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
Short answer: CoreWeave can be less expensive for a comparable GPU when its regional on-demand or spot capacity is available, but the published figures are not a universal price ranking. CoreWeave’s table mixes GPU and node configurations; AWS’s clearest H100 figure here is an eight-GPU Capacity Block reservation; Google Cloud’s A3 prices include a machine family and region; and Azure’s ND H100 v5 price is calculator- and scope-dependent.
For production, there is no defensible provider-wide reliability winner in the available public data. Choose by measured provisioning success, quota lead time, interruption and recovery behavior, networking, storage, support, and the failover options available in the regions and GPU SKUs your workload actually needs.
Executive verdict
- Lowest published spot examples: CoreWeave lists L40 at $6.27 per hour, L40S at $7.88, A100 at $9.65, and H200 at $20.93 in its 2026 North America table. Those are regional rates and may represent different billing scopes, so they cannot be compared directly with every cloud’s instance price.
- Most transparent H100 reservation example: AWS lists a P5.48xlarge Capacity Block in the US East (Atlanta) Local Zone at $41.528 per instance-hour. Because the instance has eight H100 accelerators, that is $5.191 per H100 accelerator-hour when the full instance is utilized. The Capacity Block rate is fixed at purchase, but the product has a defined term and region.
- Google Cloud A3: the displayed A3 High eight-H100 price is $88.490000119 per hour on demand and $50.458920548 per hour on spot. Google also lists an A3 High one-H100 spot configuration at $6.307365068 per hour. Machine type, region, and the rest of the VM configuration still affect the total.
- Azure: ND H100 v5 is an eight-H100 VM with substantial host memory and interconnect bandwidth, but an hourly rate is not stated in the public material used here; Azure directs buyers to its pricing calculator and the applicable deployment scope.
- Best fit depends on workload: a single-GPU inference service, an eight-GPU training job, a spot-tolerant batch queue, and a multi-region production service have different cost and availability requirements. Benchmark the exact SKU, region, billing mode, and software stack before committing.
Why a GPU-hour headline is often misleading
A quoted “GPU-hour” can describe a whole VM, an eight-GPU node, one accelerator inside that node, or a reserved capacity product. It can also omit host CPU, system memory, local NVMe, network fees, storage, and data transfer. Before comparing offers, align all of these dimensions:
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →- GPU generation and memory: an H100, H200, A100, L40, and L40S have different memory capacities and performance characteristics. Even two H100 offers can differ by memory variant, host configuration, or interconnect.
- Billing unit: determine whether the number is per accelerator, per VM, per node, or per reservation block. Divide a node rate by the number of GPUs only when you will keep the entire node busy.
- Region and date: prices and capacity are regional and change over time. Record the region, currency, retrieval date, and whether the number is promotional, spot, on-demand, reserved, or committed.
- Host resources: include vCPUs, system RAM, local NVMe, operating-system disks, and any minimum VM size. Google Cloud explicitly says its GPU price must be combined with the machine-type configuration.
- Interconnect: distributed training can be limited by GPU-to-GPU and node-to-node bandwidth rather than the accelerator’s theoretical compute rate.
- Interruption policy: spot capacity can be reclaimed. A lower rental price is only cheaper after adding checkpointing, restart time, lost work, queueing, and any on-demand fallback.
- Data movement: model downloads, dataset reads, checkpoint storage, cross-zone traffic, and egress can exceed the apparent accelerator saving.
Published pricing snapshot for 2026
The table below preserves the billing scope and qualifications of the published examples. It is a dated snapshot, not a normalized ranking.
#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 and offer | Hardware or scope | On-demand or fixed-rate price | Spot price | Important qualification |
|---|---|---|---|---|
| CoreWeave L40 | Configuration scope as listed in CoreWeave’s North America table | $10.00/hour | $6.27/hour | Regional table; the published examples may be node- or GPU-level configurations. |
| CoreWeave L40S | Configuration scope as listed in the North America table | $18.00/hour | $7.88/hour | Regional rate; confirm the number of GPUs and host resources in the selected configuration. |
| CoreWeave H200 | Configuration scope as listed in the North America table | $50.44/hour | $20.93/hour | Do not compare with an H100 or with a differently sized node without normalizing hardware and utilization. |
| CoreWeave A100 | Configuration scope as listed in the North America table | $21.60/hour | $9.65/hour | Regional rate; verify memory variant, host resources, and interruption terms. |
| CoreWeave GB300 and HGX B200 entries | Current-generation systems | Contact sales | Contact sales | No public hourly figure is stated for these entries. |
| AWS P5.48xlarge Capacity Block, US East (Atlanta) Local Zone | Eight H100 GPUs per P5.48xlarge instance | $41.528 per instance-hour | Not stated for this Capacity Block example | Equivalent to $5.191 per H100 accelerator-hour if all eight GPUs are used; the prevailing rate is fixed at purchase and the block has a defined term. |
| Google Cloud A3 High, eight H100s | Eight H100 80GB GPUs in an A3 High machine | $88.490000119/hour | $50.458920548/hour | Displayed accelerator-optimized pricing; region and machine-type configuration apply. |
| Google Cloud A3 High, one H100 | One H100 in an a3-highgpu-1g configuration | Not stated in the displayed data | $6.307365068/hour | Spot price shown for the one-GPU configuration. |
| Azure ND H100 v5 | Eight H100 GPUs, 96 vCPUs, 1,900 GiB memory, 3.2 Tbps interconnect | Not stated; use Azure’s calculator for the deployment scope | Not stated | Price depends on region, billing scope, and the selected Azure service. |
How to read the CoreWeave figures
CoreWeave publishes separate regional tables, so a North America number does not establish availability or price in another region. The table includes both on-demand and spot figures for L40, L40S, H200, and A100, while some GB300 and HGX B200 systems are marked “Contact sales.” Confirm whether the selected line item is priced per GPU or per host before calculating a per-accelerator rate.
