“Is CoreWeave cheaper than buying an H100?” “How many hours do I need to run a GPU before owning beats renting?” “Should I rent GPUs or build an on-prem AI server?” All three questions have the same answer: compare productive capacity and total cost of ownership, not a cloud hourly rate with a graphics-card price.
Using CoreWeave’s published North America HGX H100 price as a reference, an eight-GPU node costs $49.24 per on-demand hour, or $6.155 per GPU-hour when divided evenly across eight GPUs. Ownership can be cheaper at sustained utilization, but power, cooling, the rest of the server, staffing, financing, downtime and idle capacity can erase that apparent advantage. The method below lets you substitute your own quote, workload and facility costs.
As an Amazon Associate I earn from qualifying purchases.
The answer in one calculation
Define two costs for the same amount of useful work:
- Rental: the effective price per productive GPU-hour, plus storage, data transfer, networking, orchestration, monitoring, support and any commitment.
- Ownership: annualized hardware and facility costs, plus the variable cost of electricity, maintenance, software, financing and operations.
If F is annual fixed ownership cost, V is owned variable cost per productive hour and R is the rental cost per productive hour, the productive-hour crossover is:
#1 Best Overall
- ATTENTION BUYERS: EXPORTING this item outside the U.S. requires strict compliance with the Bureau of Industry and Security (BIS) U.S. Export Administration Regulations (EAR). Non-compliance carries fines up to $1 MILLION PER VIOLATION AND UP TO 20 YEARS IMPRISONMENT for both the buyer and seller. To uphold our regulatory obligations, we strictly follow BIS "Know Your Customer" guidance by screening each order and immediately aborting transactions showing unresolved regulatory red flags. ***
- [Double-Flow-Through Design] The RTX PRO 6000 Blackwell features a double-flow-through cooling design, optimizing efficiency and airflow to sustain peak performance under 600W power loads. | [5th Gen Tensor Cores] Deliver up to 3X the performance of the previous generation and support for FP4 precision for faster AI model processing times with reduced memory usage, enabling local fine-tuning of LLMs and generative AI | [4th Gen Ray Tracing Cores] Double the ray-triangle intersection rate of the previous generation to create photoreal, physically accurate scenes and immersive 3D designs with RTX Mega Geometry, which enables up to 100X more ray-traced triangles.
- [PCIe Gen 5] Support for PCIe Gen 5 provides double the bandwidth of PCIe Gen 4, improving data-transfer speeds from CPU memory and unlocking faster performance for data-intensive tasks like AI, data science, and 3D modeling. | [GDDR7 Memory] With 96 GB of GPU memory and 1.8 TB ps bandwidth, it can tackle massive 3D and AI projects, fine-tune AI models locally, explore large-scale VR environments, and drive larger multi-app workflows.
- [DisplayPort 2.1] Achieve unparalleled visual clarity and performance, driving high resolution displays at up to 8K at 240 Hz and 16K at 60 Hz. Increased bandwidth enables seamless multi-monitor setups while HDR and higher color depth support ensures superior color accuracy for precision work, such as video editing, 3D design, and live broadcasting.
- [NVIDIA Blackwell Streaming Multiprocessor] The new SM features increased processing throughput, and new neural shaders that integrate neural networks inside of programmable shaders | DLSS 4: Multi Frame Generation ensures ultra-smooth frame pacing for lifelike simulations. | [Universal MIG] Divide a single RTX PRO 6000 Blackwell into multiple isolated instances, each with dedicated resources, allowing for concurrent execution of multiple workloads, optimized GPU utilization, and secure isolation of different applications or users. [WARRANTY] 3 YR Manufacturer's Warranty. Bulk OEM Packaging. Retail Packaging is NOT included.
Break-even hours = F ÷ (R − V)
When R is less than or equal to V, ownership never breaks even under those assumptions. When utilization is below the crossover, renting normally costs less because you avoid paying for idle hardware.
Start with an apples-to-apples unit
Match the GPU and the complete system
Choose the same GPU class, VRAM, interconnect, CPU and memory capacity, software stack, geographic region, workload and evaluation period. A single NVIDIA H100 card is not equivalent to an eight-GPU HGX H100 node. CoreWeave’s published $49.24 figure is for the complete eight-GPU configuration, including vCPUs, RAM, local storage and the platform configuration, rather than eight bare cards.
