The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →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 universally cheapest GPU cloud. The winning offer depends on the same GPU count, VRAM, interconnect, region, host CPU and RAM, storage, network charges, and billing mode. A low spot quote can be unsuitable for a production job, while a higher on-demand node may include the capacity and interconnect that a distributed training run requires.
For a current shortlist, RunPod is oriented toward flexible, per-second rentals and many individual GPU models; CoreWeave publishes detailed regional multi-GPU node prices; Google Cloud adds hyperscaler services and committed-use rates; and Oracle Cloud Infrastructure lists a broad NVIDIA and AMD accelerator catalog. Inventory and prices change frequently, so check live capacity immediately before you deploy.
What makes GPU cloud prices comparable?
A “GPU-hour” is only a useful comparison after the hardware and commercial terms are normalized. Providers may quote one card, a virtual-machine line item, or an eight-GPU bare-metal node. Those are different products even when the accelerator name is identical.
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 reinstallCrashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteNormalize GPU count and billing unit
Record whether the quote covers one GPU or a complete host. CoreWeave’s published HGX figures, for example, are prices for eight-GPU nodes. Dividing the node price by eight gives a rough card-hour figure, but it does not remove the value of the node’s CPUs, memory, local disks, PCIe/NVLink topology, or network fabric. A single-GPU virtual machine from another provider cannot be assumed equivalent.
#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
Match VRAM and interconnect
An A100 40 GB is not interchangeable with an A100 80 GB when a model, batch size, or embedding index needs the larger memory pool. The same applies to H100 and H200 configurations. Multi-GPU training can be limited by inter-GPU bandwidth, so compare the complete system and its interconnect rather than the accelerator label alone.
Include the whole bill
GPU line items usually exclude some combination of host CPU, system RAM, boot and data disks, snapshots, object storage, IP addresses, network transfer, and egress. Google Cloud explicitly notes that its GPU price is only one VM component. Calculate the expected runtime, attached storage, checkpoint traffic, and data movement before declaring a provider cheaper.
Use the same region and currency
Prices are regional. Data-residency rules, taxes, currency conversion, and transfer distance can outweigh a small difference in the accelerator rate. Compare offers for the region where data may legally reside and where users or downstream services will consume the results.
Published price examples
The following figures are directly comparable only within the same provider table: each CoreWeave value is a US-region, eight-GPU node price shown in 2026. Spot values are interruptible and should not be treated as guaranteed capacity.
| Configuration | GPU memory | On-demand node price | Spot node price | Approximate price per GPU-hour* |
|---|---|---|---|---|
| HGX H100 (8 GPUs) | 80 GB per GPU | $49.24/hour | $19.71/hour | $6.16 on-demand; $2.46 spot |
| HGX H200 (8 GPUs) | 141 GB per GPU | $50.44/hour | $20.93/hour | $6.31 on-demand; $2.62 spot |
| L40S (8 GPUs) | 48 GB per GPU | $18.00/hour | $7.88/hour | $2.25 on-demand; $0.99 spot |
| A100 (8 GPUs) | 80 GB per GPU | $21.60/hour | $9.65/hour | $2.70 on-demand; $1.21 spot |
*Dividing a node quote by eight is a rough arithmetic comparison, not a promise that an equivalent single-GPU product exists. Host resources, networking, storage, and regional terms remain part of the real cost.
Google Cloud’s GPU page presents card-level line items by region. Examples include an NVIDIA T4 at $0.35 per GPU-hour and an L4 virtual-workstation line item at about $0.56 per GPU-hour. One- and three-year committed-use prices are also published, generally below the corresponding short-term rate. The complete VM still adds machine, disk, network, and other service charges.
Provider-by-provider comparison
RunPod: flexible rentals and per-second billing
RunPod separates Pods, Serverless, and Clusters. Its documentation says, “We offer 30+ GPU models,” including A100, H100, RTX 6000 Ada, and L4/L40 families. You choose a region, start a Pod, and are billed by the second while it runs. That granularity suits experiments, development environments, inference endpoints with irregular traffic, and jobs that stop immediately after a checkpoint.
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
Spot Pods are discounted, but they may be evicted when demand spikes. The dashboard and API expose current spot availability and pricing, allowing an automation layer to select a replacement region or GPU. Treat a spot Pod as disposable: write checkpoints to durable storage, make startup reproducible, and design for restart rather than assuming a reservation.
