CoreWeave is a specialized cloud provider built around NVIDIA GPU computing for training and serving AI models. It combines GPU instances with high-throughput storage, networking, caching, and orchestration features. The practical comparison with AWS, Azure, and Google Cloud is therefore not simply an hourly-rate contest: GPU model, cluster topology, regional capacity, utilization, data movement, support, commitments, and the rest of your cloud stack determine the result.
CoreWeave can be a strong fit when most of your spending and engineering effort center on large NVIDIA GPU jobs. A hyperscaler can be the better choice when identity, databases, analytics, compliance controls, networking, or existing enterprise agreements matter more than a specialist GPU platform. No source establishes a universal price or performance winner, so the defensible choice is a workload benchmark and an exact quote for your GPU, region, topology, storage, and egress pattern.
As an Amazon Associate I earn from qualifying purchases.
What is CoreWeave?
CoreWeave describes itself as an AI cloud provider that supplies GPU computing, storage, networking, and software for training and running AI models. The company says it was established in 2017 and completed its public listing on Nasdaq in March 2025.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
That positioning makes CoreWeave different from a general-purpose cloud in emphasis rather than in the basic idea of cloud computing. You still provision infrastructure remotely and pay for capacity, but the platform is designed around the demanding characteristics of modern machine-learning workloads: large NVIDIA GPU clusters, fast movement of training data, distributed jobs that span many machines, and inference services that need predictable latency.
#1 Best Overall
- NVIDIA Volta GV100 Architecture — 4,608 CUDA Cores, 640 1st-Gen Tensor Cores delivering 14 TFLOPS FP32 and 112 TFLOPS deep learning performance for AI training, inference, HPC, and scientific computing workloads
- 32GB HBM2 ECC Memory — 900 GB/s Bandwidth — High-bandwidth memory on a 4096-bit bus with ECC error correction provides the memory capacity and throughput required for the largest AI models, simulations, and datasets
- PCIe 3.0 x16 Interface — 250W TDP — Standard PCIe Gen3 connectivity with passive cooling designed for enterprise rack server deployment in HPE ProLiant, Dell PowerEdge, and Supermicro platforms with adequate chassis airflow
- NVLink — Scale to 96GB Unified Memory — Connect two V100 GPUs via NVLink at 300 GB/s bi-directional bandwidth to scale GPU memory from 32GB to 96GB for larger AI training and HPC workloads
- Multi-Precision Computing — Supports FP64 (7 TFLOPS), FP32 (14 TFLOPS), FP16 (112 TFLOPS) and INT8 precision modes for flexible deployment across training, inference, and scientific simulation workloads
CoreWeave reported $5 billion in annual revenue and 168% year-over-year growth in its 2025 shareholder letter. Those are company-reported figures, not independent market estimates, and they describe the provider’s business rather than a guarantee of capacity or performance for an individual customer.
What the CoreWeave platform includes
GPU compute for training and inference
The central product is GPU capacity, including configurations built around NVIDIA L40, L40S, A100, H200, and GB200 NVL72 hardware in the published examples. GPU choice affects memory capacity, memory bandwidth, precision support, interconnect behavior, and the number of devices needed for a job. A single-GPU development workload has a very different cost and scheduling profile from a multi-node training run.
Object and file storage
CoreWeave’s SEC filing describes purpose-built object and file storage for AI workloads. Object storage is useful for datasets, checkpoints, and durable artifacts; file storage can support shared access patterns expected by distributed training jobs. Storage performance matters when workers repeatedly read samples or write checkpoints, because a cheap GPU sitting idle while it waits for data is still an expensive GPU.
Recommended Free Tools
GPU-local caching
The platform also describes GPU-local caching. Keeping frequently accessed data closer to the accelerators can reduce repeated transfers from shared storage and improve effective throughput. The benefit depends on the workload’s data-reuse pattern, cache size, preprocessing pipeline, and whether the cache remains warm between jobs.
Networking and cross-cloud reach
CoreWeave presents networking and cross-cloud connectivity as part of the AI infrastructure stack. Cross-cloud reach can help a team place compute near an existing data estate or use a specialist GPU provider while retaining services elsewhere. It also introduces design questions about bandwidth, latency, transfer charges, routing, identity, and failure domains; those costs must be measured rather than assumed away.
