GPU Rental Cost Calculator
Compare GPU cloud rental vs ownership 2026: RTX 4090, A100, H100 rates on Vast.ai, RunPod & Lambda. Find your break-even month. Free online.
Calculator Free · Private
Your recent calculations
Data updated:
Reviewed by: Martín Rodríguez (editorial policy ) · Last reviewed:
Result
Monthly Rental Cost
Monthly Ownership Cost (amortization + electricity + overhead)
Monthly Electricity Cost (if owning)
Monthly Amortization
Break-Even Month (owning beats renting)
Annual Rental Cost
Recommendation
See step-by-step calculation
Was this calculation helpful?
Thanks! Your feedback helps us improve.
Your result
—
Every US-based ML practitioner faces the same choice: rent a GPU at $0.40–$2.50/hr on Vast.ai, RunPod, or Lambda Labs, or buy hardware for $1,800–$30,000+? This calculator models both sides — rental rates from leading US providers (vast.ai HQ in Berkeley CA, RunPod HQ in Wilmington DE, Lambda Labs in San Francisco), ownership costs using the U.S. Energy Information Administration (EIA) national average electricity rate of $0.16/kWh, and IEEE-cited TDP specs for RTX 4090, A100, H100, and more. Output: USD monthly cost comparison and break-even month for any US setup. Compare also to AWS p4/p5 and GCP A3 hyperscaler rates.
When to use this calculator
- Solo ML Engineer — RTX 4090 Rental vs. Workstation Build — Maya runs fine-tuning jobs averaging 6 hours/day, 22 days/month on a rented RTX 4090 at $0.42/hr on RunPod. Her monthly rental bill: roughly $55.44. A new RTX 4090 costs $1,899, her electricity rate is $0.13/kWh, and she estimates 8% overhead. Amortized over 24 months, her monthly ownership cost comes to about $46. Break-even: month 22. Marginal — but she values the flexibility of scaling to 4× GPUs on demand, so she keeps renting for now.
- AI Startup — Inference Budget for 2026 Roadmap — A 4-person startup serves an LLM inference API using two A100 80GB instances on Lambda Labs at $2.00/hr each, running 20 hours/day, 30 days/month. Monthly cloud bill: $2,400. Buying two A100s at $10,000 each plus a $3,000 server chassis, amortized over 36 months with $0.10/kWh electricity and 15% overhead, yields a monthly ownership cost of roughly $870. Break-even arrives around month 13. The CFO greenlights the hardware purchase.
- Academic Researcher — Spot Instances for Batch Training — Carlos runs NLP experiments using Vast.ai spot RTX 4090 instances at $0.22/hr (off-peak), 4 hours/day, 20 days/month. His monthly cost: $17.60. At that usage level, owning an RTX 4090 wouldn't break even for over 60 months — well past the card's useful life. The calculator confirms spot renting is the right call for his light, intermittent workload.
- Stable Diffusion Power User — Used GPU Economics — Jordan wants to run Stable Diffusion XL locally 8 hours/day, 30 days/month. A used RTX 3090 is available for $650, TDP 350W, electricity at $0.16/kWh, 6% overhead, 24-month amortization. Monthly ownership cost: ~$42. Comparable cloud rental (RTX 3090 on Vast.ai at $0.25/hr) would run $60/month. Break-even: month 18. Jordan buys the used card.
- Enterprise Team — H100 SXM Cluster Decision — A computer vision team needs 8× H100 SXM GPUs running 16 hours/day, 30 days/month. Cloud cost at $3.50/hr per GPU: $13,440/month. Purchasing 8× H100s at $28,000 each ($224,000 total), amortized over 48 months with $0.09/kWh datacenter power and 18% overhead: roughly $6,200/month. Break-even: month 18. The team secures a capital lease and cuts annual compute spend by over $87,000.
- Freelance Data Scientist — Variable Workload Planning — Priya's GPU usage swings from 2 hours/day during slow weeks to 10 hours/day during crunch periods. She runs the calculator at both extremes. At 2 hrs/day with a rented RTX 4090 at $0.40/hr: $24/month. Ownership at that pace breaks even past month 55. At 10 hrs/day: rental hits $120/month and ownership breaks even at month 11. The calculator helps her set a threshold: if she expects to average 7+ hours/day consistently, it's time to buy.
- Colocation vs. Home Lab — Overhead Rate Comparison — A hobbyist debates placing his RTX 4090 in a home closet (PUE ~1.2, 5% overhead) versus a local colo facility ($80/month rack fee, PUE 1.4, 12% overhead). By running both scenarios in the calculator and adjusting the electricity rate to reflect each PUE, he discovers the home setup saves $28/month in total ownership cost — but loses on reliability and thermal management for long training runs.
