There is no universally cheapest cloud for AI GPUs. CoreWeave’s published North America rate card is aggressive for several eight-GPU nodes, but AWS or Azure can cost less after committed-use discounts, spot capacity, enterprise agreements, regional availability, storage, data transfer and managed services are included. The right answer depends on your exact GPU model, region, billing plan and measured utilization.
For a defensible decision, compare the same accelerator and memory capacity under the same geography and purchasing terms, then calculate cost per training step, generated token or inference request. A low hourly number is only the starting point.
As an Amazon Associate I earn from qualifying purchases.
CoreWeave’s published GPU rates at a glance
CoreWeave’s 2026 North America rate card lists the following hourly prices. Each figure is for an entire instance with eight GPUs, not a single accelerator. Spot pricing, availability and other regions can differ.
Free tools Windows power users keep installed
One-click scans. No signup required.
| Instance | GPUs per instance | Published hourly rate | Approximate rate per GPU* | 730-hour month, GPU rental only |
|---|---|---|---|---|
| HGX H200 | 8 | $50.44 | $6.305 | $36,821.20 |
| HGX H100 | 8 | $49.24 | $6.155 | $35,945.20 |
| A100 | 8 | $21.60 | $2.700 | $15,768.00 |
| L40S | 8 | $18.00 | $2.250 | $13,140.00 |
*The per-GPU column is simple division of the node rate by eight. It is not a separately quoted CoreWeave tariff and does not account for the CPU, memory, local storage or network included in the node.
#1 Best Overall
- Read Before You Buy — No Video Output: These adapters support charging and USB 2.0 data transfer, but cannot transmit video signals. Except for standard USB webcams (which use USB data only), they are not compatible with HDMI/DisplayPort cables, video-capable USB-C hubs, or docking stations with video output.
- Convert USB-A Ports to USB-C: Designed to connect USB-C earphones, cables, flash drives, card readers, and other USB-C accessories to standard USB-A ports. Plug-and-play with no drivers or software required.
- Aluminum Alloy Housing: Built with a sturdy aluminum alloy shell that aids in heat dissipation and protects against daily wear and scratches. Designed to maintain a stable and secure connection.
- Compact & Travel-Friendly: The ultra-compact design allows the adapter to stay plugged into your device without blocking adjacent ports or adding bulk, reducing wear and tear on your original USB ports.
- 12-Month Warranty: Backed by a 12-month manufacturer warranty for peace of mind. Designed to meet strict quality control standards for reliable everyday performance.
The monthly figures assume 730 hours of continuous use and contain no storage, egress, orchestration, support or other operating costs. They are useful for scale, not a final invoice estimate.
Why an hourly comparison can be misleading
A GPU instance is a bundle of an accelerator, host CPU, system memory, local disks, networking and a software environment. Two instances described as “H100” can have different GPU counts, interconnects, host specifications or virtualization models. A training job may therefore finish faster or slower even when the advertised GPU name is identical.
Cloud prices also change with purchase channel. On-demand capacity is flexible but usually carries the highest rate. Reserved capacity, Savings Plans, committed-use contracts and enterprise agreements trade flexibility for a lower effective price. Spot or preemptible capacity can be much cheaper, but the provider may interrupt it with limited notice. A comparison that mixes one provider’s on-demand rate with another provider’s commitment or spot rate is not meaningful.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Finally, the GPU is only one line on the bill. Checkpoint storage, object-store requests, inter-region traffic, public-egress charges, load balancers, observability, support, orchestration and idle instances can outweigh a small difference in accelerator price.
What AWS and Azure offer
AWS EC2 provides accelerated-computing families including P5 and P4 for machine-learning and high-performance-computing workloads. The price depends on the exact instance size, AWS region and purchase option, so there is no single AWS number that can be placed beside CoreWeave’s North America node rates without choosing those variables.
Microsoft Azure documents NC and ND GPU-optimized virtual-machine families for AI and HPC. Azure exposes regional prices through its pricing tools, and Microsoft agreements or other commitments can change the effective rate. ND and NC configurations also differ in GPU generation, count, memory and interconnect; use the exact VM SKU rather than the family name when comparing.