How to read the AWS Capacity Block figure
The $41.528 number is for one P5.48xlarge instance-hour in the US East (Atlanta) Local Zone, not a general AWS H100 rate. Dividing by eight produces $5.191 per accelerator-hour only when the complete eight-GPU instance is allocated and used. A Capacity Block is a reservation product: its term, purchase timing, region, and fixed prevailing rate matter. AWS also offers P5 configurations with one or eight H100 GPUs and P5e and P5en configurations with eight H200 GPUs, so another P5 family member will not necessarily have this price.
How to read the Google Cloud A3 figures
Google’s displayed A3 High eight-H100 rate is quoted per machine hour. For 1,000 machine-hours, the arithmetic is $88,490.000119 on demand or $50,458.920548 on spot before VM configuration, storage, and network charges. Dividing the eight-GPU spot figure by eight gives approximately $6.3073650685 per H100-hour, consistent with the displayed one-GPU spot price after rounding. That derived number does not remove the need to pay for the machine configuration and does not prove equal availability in every region.
Free tools Windows power users keep installed
One-click scans. No signup required.
What is known about Azure pricing
Azure’s ND H100 v5 specification is public, but an hourly price is not fixed in the material used for this comparison. Azure pricing varies with region, purchase model, service scope, and agreement. Microsoft Foundry managed compute bills per accelerator, with rates varying by accelerator family and deployment scope; that managed-compute price should not be substituted for an ND H100 v5 VM price.
Hardware and architecture differences
AWS P5, P5e, and P5en
AWS documents P5 instances with one or eight H100 GPUs. P5e and P5en provide eight H200 GPUs. The documented design supports up to 3,200 Gbps Elastic Fabric Adapter networking and 900 GB/s NVSwitch peer-to-peer bandwidth. AWS says P5, P5e, and P5en can run in EC2 UltraClusters scaling to as many as 20,000 H100 or H200 GPUs. Those capabilities matter for large distributed jobs, but they do not guarantee that a particular SKU, quota, or cluster location will be available when you request it.
Azure ND H100 v5
Each ND H100 v5 VM contains eight H100 GPUs, 96 vCPUs, and 1,900 GiB of memory. Azure lists 3.2 Tbps of interconnect bandwidth, a key consideration for all-reduce-heavy training. Compare that complete VM with a complete eight-GPU node elsewhere; comparing only the H100 count can hide differences in host memory, CPU balance, storage, and network topology.
Google Cloud A3 High
Google’s A3 High machines use H100 80GB GPUs. Google directs customers to add the machine-type configuration to the accelerator price when calculating total VM cost. The one-GPU and eight-GPU spot examples demonstrate that the billing unit changes with the selected A3 shape, so reserve enough host resources for data loading, compilation, and orchestration rather than treating the GPU line as the entire invoice.
CoreWeave configurations
The published CoreWeave price lines identify accelerator families and regional on-demand or spot rates but do not, in the supplied figures, provide a common host-resource and interconnect specification for every line. Obtain the exact SKU details before comparing an L40, L40S, A100, or H200 offer with a full AWS, Azure, or Google Cloud VM.
Rank #2
- Chipset: GeForce RTX 3050
- Boost Clock / Memory: 1492 MHz / 14 Gbps
- Video Memory: 6GB GDDR6
- Memory Interface: 96-bit
- Output: DisplayPort x 1 (v1.4a) / HDMI 2.1a x 2
Availability: treat it as a live measurement
Public price lists cannot answer “Which cloud always has H100s?” Capacity changes by region, time, quota, reservation product, and demand. CoreWeave separates regional tables and marks some systems contact-sales. AWS publishes which regions and terms support Capacity Blocks. Google exposes region-specific A3 spot data. Azure documents the ND H100 v5 SKU but sends buyers to the calculator and deployment scope.