For larger deployments, RunPod describes reserved clusters with guaranteed availability, custom configurations, SLA-backed uptime, and discounted rates for enterprise-scale use. Ask for the complete cluster specification before comparing it with a public per-GPU quote; the reservation may include dedicated networking and host resources that a Pod price does not.
CoreWeave: transparent multi-GPU node tables
CoreWeave’s pricing page publishes regional tables and separates on-demand from spot. The US examples above show why node context matters: an eight-GPU H100 node is $49.24 per hour on demand, while an eight-GPU H200 node is $50.44. The H200 price buys substantially more memory per GPU, which can eliminate model sharding or reduce the number of nodes required.
European tables use different prices and include additional Blackwell and Grace Blackwell systems. Some newer SKUs are marked contact-sales or spot-only. For distributed training, inspect the listed interconnect and host shape, not just the card rate. For a single-GPU inference service, verify that the exact one-card product is offered in the required region instead of prorating a node.
Recommended Free Tools
Google Cloud: hyperscaler integration and commitments
Google Cloud maps A2 Standard to NVIDIA A100 40 GB GPUs, A2 Ultra to A100 80 GB GPUs, and G2 to NVIDIA L4 GPUs. Its catalog also includes T4, P4, P100, L4, and RTX PRO 6000 virtual-workstation rates. Prices are listed by region, with one- and three-year commitment options.
This model is useful when the GPU job must connect to existing identity, storage, data-processing, monitoring, or managed-machine-learning services. The trade-off is a more layered invoice: VM type, persistent or local disks, snapshots, networking, and egress are separate decisions. A commitment lowers the rate only when the workload will actually consume the purchased term; intermittent jobs may be better served by on-demand capacity.
Oracle Cloud Infrastructure: broad NVIDIA and AMD choices
Oracle’s price list covers bare-metal and virtual-machine shapes with H100, H200, B300, B200, GB200, A100 80 GB and 40 GB, L40S, AMD MI300X, and AMD MI355X accelerators. Bare-metal offerings are multi-GPU systems with stated interconnect bandwidth, so compare the full shape and network design rather than a card-only estimate.
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
Availability and pricing are tied to the tenancy’s geography. Check the exact region, shape size, host CPU and RAM, storage options, and network charges in the price list for that geography. AMD accelerators can be attractive when the software stack supports them, but verify framework, ROCm, container, and model compatibility before comparing their hourly rate with an NVIDIA system.
Availability, regions, and freshness
Public price pages are not inventory guarantees. RunPod directs users to its dashboard or API for current capacity and spot pricing. CoreWeave can mark a SKU contact-sales or unavailable for spot, and a listed region may have no free nodes when a large training run arrives.
Check capacity immediately before deployment and again in automation just before provisioning. Record the region, exact shape, GPU count, price, and timestamp. If the job is production-critical, confirm an on-demand or reserved path and a second region; do not rely on a spot quote that could disappear during a demand spike.
Data residency can narrow the practical region list. Keep training data, checkpoints, logs, and inference traffic inside approved jurisdictions when required. Moving data to a cheaper region can add egress fees, latency, and compliance work that exceed the GPU saving.
Supported accelerator families
| Provider | Accelerators explicitly listed | Typical comparison strength | Important qualification |
|---|---|---|---|
| RunPod | A100, H100, RTX 6000 Ada, L4/L40 families, and more than 30 GPU models | Single-GPU experimentation, inference, and flexible Pods | Spot capacity and price are live and interruptible; reserved clusters are a separate product |
| CoreWeave | H100, H200, A100, L40S; European listings also include Blackwell and Grace Blackwell systems | Multi-GPU training with published regional node tables | Figures are node prices; some newer SKUs are contact-sales or spot-only |
| Google Cloud | A100 40 GB and 80 GB, L4, T4, P4, P100, RTX PRO 6000 | GPU VMs integrated with hyperscaler services | GPU line items exclude other VM, storage, network, and egress charges |
| Oracle Cloud Infrastructure | H100, H200, B300, B200, GB200, A100 40/80 GB, L40S, AMD MI300X and MI355X | Broad NVIDIA and AMD bare-metal or VM shapes | Shape, interconnect, and regional tenancy pricing must be checked together |
How to select a provider for a real workload
- Specify the workload. Write down framework and container requirements, model size, target batch, expected GPU count, training duration, inference latency, and whether the process can restart from a checkpoint.
- Set the hardware floor. Choose minimum VRAM, required precision support, interconnect bandwidth, host RAM, local scratch space, and network throughput. Reject a cheaper card that cannot hold the model or feed it fast enough.