- Budget Forecasting — Comparing Three Providers Side by Side — A team lead runs the calculator three times for an A100 40GB workload at 12 hours/day, 30 days/month: Vast.ai spot at $0.80/hr ($288/mo), RunPod on-demand at $1.19/hr ($428/mo), and Lambda Labs reserved at $1.10/hr ($396/mo). Against an ownership cost of ~$310/month (amortized over 36 months), Vast.ai spot is the only rental option that stays cheaper — but only if uptime SLA isn't critical to the pipeline.
GPU Rental & Ownership Reference (2026 US spot pricing)
| GPU | Rental $/hr (spot) | TDP (W) | New/Used Price (USD) | Typical Break-even (8h/day, $0.16/kWh) |
|---|---|---|---|---|
| RTX 4090 | ~$0.40 | 450 | $1,800 new | ~19–39 months |
| RTX 3090 | ~$0.20 | 350 | $700 used | ~28–55 months |
| A100 40GB PCIe | ~$1.10 | 250 | $10,000 | ~13–18 months |
| A100 80GB PCIe | ~$1.40 | 300 | $12,000 | ~14–20 months |
| H100 SXM | ~$2.50 | 700 | $30,000 | ~16–22 months |
Fuente: Vast.ai, RunPod, Lambda Labs (spot rates, 2026); electricity: U.S. EIA national average $0.16/kWh (2026); TDP: NVIDIA datasheets / IEEE Spectrum
How it works
What is GPU rental cost?
GPU rental cost is the hourly or monthly fee charged by cloud providers like Vast.ai and RunPod for accessing graphics processors remotely. Prices range from $0.40/hour for an RTX 4090 to $2.50/hour for an H100. Renting eliminates upfront hardware investment, making it cost-effective for short-term projects or variable workloads.
How It Works
This calculator compares two cost models for GPU compute: cloud rental (pay-per-hour on platforms like Vast.ai, RunPod, or Lambda Labs) versus hardware ownership (purchase + electricity + overhead amortized over a set period).
---
Formulas
// Rental side
hours_per_month = hours_per_day × days_per_month
monthly_rental = hours_per_month × rental_rate_per_hour
// Ownership side
monthly_amortization = purchase_price / amortization_months
monthly_electricity = (tdp_watts / 1000) × hours_per_month × electricity_rate
monthly_overhead = (purchase_price × overhead_fraction) / 12
monthly_ownership = monthly_amortization + monthly_electricity + monthly_overhead
// Break-even
// At month N: cumulative_rental(N) = purchase_price + cumulative_ownership_opex(N)
// cumulative_rental(N) = N × monthly_rental
// cumulative_ownership_opex(N) = N × (monthly_electricity + monthly_overhead)
// purchase_price = N × (monthly_rental - monthly_electricity - monthly_overhead)
// breakeven_month = purchase_price / (monthly_rental - monthly_electricity - monthly_overhead)The break-even formula assumes you pay the full purchase price upfront and your only ongoing ownership costs are electricity and overhead (no amortization debt once you own the hardware).
---
GPU Reference Rates (2026 spot pricing, platforms vary)
| GPU | Rental $/hr | TDP | New/Used Price |
|---|---|---|---|
| RTX 4090 | ~$0.40 | 450 W | ~$1,800 new |
| A100 40GB | ~$1.10 | 300 W | ~$10,000 |
| A100 80GB | ~$1.40 | 400 W | ~$12,000 |
| H100 SXM | ~$2.50 | 700 W | ~$30,000 |
| RTX 3090 | ~$0.20 | 350 W | ~$700 used |
Spot prices fluctuate. Vast.ai prices can be 30–50% lower during off-peak hours.
---
Worked Example
Scenario: RTX 4090, 8 hrs/day, 22 days/month, $0.12/kWh, $1,800 purchase, 24-month amortization, 10% overhead.
In this scenario, renting is cheaper unless you plan to use the GPU for 39+ months continuously.
---
Limitations
Frequently asked questions
What is the cheapest GPU to rent for AI inference in 2026?
For FP16 inference, the RTX 4090 consistently offers the best cost-per-TFLOP among consumer-class GPUs, with Vast.ai spot prices ranging from $0.18/hr (off-peak) to $0.55/hr (peak demand) as of early 2025. Its 24 GB of GDDR6X VRAM handles most 7B–13B parameter models comfortably. For larger models requiring more VRAM — think 70B+ parameter inference or multi-tenant serving — an A100 80GB at ~$1.40–$2.00/hr may be more cost-effective on a per-token basis because it eliminates CPU offloading overhead. H100s deliver faster throughput but at $3–$4/hr, they typically only win for latency-sensitive production APIs or large batch jobs where raw speed reduces total job time enough to offset the higher hourly rate. Always calculate cost-per-output, not just cost-per-hour.
