit: a cluster that cannot launch the required number of GPUs atomically may waste paid allocation while jobs wait for peers.
A practical shortlist by workload
| Workload situation | Starting shortlist | Decision test |
|---|---|---|
| Short experiments or development on one to a few GPUs | RunPod; DigitalOcean | Compare per-second or GPU-hour billing, startup time, persistent storage, and the exact GPU variant |
| Eight-GPU node training with a published specialist rate | CoreWeave; DigitalOcean | Compare full-node cost, NVLink/NVSwitch topology, Spot interruption behavior, and storage |
| Data already in a hyperscaler estate | AWS or Google Cloud | Price data access, IAM, filesystem, scheduler, and network integration together with GPUs |
| Very large synchronized training | AWS, Google Cloud, CoreWeave, Lambda, or OCI | Obtain guaranteed capacity and test collective communication at the intended scale |
| Bare-metal control or AMD MI300X evaluation | OCI | Validate software porting, RDMA topology, support, and a current regional quote |
Bottom line for a buying decision
For the lowest transparent single-GPU entry price in the published figures, RunPod’s H100 PCIe rate and DigitalOcean’s H100 rate are the first numbers to test, but neither proves the lowest completed training cost. CoreWeave’s H100 and H200 Spot node rates can be compelling when you need all eight GPUs and can tolerate interruption. AWS and Google Cloud are often the rational choices when their storage, identity, scheduler, and high-speed networking already surround your data. Lambda and OCI become more relevant when you need a specialist interconnected reservation, bare-metal/RDMA control, or an alternative accelerator mix. The defensible choice comes from a normalized, region-specific run estimate followed by a short benchmark on the exact topology.
Frequently Asked Questions
What is the cheapest cloud for H100 training?
The published figures show RunPod at $2.89 per GPU-hour for an H100 PCIe and DigitalOcean at $4.41 per GPU-hour, while CoreWeave’s eight-GPU H100 node works out to about $6.16 on demand or $2.46 Spot per GPU-hour. These are different billing units and conditions, so the cheapest completed run depends on topology, capacity, storage, transfer, and interruption recovery.
How should I compare CoreWeave’s H100 price with a single-GPU Pod?
CoreWeave’s $49.24 on-demand and $19.71 Spot figures are eight-GPU node-hours. Divide by eight for a per-GPU comparison only when your job uses the entire node, then add storage and account for the node’s networking and capacity characteristics.
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 minuteAre AWS and Google Cloud more expensive than specialist GPU clouds?
Their public pages do not provide one universally comparable hourly figure. They may have higher or lower all-in cost depending on region, commitments, data location, storage, and utilization. Their integrated filesystems, object storage, identity, schedulers, and high-speed fabrics can reduce operational and data-movement costs.
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.
