There is no single cheapest GPU cloud provider. The lowest useful price depends on the same accelerator count, VRAM, region, host CPU and RAM, storage, network, billing mode, and interruption risk. A quoted “GPU-hour” can describe one card, an eight-GPU node, or only the accelerator line item.
For a practical shortlist: RunPod suits short experiments and fine-grained per-second billing; CoreWeave publishes unusually detailed multi-GPU node prices and spot tiers; Google Cloud fits teams that need hyperscaler networking, identity, and committed-use pricing; Oracle Cloud Infrastructure offers a broad NVIDIA and AMD catalog, including current H100/H200 and newer Blackwell and MI-series shapes. Check live capacity immediately before deployment because public price tables do not guarantee inventory.
As an Amazon Associate I earn from qualifying purchases.
What a comparable GPU price must include
Start by defining the workload, then compare identical resources. A price that looks cheap may represent a single GPU with little host memory, while another may include an eight-GPU server with high-bandwidth interconnects.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Repair Windows errors before they cause bigger problems3Scan for outdated or missing drivers - takes under a minuteNormalize the denominator
- GPU count: record whether the price is per card, virtual machine, or complete multi-GPU node.
- Accelerator and VRAM: an A100 40 GB is not interchangeable with an A100 80 GB, and an H100 80 GB is not an H200 141 GB.
- Interconnect: multi-GPU training can depend on NVLink, InfiniBand, or another fabric; PCIe-only connectivity may change scaling efficiency.
- Region and currency: regional prices, taxes, data-residency rules, and inventory differ.
- Billing mode: compare on-demand, reserved or committed use, and spot separately.
- Host resources: include CPU cores, system RAM, local NVMe, persistent disks, and network throughput.
- Data movement: add object-storage operations, inter-zone traffic, and internet egress where applicable.
- Availability risk: a discounted instance that can disappear during training is not equivalent to guaranteed capacity.
| Comparison field | Why it changes the decision |
|---|---|
| Accelerator and VRAM | Determines model fit, batch size, and memory headroom. |
| GPU count and node shape | Changes parallel-training performance and the effective GPU-hour. |
| Region | Affects latency, residency, price, and whether the SKU is actually obtainable. |
| Billing mode | Sets price and the probability of interruption or a long-term commitment. |
| CPU, RAM, storage, and network | Can dominate preprocessing, checkpointing, and data-transfer costs. |
RunPod: flexible experiments and per-second billing
RunPod’s pricing page, updated October 8, 2026, separates Pods, Serverless, and Clusters. Its documentation says, “We offer 30+ GPU models,” including A100, H100, RTX 6000 Ada, and L4/L40 families. That breadth makes it useful when a project needs to move between consumer, workstation, and data-center accelerators rather than standardize on one hyperscaler shape.
#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
How the billing and capacity model works
You choose a region for a Pod and are billed by the second while it runs. That granularity is useful for interactive development, short inference jobs, and bursty experiments where hourly rounding would waste money. The dashboard and API expose current spot availability and pricing, so the number shown at launch is more meaningful than a static estimate.
Spot instances are discounted but may be evicted when demand spikes. Treat them as disposable workers: write checkpoints to durable storage, make startup scripts repeatable, and use a scheduler that can retry failed work. A spot price should not be used for a deadline-sensitive run unless the training system can resume safely.
When a reserved cluster is justified
RunPod says reserved clusters provide guaranteed availability, custom configurations, SLA-backed uptime, and discounted rates for large enterprise deployments. Those benefits address the main weaknesses of on-demand and spot capacity, but they are intended for sustained or enterprise-scale use rather than a one-hour test. Ask for the complete cluster shape, networking, storage, and term before comparing its discount with an hourly alternative.
CoreWeave: transparent multi-GPU node and spot tiers
CoreWeave publishes regional tables that separate on-demand from spot pricing. The following examples are from its United States table in 2026 and are prices for complete eight-GPU nodes, not individual cards.
| Configuration | GPUs and VRAM | On-demand node price | Spot node price | Rough price per GPU-hour* |
|---|---|---|---|---|
| HGX H100 | 8 × 80 GB | $49.24/hour | $19.71/hour | $6.16 on-demand; $2.46 spot |
| HGX H200 | 8 × 141 GB | $50.44/hour | $20.93/hour | $6.31 on-demand; $2.62 spot |
| L40S | 8 × 48 GB | $18.00/hour | $7.88/hour | $2.25 on-demand; $0.99 spot |
| A100 | 8 × 80 GB | $21.60/hour | $9.65/hour | $2.70 on-demand; $1.21 spot |
*The per-GPU figures are simple division of the published node price by eight. They are not standalone single-GPU offers and do not remove the value of the host CPU, RAM, storage, or interconnect.