Rank #2
- 5-in-1 USB-C Hub: Experience comprehensive connectivity featuring a Power Delivery input, two USB-A 2.0 ports, a USB-A 3.0 port, and an HDMI port. (Note: The USB-C power delivery input port is only for connecting an external wall charger to power your laptop and cannot power peripheral devices.)
- 90W Pass-Through Charging: Achieve optimal charging with 90W pass-through power to your laptop, supported by a total input of 100W, with the hub reserving 10W for operational efficiency. (Note: Wall charger not included.)
- Quick Data Transfers: Accelerate your productivity with rapid data transfers using a high-speed 5Gbps USB 3.0 port and two 480Mbps USB 2.0 ports.
- 4K HDMI Display: Enhance your visual experience with a hub capable of delivering 4K resolution at 30Hz in both mirror and extend modes. Please note that this hub is compatible with MacBook (macOS 12 and newer), Windows 10 and 11, ChromeOS, and laptops equipped with DP Alt Mode and Power Delivery. Note: This device is not compatible with Linux.
- What You Get: Anker USB-C Hub (5-in-1, 4K HDMI), welcome guide, 18-month warranty, and our friendly customer service.
| Provider | GPU families cited for AI/HPC | Comparable current hourly figure in the available data | What must be selected before pricing |
|---|---|---|---|
| CoreWeave | HGX H200, HGX H100, A100, L40S | Published North America node rates shown above | Exact node, region, billing tier, storage and network terms |
| AWS | P5 and P4 EC2 families | Not stated as one apples-to-apples figure; varies by SKU, region and purchase option | Instance type, region, on-demand/commitment/spot choice and adjacent services |
| Azure | NC and ND VM families | Not stated as one apples-to-apples figure; varies by SKU, region and agreement | VM SKU, region, pricing agreement, Azure ML and storage configuration |
The absence of one universal AWS or Azure number is not a missing feature: both catalogs contain many configurations. It means you must define the workload before a price ranking is honest.
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 minuteNormalize the hardware before comparing prices
Match accelerator generation and memory
Start with the GPU model and memory capacity your software requires. An H200, H100, A100 and L40S are not interchangeable for every model. Memory limits can force a larger node, model sharding or a different batch size, changing total cost even if the nominal hourly rate is lower.
Match GPU count and interconnect
Compare eight-GPU nodes with eight-GPU nodes, or convert both offers to the same number of accelerators only after confirming that the host and interconnect are equivalent. Distributed training can be highly sensitive to GPU-to-GPU bandwidth and topology. A cheaper node that scales poorly may cost more per completed training step.
Include host resources
Record vCPU count, system RAM, local NVMe capacity, attached block storage, network bandwidth and any license charges. If one provider includes local scratch storage and another requires a separate disk, add the second disk before deciding.
Hold geography constant
Choose the same country or an equivalent low-latency region. Region changes affect price, data-transfer distance, service availability and data-residency compliance. CoreWeave’s figures above are specifically for North America; they should not be generalized to every CoreWeave location.
Outdated 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 matchWindows 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 reinstallSeparate on-demand, committed and interruptible capacity
| Pricing mode | Best for | Main benefit | Main risk |
|---|---|---|---|
| On-demand | Short projects, experiments and uncertain demand | No long-term commitment; capacity can be released when idle | Usually the highest hourly price and possible capacity shortages |
| Reserved or committed | Predictable, sustained workloads | Lower effective rate when usage is maintained | You can pay for unused commitment or lose flexibility |
| Spot or preemptible | Checkpointed training, batch inference and fault-tolerant jobs | Potentially large discount | Interruption, queueing and uncertain availability |
Quote each mode separately. Do not average them into one “cloud price.” A production service that cannot tolerate interruption should not be costed on a spot rate, while a checkpointed research job may be a good candidate for one.
Rank #3
- Sleek 7-in-1 USB-C Hub: Features an HDMI port, two USB-A 3.0 ports, and a USB-C data port, each providing 5Gbps transfer speeds. It also includes a USB-C PD input port for charging up to 100W and dual SD and TF card slots, all in a compact design.
- Flawless 4K@60Hz Video with HDMI: Delivers exceptional clarity and smoothness with its 4K@60Hz HDMI port, making it ideal for high-definition presentations and entertainment. (Note: Only the HDMI port supports video projection; the USB-C port is for data transfer only.)