Software for the AI lifecycle
CoreWeave groups compute, storage, networking, and software for the lifecycle from training to inference. When evaluating the software layer, inspect the scheduler and orchestration experience, container and Kubernetes support, image distribution, observability, checkpoint handling, autoscaling, and controls for inference endpoints. A lower GPU rate does not help if engineers spend more time adapting jobs or operating missing platform services.
CoreWeave published GPU prices
CoreWeave’s pricing page lists regional on-demand and spot configurations. The following examples come from the North America table and are listed hourly. Prices, regions, and availability can change, so treat them as published examples rather than a standing quote.
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 | Configuration | On-demand price | Spot price |
|---|---|---|---|
| NVIDIA L40 | 8 GPUs | $10.00 per hour | $6.27 per hour |
| NVIDIA L40S | 8 GPUs | $18.00 per hour | $7.88 per hour |
| NVIDIA A100 | 8 GPUs | $21.60 per hour | $9.65 per hour |
| NVIDIA H200 | 8 GPUs | $50.44 per hour | $20.93 per hour |
The same page lists an NVIDIA GB200 NVL72 configuration with four GPUs at $42.00 per hour on demand. The published examples also include entries marked for contacting sales, so not every configuration has a public number.
For scale, multiplying the listed hourly rates by a hypothetical 730-hour month gives $7,300 for the L40 on demand configuration, $13,140 for L40S, $15,768 for A100, and $36,821.20 for H200. Those are arithmetic illustrations of continuous reservation for 730 hours; they exclude storage, networking, egress, support, taxes, discounts, idle time, and any interruption or availability conditions attached to spot capacity. A real monthly bill depends on actual runtime and the terms of the selected region and contract.
How AWS, Azure, and Google Cloud differ
AWS, Azure, and Google Cloud offer GPU virtual machines inside much broader cloud platforms. Their advantage is often the surrounding ecosystem: identity, databases, analytics, networking, security controls, observability, and enterprise procurement. Their GPU economics and available models still require a provider-, region-, and purchase-option check.
| Decision axis | CoreWeave | AWS, Azure, or Google Cloud |
|---|---|---|
| Primary orientation | AI- and GPU-focused cloud stack | Broad cloud platforms with GPU virtual-machine families |
| GPU price visibility | Regional on-demand and spot tables; some configurations require contacting sales | Published catalogs and pricing components; exact SKU and region selection remains necessary |
| Software breadth | Specialized AI compute, storage, networking, and inference tooling | Broader databases, identity, analytics, enterprise integration, and application services |
| Capacity questions | Reserved capacity, cluster size, interconnect, and support contract | Quota, reservation or commitment, region, and fit with the existing cloud estate |
| Cost method | GPU rate plus storage, egress, orchestration, and idle time | VM and GPU charges plus storage, networking, commitments, and spot or preemptible risk |
AWS GPU pricing context
AWS announced up to a 45% price reduction for selected EC2 NVIDIA GPU-accelerated P4 and P5 instances in 2025. The announcement identifies P5 as the H100 family and P5en as the H200 family. The reduction is not a universal AWS GPU price: verify the exact instance, region, operating system, tenancy, purchase option, and current catalog before using it in a business case.
Google Cloud GPU pricing context
Google Cloud states that GPU charges are added to the VM machine type and vary by region. Its pricing model includes on-demand, sustained-use or committed-use, and spot options. Google lists A3 H100 and A3 Ultra H200 accelerator-optimized virtual machines, with hourly prices and commitment or spot variants. Google summarizes the billing relationship this way: "Each GPU adds to the cost of your instance in addition to the cost of the machine type." Compare the complete VM price, not just an accelerator line item.
Azure pricing context
Azure’s current price depends on the exact GPU SKU, region, operating system, purchase model, and available quota. A current Azure price was not established in the material used for this comparison, so an Azure number would be misleading. Obtain a quote for the same GPU class, topology, runtime, storage, and network pattern used for the other providers.
Is CoreWeave cheaper than AWS, Azure, or Google Cloud?
There is no defensible universal answer. A listed CoreWeave H200 rate cannot be compared fairly with an AWS, Azure, or Google Cloud number unless all of the following match:
- The same accelerator generation and memory profile.
- The same number of GPUs and the same multi-node topology.
- The same region and availability assumptions.