For training, interconnect matters as much as the accelerator. A server with slower links may take longer to finish the same job, so compare cost per completed training run or token, not only cost per clock hour. Keep the comparison to one currency and the same tax treatment.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Clear out junk files and repair common Windows errors3Fix the driver behind crashes, sound loss and screen glitchesNormalize the published rental prices
The current North America card lists these eight-GPU prices:
| Commercial product | Eight-GPU node price | Simple eight-way equivalent | How to use it |
|---|---|---|---|
| On-demand HGX H100 | $49.24 per node-hour | $6.155 per GPU-hour | Baseline for capacity that can be started and released without a long commitment |
| Spot | $19.71 per node-hour | $2.46375 per GPU-hour | Separate interruption-sensitive scenario; do not mix with on-demand results |
| Inference | $6.16 per node-hour | $0.77 per GPU-hour | Separate inference product, not a discount on training capacity |
The per-GPU figures are arithmetic allocations of a node price. Your actual workload may require the whole node, so use the node rate if eight GPUs are needed together. Verify the live price, region, billing granularity and availability before making a purchase or commitment.
Add every rental-side charge
Multiply the effective rate by billed hours, then add persistent storage, checkpoint traffic, ingress and egress, replication, attached CPU or memory, networking, orchestration, observability, support and any minimum commitment. Storage on CoreWeave’s Classic offering is presented as a separate network-attached service, so it should not silently disappear from a compute-only comparison.
For a bursty project, model the hours the instance is actually allocated. For a service that must remain warm, include idle serving hours, replicas, autoscaling headroom and recovery capacity. Spot capacity also needs an interruption allowance: checkpoints, restart time and lost work are real costs even if the spot rate is lower.
Recommended Free Tools
Rank #2
- NVIDIA Ampere Architecture-based CUDA Cores - Double-speed processing for single-precision floating point (FP32) operations and improved power efficiency provide significant performance improvements for graphics and simulation workflows, such as complex 3D computer-aided design (CAD) and computer-aided engineering (CAE), on the desktop.
- Second-Generation RT Cores - With up to 2X the throughput over the previous generation and the ability to concurrently run ray tracing with either shading or denoising capabilities, second-generation RT Cores deliver massive speedups for workloads like photorealistic rendering of movie content, architectural design evaluations, and virtual prototyping of product designs. This technology also speeds up the rendering of ray-traced motion blur for faster results with greater visual accuracy.
- Third-Generation Tensor Cores - New Tensor Float 32 (TF32) precision provides up to 5X the training throughput over the previous generation to accelerate AI and data science model training without requiring any code changes. Hardware support for structural sparsity doubles the throughput for inferencing. Tensor Cores also bring AI to graphics with capabilities like DLSS, AI denoising, and enhanced editing for select applications.
- Third-Generation NVIDIA NVLink - Increased GPU-to-GPU interconnect bandwidth provides a single scalable memory to accelerate graphics and compute workloads and tackle larger datasets.
- 48 Gigabytes (GB) of GPU Memory - Ultra-fast GDDR6 memory, scalable up to 96 GB with NVLink, gives data scientists, engineers, and creative professionals the large memory necessary to work with massive datasets and workloads like data science and simulation.
Calculate the full cost of owning hardware
Annualize the purchase
Use this formula for the hardware portion:
Annualized capex = (purchase price − residual value) ÷ useful life in years
For an hourly view that allocates depreciation only over productive operation, use:
Capex per productive hour = (purchase price − residual value) ÷ useful-life productive hours
Residual value should reflect what you can realistically sell or redeploy at the end of the period. A current quote for a complete server, not just a card, is required. The published figures used here do not establish a market purchase price, useful life or resale value for an owned H100 system.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Measure electricity at the wall
GPU board TDP is not the same as system draw. Measure the server at the rack with a power meter or, for a multi-server installation, a suitable three-phase energy meter. Include power-supply conversion losses, fans or liquid-cooling pumps, CPU and memory draw, networking equipment and the facility’s power-usage effectiveness (PUE).
The direct electricity formula is:
Electricity cost per hour = measured wall draw in kilowatts × electricity price per kWh
Multiply that result by PUE if your electricity figure excludes cooling and facility overhead. Use the tariff that actually applies to the site, including demand charges where relevant.