Regional and SKU caveats
European tables contain separate values and additional Blackwell and Grace Blackwell systems. Some newer SKUs are listed as contact-sales or spot-only. Therefore, a United States H100 calculation cannot be used as a European quote, and a table entry does not prove that a single-GPU shape is available.
Rank #2
- 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
CoreWeave is a strong candidate for distributed training when the full node, high-speed fabric, and predictable shape matter. Spot can cut the listed node price substantially, but the job must tolerate eviction and the scheduler must understand the provider’s interruption behavior.
PC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Outdated 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 matchGoogle Cloud: hyperscaler integration and committed-use rates
Google Cloud lists GPU prices by region and publishes one-year and three-year commitment prices. Its accelerator families map as follows:
| Google Cloud family | Accelerator mapping |
|---|---|
| A2 Standard | NVIDIA A100 40 GB |
| A2 Ultra | NVIDIA A100 80 GB |
| G2 | NVIDIA L4 |
| Other listed lines | T4, P4, P100, L4, and RTX PRO 6000 virtual-workstation rates |
Examples on the 2026 pricing page include a T4 line item at $0.35 per GPU-hour and an L4 virtual-workstation line at about $0.56 per GPU-hour. These are accelerator line-item prices, not complete VM bills. Machine type, boot and persistent disks, attached services, network traffic, and egress are charged separately. A committed-use rate can be lower, but it requires the corresponding one- or three-year commitment and should be compared with the utilization you can realistically sustain.
Why teams choose this model
Google Cloud is often selected when GPU jobs must share existing identity, networking, storage, monitoring, and data platforms. The trade-off is accounting complexity: the apparent GPU rate is only one component of the VM and surrounding services. Verify quota and regional stock before reserving capacity, especially for high-demand accelerators.
Oracle Cloud Infrastructure: broad NVIDIA and AMD catalog
Oracle Cloud Infrastructure’s price list covers bare-metal and virtual-machine GPU shapes. The catalog includes H100, H200, B300, B200, GB200, A100 80 GB and 40 GB, L40S, AMD MI300X, and AMD MI355X systems.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Compare complete shapes, not just cards
Oracle’s bare-metal entries are multi-GPU systems with stated interconnect bandwidth. Use the full shape price, GPU count, host specification, and networking when estimating distributed-training cost. Dividing a bare-metal quote by the number of cards can be useful for a rough screen, but it must not be presented as a purchasable single-GPU rate.
Rank #3
- 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
The published list is region-specific. Select the buyer’s tenancy geography first, then confirm that the exact shape, image, quota, and network service are offered there. AMD MI300X and MI355X can be relevant alternatives when software supports ROCm, but framework, kernel, and container compatibility should be validated before committing to a long run.
Availability, regions, and data residency
Price pages are snapshots, not reservations. RunPod directs users to its dashboard or API for current capacity and spot pricing. CoreWeave marks some SKUs as contact-sales or unavailable for spot. Google Cloud and Oracle publish regional lists, but a listed region can still have no immediately allocatable quota.
Check these items immediately before launch
- Choose the required geography and confirm that moving data there complies with your residency policy.
- Check live inventory or submit a capacity request for the exact GPU, count, VM or bare-metal shape, and zone.
- Confirm quota, account approval, and any contact-sales requirement.
- Validate the interconnect and network path needed for multi-GPU or multi-node training.
- Price storage, snapshots, checkpoints, egress, and idle resources alongside the GPU.
- Record the quote timestamp and terms; prices, regions, and accelerator catalogs can change daily.
A marketplace survey cited 158 listed on-demand prices across 14 providers in 2026, illustrating how many apparently comparable offers exist. That count does not establish that every listing is available, equivalent, or suitable for the same workload.