Does the break-even calculation include the cost of the host machine or server chassis?
No — this calculator models the GPU cost only. If you are assembling a dedicated workstation from scratch, add $500–$2,000 for a mid-range PCIe platform (CPU, motherboard, RAM, PSU, case). If you are deploying a datacenter-grade SXM GPU, a server chassis with appropriate NVLink infrastructure can add $3,000–$8,000+. If you already own a compatible machine and are upgrading the GPU, that platform cost is a sunk cost and should not be included. For the most accurate break-even, input only the marginal cost attributable to the GPU itself — which is exactly what this calculator's purchase price field is designed for.
How does Vast.ai pricing compare to RunPod or Lambda Labs?
Vast.ai operates a peer-to-peer marketplace where independent hosts compete on price, producing the lowest average rates — but with variable uptime, host reliability, and no formal SLA. Spot-like pricing means availability can disappear mid-job. RunPod sits in the middle: vetted hardware, more consistent uptime, and rates typically 20–40% above Vast.ai but well below hyperscalers. They offer both spot and on-demand tiers. Lambda Labs targets teams that need enterprise reliability and support, with on-demand pricing closer to RunPod on-demand but more predictable availability. All three are 60–80% cheaper than equivalent AWS p4 or GCP A3 instances. For cost optimization, Vast.ai wins for interruptible batch jobs; RunPod or Lambda wins for continuous inference workloads where job restarts are expensive.
What overhead percentage should I enter for my setup?
The overhead field captures annual maintenance, cooling, and ancillary costs as a percentage of your GPU purchase price. Use these benchmarks as a guide: Home desktop workstation: 5–8% — occasional thermal paste replacement, dust cleaning, minor accessories. Dedicated home server closet: 10–15% — add pro-rata share of a UPS, extra case fans, and potential repair labor. Colocation rack: 15–25% — monthly rack and power fees, redundant cooling, and potential managed services add up quickly. Enterprise on-premises: 20–30%+ — factor in maintenance contracts, spare parts inventory, and IT staff time. When in doubt, start with 10% for a home workstation and 18% for any professional rack deployment. These figures are deliberately conservative — real-world costs often run higher due to unexpected failures.
Why is there a difference between the A100's TDP in the presets and the official spec sheet?
The A100 exists in multiple form factors with significantly different power envelopes. The A100 40GB PCIe has a TDP of 250W; the A100 80GB PCIe is 300W; the A100 40GB and 80GB SXM4 variants reach 400W under full load. Our presets use the PCIe figures as defaults because most cloud rental listings and consumer-accessible hardware use the PCIe form factor. If you are renting or purchasing SXM4-based A100s (common in HGX server nodes), manually increase the TDP to 400W in the custom field to accurately model electricity costs. The difference — 300W vs. 400W at $0.12/kWh for 8 hours/day — adds roughly $3.50/month to electricity cost, which compounds meaningfully over a 36-month amortization period.
Should I use a 24-month or 36-month amortization period?
The right answer depends on the GPU type and your use case. Consumer GPUs (RTX 4090, RTX 3090): use 24 months. NVIDIA typically launches a new consumer generation every 18–24 months, and resale value drops sharply once a successor is announced — often 30–50% within six months of launch. Modeling beyond 24 months overstates the asset value. Datacenter GPUs (A100, H100): use 36–48 months. Enterprise procurement cycles are longer, buyers are less trend-sensitive, and these cards retain value better due to constrained supply. Hobbyist / budget GPUs (RTX 3080, RTX 3060): 24–30 months is reasonable, balancing depreciation against the fact that lower-end cards are used longer before replacement. If you plan to resell the GPU partway through the amortization period, you can model the residual value by shortening the period to your expected ownership horizon.
Does the electricity rate field account for cooling overhead?
No — the formula multiplies GPU TDP by hours used, then applies your entered electricity rate. Cooling and facility overhead are not automatically included. In practice, every watt a GPU consumes must also be cooled, adding 10–40% to total facility power draw depending on your setup. To account for this, apply your environment's Power Usage Effectiveness (PUE) factor before entering the electricity rate. For example, if your rate is $0.13/kWh and your home lab PUE is 1.2, enter $0.156/kWh instead. Typical PUE benchmarks: home desktop: 1.05–1.15; home server closet: 1.15–1.30; modern colocation datacenter: 1.10–1.20; older enterprise datacenter: 1.40–1.80. Using raw TDP without a PUE adjustment will underestimate your true electricity cost.