Rank #3
- 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
Include fixed and variable operating costs
Ownership costs that are easy to omit include:
- Server chassis, motherboards, CPUs, memory, NVMe storage, network adapters, racks and cabling.
- Power distribution, UPS capacity, generators, cooling equipment, floor space or colocation fees and facility PUE.
- Maintenance contracts, replacement fans, disks, power supplies, cables and spare accelerators.
- Operating-system and cluster software, security tooling, monitoring, logging and backup.
- System administration, patching, incident response, compliance work and insurance.
- Interest, lease fees, payment costs, depreciation treatment, tax credits and the opportunity cost of capital.
- Downtime while a component is being diagnosed, shipped or replaced.
Separate costs that occur every year from costs that scale with productive hours. Annual facility rent and support belong in F; electricity, per-job software charges and wear that scales with operation belong in V.
Use productive utilization, not calendar hours
A calendar year has 8,760 hours, but a machine used 60% of the time produces only 5,256 productive hours. Allocate fixed ownership costs over the productive hours you can actually sell or use. Rental billing generally follows the hours an instance is allocated, so releasing it stops the compute charge.
Utilization should include queue gaps, data preparation, software failures, maintenance windows and time waiting for human approval. A nominal “always on” plan with only 30% useful work has 2,628 productive hours per year; the remaining capacity still depreciates and occupies space.
Worked sensitivity example
The following is an illustrative model, not a quoted hardware price or an electricity-market average. It uses a hypothetical $35,000 owned H100-class system, $0.03 per productive hour for maintenance and software, and 1,500 watts of measured wall draw. The rental reference is the published $6.155 per GPU-hour on-demand equivalent, except where the table specifies spot. A committed-rate discount is not stated, so the high case uses the on-demand reference as a conservative stand-in; replace it with your contract rate.
| Variable | Low case | Base case | High case |
|---|---|---|---|
| Productive utilization | 30% | 60% | 90% |
| Hardware life | 2 years | 3 years | 5 years |
| Residual value | 0% | 20% | 30% |
| Electricity | $0.08/kWh | $0.12/kWh | $0.25/kWh |
| Rental mode | Spot | On-demand | Committed/on-demand; committed price not stated |
For each case, annual productive hours are 8,760 multiplied by utilization. Annual fixed capex is the net purchase cost divided by hardware life. Owned variable cost is electricity plus $0.03 per hour; cooling, labor, financing and downtime are intentionally excluded so the effect of the listed assumptions is visible.
| Case | Annual productive hours | Rental rate used | Owned variable cost | Annual fixed capex | Crossover productive hours | Does it occur within the modeled life? |
|---|---|---|---|---|---|---|
| Low | 2,628 | $2.46375/GPU-hour spot | $0.15/hour ($0.12 electricity + $0.03) | $17,500/year | About 7,565 hours | No; that is about 2.88 years at this utilization, longer than the two-year life |
| Base | 5,256 | $6.155/GPU-hour on-demand | $0.21/hour ($0.18 electricity + $0.03) | $9,333/year | About 1,570 hours | Yes; roughly 3.6 calendar months at 60% utilization |
| High | 7,884 | $6.155/GPU-hour on-demand proxy | $0.405/hour ($0.375 electricity + $0.03) | $4,900/year | About 852 hours | Yes; roughly 40 calendar days at 90% utilization |
The low case illustrates why a cheap spot rate can delay ownership payback: the rental price is lower while the machine still incurs depreciation and facility costs during idle time. The base and high cases cross quickly because the modeled utilization spreads capex over many productive hours. Adding a complete server, cooling, labor, financing and downtime raises the ownership cost and moves each crossover to the right.
Turn the formula into a spreadsheet or script
- Enter the comparable unit. Record GPU count, VRAM, interconnect, region, software stack and whether the workload needs a full node.
- Enter rental rates. Keep spot, on-demand and inference in separate columns. Divide a node rate by eight only when a one-GPU equivalent is genuinely useful.
- Estimate productive hours. Use 8,760 × utilization × number of years. Subtract planned maintenance and expected interruption time rather than assuming 100% availability.
- Build rental total. Compute rate × billed hours, then add storage, network, data transfer, orchestration, observability, support and commitments.