- The same purchase type: on-demand, reserved or committed, and spot or preemptible.
- The same operating-system, host, and VM charges.
- The same storage throughput, checkpoint frequency, data-transfer path, and egress.
- The same utilization, including time spent provisioning, waiting in a queue, or sitting idle.
- The same support, software, licensing, and contract terms.
CoreWeave’s public table makes initial GPU-rate screening straightforward. Hyperscaler catalogs can become more economical when a team uses commitments, sustained-use discounts, existing credits, or integrated services. Conversely, a nominally cheaper VM can lose its advantage if quota delays, data movement, lower effective utilization, or engineering work reduce completed training steps per dollar.
Which cloud is best for H100 or H200 training?
H100 and H200 training is primarily a capacity, topology, and utilization decision. The best provider is the one that can deliver the required cluster with the required interconnect, keep the data pipeline fed, and provide acceptable economics under your job’s failure and checkpoint behavior.
Rank #3
- Colour: brown
- Brand: Nvidia
- Packed with features
- Best product in its class
When CoreWeave is a strong candidate
- Your workload is dominated by large NVIDIA GPU training or inference runs.
- You need an AI-specialized platform rather than a broad catalog of unrelated application services.
- You can use the published H200 configuration or negotiate the capacity and topology your run requires.
- Your team values a single stack for GPU compute, AI-oriented storage, networking, caching, and inference operations.
- You are prepared to validate regional availability, support response, and contract commitments.
When AWS may be a better fit
- Your organization already operates heavily in AWS and depends on its identity, networking, data, and monitoring services.
- The exact P5 H100 or P5en H200 instance is available in your region with sufficient quota.
- An AWS commitment, enterprise agreement, or announced reduction for a qualifying instance changes the total-cost calculation.
- You need to keep training data and downstream production services within the same established environment.
When Google Cloud may be a better fit
- A3 H100 or A3 Ultra H200 capacity is available in the required region.
- You can benefit from Google’s machine-learning and data tooling alongside the accelerator VM.
- Sustained-use, committed-use, or spot pricing matches the duration and interruption tolerance of your jobs.
- Your data pipeline already uses Google Cloud services, reducing transfer and integration work.
When Azure may be a better fit
- Your identity, governance, data, and networking controls are centered on Azure.
- The required GPU SKU and quota are available where your data or users must remain.
- Enterprise procurement or compliance requirements favor Azure even when a specialist provider has a lower visible GPU rate.
How to calculate the real workload cost
Use completed work as the unit of comparison, not only dollars per GPU hour. A useful first model is:
Total cost = GPU and VM runtime + storage + data transfer and egress + orchestration and software + support and commitment costs + idle or failed-job time
Then divide by a workload result such as completed training steps, tokens processed, model checkpoints produced, or successful inference requests. Record the same measurements on every provider.
Free tools Windows power users keep installed
One-click scans. No signup required.
Training cost worksheet
- Specify the model, dataset, precision, batch size, sequence length, checkpoint interval, and target number of steps.
- Specify the GPU count, node count, network topology, and maximum acceptable job interruption.
- Measure data-read throughput, preprocessing time, GPU utilization, synchronization wait, checkpoint time, and restart time.
- Record the exact hourly rates and all attached VM, storage, and network charges for the chosen region and purchase model.
- Run enough work to capture warm-up, steady state, checkpoint, and failure-recovery behavior.
- Compare cost per completed step or token and the elapsed time to the target model quality.
Inference cost worksheet
- Define request rate, prompt and output length, latency target, batch policy, concurrency, and availability target.
- Measure tokens per second, time to first token, tail latency, GPU memory headroom, and replica scale-up time.
- Include idle capacity required to absorb peaks and the storage or network cost of model artifacts.
- Compare cost per million tokens or per successful request, not just cost per running GPU.
Capacity, networking, and storage checks
A provider can have an attractive price and still fail the project if the needed cluster is unavailable. Ask each provider for written answers to these questions:
- How many GPUs can be delivered together in the target region and availability zone?
- What interconnect is available between GPUs and nodes, and what topology will the job receive?
- Can capacity be held for the training window, and what happens if a reservation is delayed?
- What storage throughput and shared-file semantics are available for the dataset and checkpoints?
- Can data be cached close to the GPUs, and how is cache warming handled after a restart?