Recommended Free Tools
On-demand, reserved, and spot: choose by failure tolerance
| Mode | Best use | Main advantage | Main risk |
|---|---|---|---|
| On-demand | Deadlines, interactive work, variable demand | No long commitment; normally the simplest launch path | Highest published hourly rate and possible capacity shortage |
| Reserved or committed | Predictable, sustained utilization | Lower effective rate and, for reserved clusters, stronger availability guarantees | Contract term, utilization risk, and less flexibility |
| Spot | Checkpointed training, batch inference, fault-tolerant jobs | Large discounts, such as the CoreWeave US node examples | Eviction, changing inventory, and restart engineering |
Spot reliability is therefore workload-dependent, not a single percentage. A resilient job can achieve a low effective cost if it checkpoints frequently and restarts automatically; an uncheckpointed run can lose more time and money than the discount saves.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.A practical cost-normalization worksheet
Before selecting a provider, create one row per offer and fill every field. Use the same training plan for each row.
| Field | Value to record |
|---|---|
| Provider and retrieval date | Provider name, region, currency, and timestamp checked |
| Accelerator | Exact model, VRAM, count, and whether the quote is per GPU or per node |
| Host | CPU model or cores, system RAM, local storage, and attached disk pricing |
| Network | Interconnect type, bandwidth, intra-region charges, and internet egress |
| Billing | On-demand, spot, reservation or commitment term, and billing granularity |
| Availability | Live capacity status, quota, eviction policy, and replacement options |
| Effective total | GPU or node cost plus storage, data transfer, orchestration, and expected retries |
For CoreWeave’s eight-GPU US HGX H100 example, dividing $49.24 by eight gives about $6.16 per GPU-hour on demand, while $19.71 divided by eight gives about $2.46 on spot. Those calculations are useful only when your workload needs the entire node and its host resources. The same method applied to a single-GPU VM from another provider can create a misleading “winner.”
Rank #4
- Powered by the NVIDIA Blackwell architecture and DLSS 4. System Requirements: Minimum 850W PSU with 16-pin 12V-2x6 (12VHPWR) connector required. Verify before purchasing.
- Military-grade components deliver rock-solid power and longer lifespan for ultimate durability. Compatibility: 348mm (13.7") length, 3.6 slots, 4.3 lbs. Confirm case clearance and slot spacing. GPU bracket included.
- Protective PCB coating helps protect against short circuits caused by moisture, dust, or debris
- 3.6-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
Which provider fits common workloads?
Short experiments or intermittent inference
Start with a provider offering per-second billing and a region with immediate inventory. RunPod’s Pod model is designed for this pattern; compare the complete Pod configuration, not only its accelerator label.
Eight-GPU distributed training
Examine CoreWeave’s HGX configurations or an equivalent full-node shape from another provider. Focus on interconnect, host resources, checkpoint storage, and whether spot eviction can be tolerated.
Existing hyperscaler platform
Google Cloud is most compelling when your data, IAM, observability, and network controls already live there. Include all VM and service line items in the estimate and compare one- or three-year commitments only against credible utilization.
Need for NVIDIA and AMD options
Oracle Cloud Infrastructure has one of the broadest listed catalogs in this comparison, spanning NVIDIA H100/H200, Blackwell systems, A100 and L40S, and AMD MI300X/MI355X. Confirm software compatibility and the complete bare-metal or VM shape before choosing by card price.
Common mistakes and how to avoid them
- Comparing a node with a card: label every number as GPU-hour, VM-hour, or node-hour.
- Ignoring VRAM: match memory capacity before comparing model names.
- Assuming a listed SKU is deployable: check live capacity, quota, and sales status immediately before launch.
- Treating spot as guaranteed: implement checkpointing, retries, and an alternate shape or region.
- Leaving out egress and storage: estimate the full data lifecycle, including checkpoints and final artifact downloads.
- Using commitment discounts for uncertain demand: model utilization and exit scenarios before signing a one- or three-year term.
- Moving regulated data for a cheaper region: verify residency, contractual, and security requirements first.
Bottom line
Choose by normalized workload cost and dependable capacity, not by the largest discount or the smallest GPU-hour headline. RunPod is a practical starting point for short, flexible jobs; CoreWeave provides clear eight-GPU H100, H200, A100, and L40S node comparisons; Google Cloud adds regional and committed-use integration; and Oracle Cloud Infrastructure offers broad NVIDIA and AMD shape coverage. Recheck inventory, price, quota, and region immediately before every important deployment.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →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.