- Double Up on Efficiency: The two USB-A 3.0 ports and a USB-C port support a fast 5Gbps data rate, significantly boosting your transfer speeds and improving productivity.
- Fast and Reliable 85W Charging: Offers high-capacity, speedy charging for laptops up to 85W, so you spend less time tethered to an outlet and more time being productive.
- What You Get: Anker USB-C Hub (7-in-1), welcome guide, 18-month warranty, and our friendly customer service.
A repeatable method for calculating true GPU cost
- Define the workload. Write down model size, precision, target batch size, training duration, expected inference volume, latency target and whether interruptions are acceptable.
- Select a reference configuration. Specify GPU model, GPU count, memory, vCPU, RAM, local storage and interconnect. Use the same specification for every provider where possible.
- Choose a region or equivalent geography. Keep the data source, users and compliance requirements in the same region for each quote.
- Collect three price modes. Record on-demand, committed and spot/preemptible prices with the quote date and any minimum term. A dynamic price should be refreshed before purchase.
- Add non-GPU resources. Include boot volumes, checkpoint disks, object storage, snapshots, CPUs, RAM, software licenses and orchestration workers.
- Estimate transfer. Count data entering and leaving the GPU region, transfers between zones, model downloads, checkpoint uploads and user-facing egress. Apply each provider’s current regional terms.
- Measure performance. Run a representative benchmark and capture training steps per hour, tokens per second, inference requests per second and peak memory use. A synthetic benchmark that does not exercise your model can mislead.
- Convert to a unit cost. Use
cost per training step = total run cost ÷ completed steps,cost per million tokens = total inference cost ÷ (generated tokens ÷ 1,000,000)orcost per request = total service cost ÷ successful requests. - Model utilization. Separate wall-clock uptime from useful GPU time. Include data-loading stalls, queue time, debugging, autoscaling lag and idle periods.
- Add operational labor and risk. Account for monitoring, incident response, engineering time, migration work, support contracts and the expected cost of an interruption.
This process turns an hourly rate into the number your finance or engineering team actually needs.
Provider-specific checks
CoreWeave
- Verify that the quoted HGX node has the GPU generation, memory and interconnect your framework expects.
- Confirm queue behavior and available capacity for the region and time window you need.
- Read storage and network terms for the exact SKU; a low accelerator rate can be offset by checkpoint or egress charges.
- Ask whether the quoted figure is on-demand, committed or spot, and record interruption terms for any discounted capacity.
- Benchmark multi-GPU scaling rather than assuming eight GPUs deliver eight times single-GPU throughput.
AWS
- Identify the exact P5 or P4 EC2 instance type and region; the family label alone is insufficient.
- Price on-demand, Savings Plans or reservations, and spot separately.
- Include EBS or other storage, data transfer, elastic networking and any managed training or inference service.
- Check whether existing AWS commitments or enterprise pricing materially lower the effective rate.
- Value integrated identity, monitoring, data services and deployment tooling only if they replace work or services you would otherwise buy.
Azure
- Choose the precise NC or ND VM SKU, GPU count and region.
- Check Microsoft agreements and other committed-use terms before comparing with public rates.
- Include managed-disk, Blob Storage, networking and Azure Machine Learning charges.
- Verify regional capacity and quota lead time; a nominally cheaper SKU is not useful if you cannot obtain enough GPUs.
- Account for existing Microsoft identity, security and data-platform integrations when estimating migration and operating effort.
Utilization changes the answer
Suppose an eight-GPU H100 node is rented continuously at CoreWeave’s published $49.24 per hour. Using a 730-hour month gives $35,945.20 before every other charge. If the node is powered on for 730 hours but performs useful work for only half that time, the effective accelerator cost during useful work doubles to $98.48 per productive hour. The same arithmetic applies to any provider and makes idle capacity visible.