- What are the ingress, egress, cross-cloud, and inter-region charges?
- Which scheduler, Kubernetes integration, image registry, metrics, and logging paths are supported?
- What support response and escalation path apply to a failed multi-node run?
For distributed training, network behavior is part of compute performance. A cluster that spends a large fraction of each step synchronizing or waiting for storage may deliver fewer useful steps per dollar than a more expensive cluster with better effective utilization.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Spot, on-demand, and commitments
On-demand capacity
On-demand is the simplest baseline for a benchmark because the price is visible and there is no long-term commitment. It is usually the right starting point for measuring a new workload, validating software, or running jobs whose schedule cannot tolerate capacity changes. The published CoreWeave examples above are on-demand prices for the listed configurations.
Spot or preemptible capacity
Spot pricing can materially reduce the hourly rate, as the CoreWeave examples show, but the economic benefit depends on interruption behavior, checkpoint frequency, restart time, and the availability of replacement capacity. Test a realistic interruption: stop a worker, restore from the latest checkpoint, and measure lost work and recovery time before assigning a production deadline to spot capacity.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Reserved and committed capacity
Commitments can lower the effective rate when utilization is predictable, but they create a utilization risk. Compare the discounted price with the cost of unused hours, the term length, transfer restrictions, and the probability that your GPU requirement will change. A flexible on-demand baseline is useful even after negotiating a commitment.
Rank #4
- Bulk Pack without retail box
A practical provider-selection process
- Classify the workload. Separate interactive development, batch training, distributed training, fine-tuning, and latency-sensitive inference.
- Set hardware and topology requirements. Record GPU memory, GPU count, node count, interconnect, region, storage throughput, and quota.
- Build an apples-to-apples price sheet. Include GPU and VM charges, storage, data transfer, support, commitments, and expected idle time.
- Run a representative benchmark. Use the production container, dataset shard, precision, batch settings, and checkpoint process.
- Test failure and recovery. Measure checkpoint restoration, worker replacement, queue delays, and lost work under the intended purchase model.
- Evaluate operational fit. Score identity integration, networking, observability, scheduling, support, compliance, and the team’s existing skills.
- Request an exact quote. Ask for the GPU model, topology, region, term, minimum commitment, storage, egress assumptions, and support level in writing.
- Choose a primary and fallback. Document how data, images, checkpoints, and infrastructure definitions can move if capacity or pricing changes.
Common mistakes and how to avoid them
Comparing a GPU rate with a complete instance rate
Some providers add GPU charges to a VM machine type, while a specialist table may present a configured GPU instance. Rebuild each quote into the same components before comparing.
Ignoring the region
GPU availability and price are regional. A rate in a North America table does not establish the price or capacity in another geography. Put region and availability zone beside every number in your spreadsheet.
Pricing a cluster without measuring utilization
Eight GPUs at high utilization can be cheaper per completed step than fewer GPUs that spend time waiting for data or synchronization. Measure the pipeline and report useful work, not only allocated capacity.
Leaving egress out of the model
Cross-cloud designs can move data charges from compute to networking. Count dataset ingress, checkpoint egress, replication, model distribution, and traffic to production consumers.
Assuming public capacity equals guaranteed capacity
A published configuration does not guarantee that the required number of GPUs will be available for your launch window. Confirm quota, reservation mechanics, lead time, and escalation procedures with sales or support.
Using spot without a recovery test
Low spot prices are meaningful only if the job can resume safely. Verify checkpoint integrity, restart automation, duplicate-worker handling, and the amount of work lost after an interruption.
Bottom line: who should use CoreWeave?
Choose CoreWeave when your main requirement is reliable access to large NVIDIA GPU training or inference capacity and an AI-focused stack is more valuable than a broad general-cloud catalog. Its published regional GPU tables make initial rate comparisons easy, and its platform covers storage, networking, caching, and software intended for the AI lifecycle.
Choose AWS, Azure, or Google Cloud when existing enterprise integration, identity, compliance, data services, commitments, or operational familiarity outweigh the benefits of a specialist provider. For H100 and H200 projects, make the final decision from a benchmark and a provider quote that match the exact GPU, topology, region, utilization, storage, and egress pattern. That process is more reliable than declaring any cloud the cheapest in the abstract.
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.