- Normalize quotes. Convert node prices to the same GPU count only for a first estimate, then add CPU, RAM, disks, snapshots, network transfer, and egress. Keep region and currency constant.
- Choose a failure model. Use on-demand for jobs that cannot tolerate eviction, spot for checkpointed or elastic work, and reserved or committed capacity when steady utilization justifies a term.
- Check live inventory. Query the provider dashboard or API for the exact region and shape immediately before launch. Save the response so an unavailable SKU does not silently turn into a different GPU.
- Run a small acceptance test. Confirm driver and framework versions, GPU-to-GPU bandwidth, storage throughput, dataloader performance, checkpoint time, and actual network costs before committing to a long run.
- Plan fallback capacity. Keep a compatible region, shape, or accelerator and a tested restart image. For spot workloads, automate checkpointing and eviction handling.
Is spot GPU pricing reliable for training?
Spot pricing is reliable as a discount mechanism, not as uninterrupted capacity. RunPod explicitly warns that spot instances can be evicted when demand rises. CoreWeave publishes spot tiers, but some SKUs may be spot-only or unavailable at a given moment. The effective cost is therefore GPU time plus the time and storage needed to recover from interruptions.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Spot works well for queued batch jobs, hyperparameter sweeps, rendering, and training that checkpoints frequently. Use short checkpoint intervals, durable object or network storage, idempotent startup scripts, and a scheduler that can retry in another region. Avoid spot for a single uncheckpointed fine-tune, a live database-backed service, or a deadline that cannot absorb an unknown restart.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Practical cost calculations
Suppose an eight-GPU H100 job runs for 100 hours on the CoreWeave US on-demand figure. The raw node charge is 100 × $49.24, or $4,924, before storage, network, and other services. The same duration at the published spot figure is 100 × $19.71, or $1,971, but an eviction can add recovery time and capacity risk. The rough per-GPU arithmetic is $6.16 and $2.46 respectively, while the job still consumes an eight-GPU host.
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
For a Google Cloud T4 line item, 1,000 GPU-hours at $0.35 is $350 before the VM, disks, and network. That number is useful only if the workload fits a T4 and the surrounding VM costs are included in the final estimate. A lower accelerator rate can lose its advantage when the job requires more GPUs, larger memory, or substantial cross-region egress.
A decision guide by workload
| Workload need | Most relevant option to investigate | Why | Check before purchase |
|---|---|---|---|
| Short experiment or bursty inference | RunPod Pods | Per-second billing and many GPU models reduce idle time | Live regional inventory, image startup time, and spot eviction behavior |
| Large, tightly coupled training job | CoreWeave multi-GPU nodes | Regional node tables expose H100, H200, A100, and L40S system pricing | Interconnect, exact node availability, storage throughput, and on-demand fallback |
| GPU workload integrated with existing cloud services | Google Cloud GPU VMs | Regional pricing and one- or three-year commitments fit established hyperscaler operations | Total VM invoice, commitment utilization, quotas, and egress |
| Need for NVIDIA and AMD choices or bare metal | Oracle Cloud Infrastructure shapes | Catalog includes current NVIDIA systems and MI300X/MI355X | Software compatibility, interconnect, tenancy region, and full-shape price |
What the market listings do—and do not—tell you
A 2026 market listing from GPUCostLab counted 158 listed on-demand prices across 14 providers. Such catalogs are useful for discovering candidates, but a count of prices does not establish capacity, identical hardware, or total cost. Treat every public number as a timestamped starting point and validate the exact product in the provider’s control plane.
Frequently Asked Questions
Can I compare an eight-GPU node quote with a single-GPU cloud VM?
Only as a rough arithmetic estimate. A node price also buys host CPUs, RAM, local storage, interconnect, and networking, so a prorated GPU-hour is not an equivalent single-GPU product.
Which provider has H200 capacity right now?
No public price page can guarantee inventory at a future moment. Check the provider dashboard or API for the required region and shape immediately before provisioning, and keep an on-demand or alternate-region fallback.
When do committed-use discounts make sense?
They make sense when utilization is predictable for the full one- or three-year term and the workload remains in the quoted region. Bursty or experimental jobs may pay less overall with on-demand or interruptible capacity.
Are AMD MI300X and MI355X drop-in replacements for NVIDIA GPUs?
No. Confirm ROCm, framework, container, kernel, and model support, then benchmark memory use and throughput on the exact shape before switching accelerator families.
Free tools Windows power users keep installed
One-click scans. No signup required.
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.