For bursty inference, autoscaling can reduce the paid hours but may introduce cold-start latency and queueing. For long training runs, a commitment can lower the rate, yet a failed experiment or delayed dataset leaves the commitment unused. Model at least three utilization cases—low, expected and high—rather than relying on one forecast.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Storage, networking and operations can dominate
| Cost area | Questions to answer | Why it matters |
|---|---|---|
| Checkpoint and dataset storage | How many terabytes, how many versions and how long are they retained? | Repeated checkpoints and replicas can cost more than expected, especially for multi-month projects. |
| Data transfer | Where does training data originate, and where do outputs go? | Cross-region and public egress can turn a low GPU rate into a high total bill. |
| Orchestration | Who pays for schedulers, control-plane nodes, registries and worker services? | Clusters often run support instances while GPUs are busy or idle. |
| Observability and support | Are logs, metrics, tracing and premium support required? | Production reliability has recurring charges and engineering overhead. |
| Migration and integration | How much code, identity and networking must change? | A small hourly saving may not repay migration labor or lock-in risk. |
When local hardware beats cloud rental
A consumer GPU such as NVIDIA’s GeForce RTX 5090 can be useful for local experimentation or inference, but it is not a universal substitute for a datacenter accelerator. Consumer hardware differs in memory capacity, virtualization support, reliability features, serviceability and deployment profile.
To compare a local system fairly, calculate the purchase price plus workstation components, electricity, cooling, maintenance, depreciation and the value of your own time. Divide that total by productive GPU hours over the expected service life. Include the cost of spare parts and downtime, and account for the fact that local hardware cannot elastically scale for a short training surge.
Local equipment often wins when utilization is steady and predictable, the model fits in available memory and the team can operate the system. Cloud capacity is usually more attractive for bursty demand, large distributed jobs, rapid experimentation or workloads that require specialized datacenter interconnects.
Rank #4
- Dual Converters, Infinite Potential：Includes 2× USB C male to USB A female adapters and 2× USB A male to USB C female adapters. Perfect for a wide range of uses—tablets with Bluetooth keyboards, expand USB ports on macbook, and more. Two different converters for all your daily needs
- Next-Level 10Gbps & 3A Charging: No more slow 480Mbps, this usb to usb c adapter has a transfer speed of up to 10Gbps, allowing you to do more transferring in less time. This usb adapter fits both USB A and USB C charger, supporting up to 3A fast charging
- Upgraded Exquisite Craftsmanship: With an aluminum alloy housing and metal connector, the usbc to usb adapter is extremely durable and sturdy. Rigorously tested to withstand more than 10,000 times of plugging and unplugging, ensuring long-lasting performance
- Broad Compatible: The usb c to usb adapter widely supports all USB C/ USB A devices like laptops, tablets, cellphones, car chargers, and phone chargers. Such as compatible with MacBook Pro/Air 2023/2022, Thunderbolt 4/3 Devices,Apple MagSafe Watch 9/8/7/SE/Ultra, iPad Pro 2022/2021, Samsung Galaxy S23/S20/S10, and iPhone 17/16/15 Pro. Plug and play
- Please Note: To reach 10Gbps speed, keep the cable under 3.3 ft. For USB A Male to USB C adapters, try flipping the USB C connector. USB C Male to USB A adapters support bidirectional 10Gbps transfer within 3.3 ft
Workload-based decision guide
| Workload pattern | Usually the strongest starting option | Checks before committing |
|---|---|---|
| Short experiments and uncertain demand | On-demand capacity from the provider with available GPUs | Release idle instances promptly; compare startup time and regional capacity. |
| Long, repeatable training | Committed capacity or a negotiated agreement | Confirm utilization forecast, cancellation terms and benchmarked throughput. |
| Checkpointed batch training | Spot or preemptible capacity | Test restart logic, checkpoint frequency and interruption behavior. |
| Latency-sensitive production inference | Reliable reserved or on-demand nodes near users and data | Measure tail latency, autoscaling response and failover capacity. |
| Steady single-GPU development | Possibly local hardware | Include electricity, cooling, depreciation, maintenance and security. |
| Large distributed training | Provider with suitable multi-GPU topology and capacity | Measure scaling efficiency, network bandwidth and queue time, not just hourly price. |
FinOps controls that prevent GPU waste
Variable GPU spending needs active management rather than a monthly surprise. Apply visibility, allocation and optimization practices to every project and team.