- Build ownership total. Add net capex, electricity, cooling, facility, maintenance, software, labor, financing, insurance, spares and downtime.
- Calculate both views. Divide total cost by productive hours for cost per useful hour, and divide by completed jobs or generated tokens for performance-adjusted cost.
- Run sensitivities. Change utilization, power price, hardware life, residual value, rental mode and job throughput. A decision that changes under a small utilization shift is not robust enough for a long commitment.
Costs that commonly invalidate a simple GPU comparison
Whole-node mismatch
Comparing $6.155 for one cloud-equivalent GPU with the price of one bare card omits the host system and the other seven accelerators that may be required for the workload. Conversely, buying a complete eight-GPU server for a job that uses one card creates stranded capacity. Match the smallest deployable unit on both sides.
Rank #4
- Discrete graphics card memory 40 GB
- Memory bandwidth (max) 1555 GB/s
- Graphics processor family NVIDIA
- Graphics processor A100
Power delivery and cooling
High-density servers can require upgraded circuits, distribution, cooling and backup power. The energy bill is only part of the cost; construction, colocation and demand charges can be larger than the GPU’s incremental electricity.
Idle capacity
Owned hardware depreciates and occupies space when no job is queued. Rental usually stops when the instance is released, although persistent disks and reserved services may continue billing. Model the actual queue and serving pattern.
Data and storage
Training repeatedly reads datasets and writes checkpoints. Include persistent capacity, replication, snapshot retention, ingress, egress and the time spent moving data. A lower compute price can disappear if the data path is expensive or slow.
Reliability and operations
Budget for failed components, spare parts, patching, monitoring, security reviews, cluster administration and incident downtime. Cloud support and managed orchestration have an explicit price; an internal team has a payroll and opportunity cost.
Financing and tax
Compare cash purchase, lease or debt on the same basis. Include interest, payment fees, depreciation, tax credits and the return the capital could earn elsewhere. A tax treatment available to one buyer or jurisdiction may not apply to another.
Performance parity
If the owned system trains 20% slower because of its network, storage or software stack, its cost per completed run is higher even when its hourly electricity bill is lower. Record time-to-train, tokens per second, failure rate and usable batch size for both environments.
Free tools Windows power users keep installed
One-click scans. No signup required.
How to interpret CoreWeave’s published TCO claim
CoreWeave’s vendor-hosted summary of a Signal65 study says the comparison covers compute, storage, networking, orchestration, observability and support over three years. It reports “up to 65% lower three-year TCO on NVIDIA GB200 NVL72” in the cloud scenario it compared.
That claim is evidence about the specified cloud comparison, not proof that renting is cheaper than buying for every organization. It does not replace your purchase quote, utilization forecast, electricity tariff, facility costs, staffing model or residual-value assumption. Treat it as a reminder to include the categories in a real TCO model, not as a universal break-even price.
Choosing rental, ownership or a hybrid
Rent when flexibility has economic value
- Demand is experimental, seasonal or difficult to forecast.
- You need more GPUs than your facility can power or cool.
- Capital is scarce, or deployment speed matters more than long-run depreciation.
- Jobs can tolerate spot interruptions and checkpoint frequently.
- You need to test a new GPU generation before committing to a purchase.
Own when utilization is high and predictable
- The workload can keep the system productively busy for most of its useful life.
- You already have suitable power, cooling, networking and staff.
- Data-residency, security or latency requirements make external capacity costly.
- You can finance the complete system and have a credible resale or redeployment plan.
Use a hybrid model when workloads differ
Keep a right-sized owned cluster for steady baseline demand and rent burst capacity for launches, retraining or temporary experiments. Compare the hybrid’s fixed base plus burst rental against both an oversized purchase and an all-cloud plan. Include the operational cost of managing two environments.
Quick Recap
Practical checks before committing
- Obtain a current quote for the complete owned server, rack, network and support package.
- Measure wall power under your actual training or inference load, not only at idle.
- Confirm cloud billing unit, minimum duration, storage pricing, transfer pricing and regional availability.
- Benchmark one representative job in each environment and record completion time and failure recovery.
- Run low, base and high utilization cases, then test a worst case with higher electricity and lower resale value.
- Revisit the model when GPU prices, power tariffs, workload mix or software efficiency changes.
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.