For every capacity request, record:
- GPU model, memory variant, quantity, and complete machine shape.
- Target region, zone or Local Zone, and fallback regions.
- Quota status and the time required to obtain or increase quota.
- Requested billing mode: on-demand, spot, reserved, Capacity Block, or commitment.
- Time from API request to a running instance, including failed attempts.
- Whether capacity was delivered in one placement group or split across hosts.
- Any provider-side status event, maintenance notice, or interruption.
A practical availability test
- Define the demand profile. Write down the exact GPU count, minimum memory, interconnect requirement, region, launch window, and maximum acceptable queue time.
- Preflight quota. Request quota before the experiment or production launch. Record the approved limit and the date it was granted; a quota approval is not a promise of physical capacity.
- Probe capacity repeatedly. Run identical provisioning attempts at representative times, using the same region and machine shape. Log successes, failures, launch latency, and error messages.
- Test fallbacks. Try the next approved region, a smaller shape, or on-demand capacity if spot is unavailable. Keep the fallback policy explicit so an automated scheduler does not silently select an incompatible GPU.
- Measure usable capacity. A VM that launches but lacks the required network topology, local storage, or placement is not a successful sample for a distributed job.
- Set an expiry on the result. Re-run the test after major product launches, quota changes, or sustained demand. Availability is not a permanent property of a price page.
Reliability for production workloads
The available public material does not provide a common, provider-wide GPU uptime statistic, so it cannot support a universal reliability ranking. Reliability is a workload-level outcome involving compute, control plane, networking, storage, scheduling, and recovery.
Metrics that belong in an SLO review
- Provisioning success rate: the percentage of requests that deliver the required shape within the allowed time.
- Quota lead time: elapsed time to obtain initial quota and an increase.
- Interruption rate and notice: how often spot jobs are reclaimed and how much warning the workload receives.
- Recovery time: time from failure or reclamation to a healthy replacement worker.
- Checkpoint recovery: lost work per interruption and the time to restore optimizer state, data-loader state, and model artifacts.
- Network behavior: bandwidth, latency, packet loss, and collective-operation stability at the required scale.
- Storage dependency: checkpoint durability, read throughput, write throughput, and behavior during storage service degradation.
- Failover: ability to move to another zone, region, provider, or GPU generation without changing the model or data pipeline.
- Support response: escalation path and response time for capacity, hardware, and networking incidents.
On-demand is not the same as interruption-free
On-demand removes the normal spot-reclamation mechanism, but it does not eliminate quota limits, hardware faults, maintenance, regional incidents, or application failures. Capacity Blocks add a fixed purchase rate for their term, yet the reservation still depends on the supported region, schedule, and the block’s scope. Build health checks, checkpointing, retries, and replacement capacity into every production design.
Designing for spot interruptions
Use spot for work that can pause, resume, or be restarted. Checkpoint to durable storage at a cadence based on the cost of lost computation, not an arbitrary timer. Keep images and dependencies reproducible, make training jobs idempotent, and maintain an on-demand or alternate-region escape route. The correct comparison is expected completed work per dollar, not the lowest spot number.
Real total cost: a calculation framework
Use this model for each provider and configuration:
Total cost = accelerator or VM rental + host-resource charges + durable storage + local storage or disk charges + data-transfer and egress fees + idle reservation time + interruption and restart cost + engineering and support cost.
The rental examples illustrate why the billing unit matters:
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 minute- 1,000 uninterrupted CoreWeave H200 accelerator-hours would be $50,440 on demand or $20,930 at the listed North America spot rate, assuming the quoted line is one accelerator-hour and no other charges apply. Confirm that assumption with the SKU.
- 1,000 AWS P5 Capacity Block instance-hours would cost $41,528 at the displayed rate and provide 8,000 H100 accelerator-hours if all eight GPUs run continuously. Expressed per accelerator-hour, the same arithmetic is $5.191, but buying only part of the node is generally not an option.
- 1,000 Google A3 High eight-GPU machine-hours would be $88,490.000119 on demand or $50,458.920548 on spot before machine, storage, and network additions. A one-GPU A3 spot shape at $6.307365068 per hour has a different capacity and host profile.
Do not subtract storage or egress from a GPU quote simply because a provider bundles some local resources. Measure dataset size, checkpoint frequency, artifact retention, cross-zone traffic, and the amount of output that leaves the region.
Rank #3
- 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
Utilization changes the winner
A full eight-GPU node can make an apparently cheap per-GPU calculation expensive if seven GPUs sit idle. Conversely, a higher per-GPU price may be cheaper when it lets a single-GPU service scale independently. Record GPU utilization, memory utilization, input pipeline stalls, and time spent waiting for distributed collectives. Compare cost per completed training step, inference request, or processed token rather than cost per rented hour alone.