Recommended Free Tools
- Tag instances with owner, project, environment, model and expiration date.
- Set automatic shutdown schedules for development and notebooks.
- Alert on low GPU utilization, stalled data loaders and instances running without a job.
- Separate training, inference, experimentation and shared-platform budgets.
- Track cost per experiment, training step, token and successful request—not only total cloud spend.
- Review storage retention and delete obsolete checkpoints, images and snapshots.
- Keep an interruption log for spot workloads and quantify restart time.
- Revisit commitments when model architecture, traffic or team usage changes.
Common comparison mistakes and fixes
Comparing different GPU generations
Symptom: one quote is dramatically cheaper but training takes longer or cannot fit the model. Fix: match memory, precision support and measured throughput before comparing rates.
Using a per-GPU number for a node-only quote
Symptom: a divided figure looks cheaper than a competitor’s advertised single-GPU price. Fix: keep node and per-GPU figures clearly labeled and include host resources.
Ignoring interruption risk
Symptom: a spot estimate looks excellent until jobs restart repeatedly. Fix: add checkpoint storage, restart time and expected lost work to the spot scenario.
Leaving idle time out
Symptom: the project’s invoice is far above the benchmark estimate. Fix: measure productive GPU hours and automate shutdowns.
Forgetting data movement
Symptom: training is cheap but exporting models or serving users is expensive. Fix: map every data path and price inter-region and public egress.
Best Value
- 5-in-1 Connectivity: Equipped with a 4K HDMI port, a 5 Gbps USB-C data port, two 5 Gbps USB-A ports, and a USB C 100W PD-IN port. Note: The USB C 100W PD-IN port supports only charging and does not support data transfer devices such as headphones or speakers.
- Powerful Pass-Through Charging: Supports up to 85W pass-through charging so you can power up your laptop while you use the hub. Note: Pass-through charging requires a charger (not included). Note: To achieve full power for iPad, we recommend using a 45W wall charger.
- Transfer Files in Seconds: Move files to and from your laptop at speeds of up to 5 Gbps via the USB-C and USB-A data ports. Note: The USB C 5Gbps Data port does not support video output.
- HD Display: Connect to the HDMI port to stream or mirror content to an external monitor in resolutions of up to 4K@30Hz. Note: The USB-C ports do not support video output.
- What You Get: Anker 332 USB-C Hub (5-in-1), welcome guide, our worry-free 18-month warranty, and friendly customer service.
Assuming capacity is guaranteed
Symptom: a selected SKU cannot be obtained when a deadline arrives. Fix: check quotas, queue times and regional supply before signing a commitment.
A practical worksheet for your own quotes
For each provider and scenario, record:
- Provider, region and quote date.
- Exact GPU SKU, count, memory and interconnect.
- vCPU, RAM, local NVMe and attached-storage configuration.
- On-demand, committed and spot rates with term and interruption conditions.
- Expected GPU hours, productive utilization and job duration.
- Storage terabytes, retention period and checkpoint frequency.
- Inbound, inter-region and outbound data volume.
- Control-plane, monitoring, support and engineering costs.
- Measured steps per hour, tokens per second or requests per second.
- Total cost and unit cost for low, expected and high utilization.
Keep the worksheet with the benchmark results. When a provider changes a price, region or instance generation, you can update one assumption instead of rebuilding the analysis.
Bottom line: which cloud is cheapest?
CoreWeave is a strong candidate when its North America node rate, GPU availability and network/storage terms match your workload. Its published 2026 prices—$50.44 per hour for eight H200 GPUs, $49.24 for eight H100 GPUs, $21.60 for eight A100 GPUs and $18.00 for eight L40S GPUs—are compelling list prices, but they are not complete workload costs.
AWS may win when Savings Plans, reservations, existing enterprise terms or adjacent managed services reduce the total cost of operating there. Azure may win when Microsoft agreements, Azure Machine Learning integration, identity and data services eliminate other spending or engineering work. A local RTX 5090 system may win for steady, modest-scale use after electricity and ownership costs are included.
The defensible choice is the provider with the lowest measured cost for your required output—training step, token or request—under the same hardware, geography, commitment and reliability assumptions. Recalculate when utilization, model size, traffic or regional availability changes.
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.