Commitments and reservations
Commitments and Capacity Blocks trade flexibility for a known rate or allocation. Buy them only after measuring a stable baseline and confirming that the term, region, machine shape, and cancellation rules match your forecast. Keep burst capacity on a separate, interruptible path so a reservation is not forced to absorb every peak.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Choosing by workload
| Workload | Primary decision criteria | Questions to answer before selecting a provider |
|---|---|---|
| Single-GPU inference or fine-tuning | Independent scaling, memory fit, startup time, and regional redundancy | Can one-GPU shapes launch reliably? Is model loading faster than the expected request interval? What is the fallback GPU? |
| Eight-GPU distributed training | Interconnect, placement, host balance, checkpoint throughput, and reservation certainty | Is the quoted price for a complete eight-GPU node? Can all workers launch together in the target zone? |
| Large synchronous training | Collective bandwidth, cluster scale, scheduler behavior, and failure recovery | Does the provider offer the required UltraCluster or equivalent topology? How quickly can a failed worker be replaced? |
| Spot-tolerant batch jobs | Spot depth, interruption handling, queueing, and restart automation | How much work can be lost? Can checkpoints be read from another region or provider? |
| Latency-sensitive production service | Region proximity, predictable provisioning, network path, and support | What is the measured p95 launch and recovery time in the customer region? Are egress and cross-zone paths acceptable? |
| Regulated or data-local workloads | Approved regions, isolation, logging, retention, and contractual controls | Does the exact GPU service satisfy the required data-residency and compliance scope? |
How to run an apples-to-apples bakeoff
- Freeze the specification. Select one GPU generation and memory size, one GPU count, a minimum CPU and RAM profile, local storage requirement, interconnect requirement, region, and software image.
- Separate purchase modes. Run independent trials for on-demand, spot, reservation, Capacity Block, and commitment. Do not blend their prices into one average.
- Measure useful work. Use the same model, dataset, batch size, precision, checkpoint interval, and evaluation procedure. Record throughput, time to first result, and cost per completed unit of work.
- Exercise failure paths. Intentionally stop a worker, revoke a spot allocation where the platform permits it, and test restart from a checkpoint. Record lost work and recovery time.
- Measure storage and network. Include dataset staging, checkpoint writes, cross-zone transfers, and final artifact egress. Keep these line items separate from accelerator rental.
- Repeat across time. Run during at least two demand periods and on more than one day. Capacity and launch latency that look good once may not represent the production window.
- Set a decision threshold. Define the maximum acceptable cost, launch latency, interruption loss, and recovery time before looking at results. This prevents a low rental price from overriding an SLO failure.
Common failure modes and fixes
Capacity is unavailable in the preferred region
Verify that the request is for the exact SKU and shape shown as available, then try a pre-approved alternate region or billing mode. If the system is marked contact-sales, obtain a capacity commitment instead of assuming an API request will succeed.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Quota is approved but instances still do not launch
Quota authorizes usage; it does not reserve physical GPUs. Check zone placement, minimum host resources, and whether the requested count can be placed together. Reduce the batch size or use a smaller shape only if the workload’s interconnect requirement still holds.
A spot job loses too much work
Shorten checkpoint intervals, write checkpoints atomically to durable storage, and make restart automatic. Compare the resulting completed-work cost with on-demand or a Capacity Block rather than comparing hourly rates in isolation.
Distributed training is slower than expected
Profile data loading, CPU and RAM pressure, GPU memory use, collective operations, and cross-node traffic. A cheaper accelerator with a weaker host or network can cost more per training step.
The invoice is higher than the GPU estimate
Reconcile machine hours against accelerator hours, then add disks, durable storage, snapshots, inter-zone traffic, egress, idle reservations, and support or managed-compute fees. Google Cloud’s A3 guidance specifically requires adding the machine-type configuration to the GPU price; the same accounting discipline is needed for every provider.
A decision rule you can defend
Select the provider and SKU that meet the workload SLO at the lowest measured cost per useful result. Use CoreWeave’s regional spot or on-demand table when the exact configuration is available and its interruption behavior fits the job. Use an AWS P5 Capacity Block when a defined term, fixed purchase rate, and eight-H100 allocation match the training schedule. Use Google Cloud A3 when its regional machine shape, host configuration, and spot or on-demand behavior fit the pipeline. Use Azure ND H100 v5 when its eight-GPU topology, memory, interconnect, regional controls, and calculator-produced price satisfy the deployment requirements.
Keep a second provider or region tested for recovery if downtime has material business cost. Recheck prices, quota, and capacity before each major training run: the figures above are 2026 snapshots, while availability and rates are live operational properties.
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.