Can I use this calculator for multi-GPU setups like a 4× or 8× rig?
Yes — the per-unit math is linear, so scaling is straightforward. Option 1 (simplest): Multiply your hours/day by the number of GPUs. If you run 4 GPUs for 8 hours/day, enter 32 hours/day. Option 2: Multiply both the rental rate and purchase price by the number of GPUs before entering them. For example, 8× H100s at $3.50/hr each become a single $28/hr entry. Note that multi-GPU setups may have non-linear overhead (NVLink bridges, higher-wattage PSUs, more complex cooling), so consider bumping your overhead percentage by 3–5 percentage points for rigs with more than two GPUs. For 8-GPU server deployments, also factor the chassis cost separately if it's a meaningful portion of total spend.
What if my spot rental rate is much lower than the presets?
Select Custom / Other GPU and enter your actual spot rate in the Custom Rental Rate field. Vast.ai RTX 4090 spot prices have been observed as low as $0.18–$0.22/hr during off-peak hours (typically 2–8 AM US time), compared to the $0.40/hr on-demand rate. At $0.20/hr for 8 hours/day, 30 days/month, your monthly rental cost drops to just $48 — meaning an RTX 4090 purchase at $1,899 wouldn't break even for over 28 months even before factoring in electricity and overhead. This dramatically changes the rent-vs-buy calculus. The key tradeoff: spot instances can be interrupted, so they're suitable for checkpoint-resumable training jobs but not for continuous inference serving.
How accurate is this calculator compared to my actual cloud invoice?
The calculator models pure compute cost only. Your actual cloud invoice may also include: data transfer / egress fees (typically $0.05–$0.12/GB out on most providers); storage costs for persisting model weights and datasets between sessions; idle reservation fees if you hold a reserved instance but don't use it fully; and taxes or platform fees. For Vast.ai and RunPod, egress is generally minimal for ML workloads that consume data but don't output large files. For Lambda Labs, network egress policies vary by plan. As a practical rule, add 5–15% to your calculated rental cost to estimate your actual all-in invoice. On the ownership side, the calculator also excludes one-time setup costs like cable management, a dedicated circuit, or cooling upgrades — factor those in separately if they apply.
Does renting a GPU make sense for short, one-off experiments?
Almost always yes. If you're running a single training experiment that takes 20–40 hours total, the rental cost at $0.40–$1.50/hr is $8–$60 — far below the carrying cost of owned hardware for such a brief use case. Ownership only wins when you have sustained, predictable usage over many months. A useful rule of thumb from this calculator: if your monthly rental bill is less than your monthly ownership cost (purchase amortization + electricity + overhead), you should keep renting. The break-even month tells you exactly when that relationship flips. For most researchers with usage below 4–5 hours/day, ownership rarely breaks even within the GPU's depreciation window — making cloud rental the rational default for light workloads.
How do I model the scenario where I rent while waiting to buy?
Run the calculator with your expected usage and current rental rate to get your monthly rental cost. Then note the break-even month for ownership. If GPU prices are high today and you expect them to fall (common after a new NVIDIA generation launches), delay the ownership decision and continue renting. For example, if break-even is month 20 and you expect RTX 5090 availability to drop RTX 4090 prices by 30% in 6 months, waiting 6 months and recalculating at the new price point may push break-even down to month 14 — a materially better deal. Use the Custom purchase price field to model different future price points and find your optimal buy timing.
Sources & references
- IEEE — Hardware Specifications and Energy Benchmarks for AI — IEEE Spectrum (2026)
- U.S. Average Retail Price of Electricity — EIA — U.S. Energy Information Administration (2026)
- Vast.ai GPU Marketplace Pricing (Berkeley, CA) — Vast.ai (2026)
- RunPod GPU Instance Pricing (Wilmington, DE) — RunPod (2026)
- Lambda Labs GPU Cloud Pricing (San Francisco, CA) — Lambda Labs (2026)
- AWS EC2 P5/P4d Pricing (Amazon, US-based) — Amazon Web Services (2026)
- NVIDIA A100 Tensor Core GPU Datasheet — NVIDIA Corporation (2023)
Methodology & trust
Editorial
Calculadora de tecnología revisada por el equipo editorial de Hacé Cuentas, contrastada con IEEE — Hardware Specifications and Energy Benchmarks for AI, según nuestra política editorial y metodología.
Updates
Última revisión: June 20, 2026. Los parámetros se verifican periódicamente con las fuentes citadas.
Privacy
Calculations run 100% in your browser. We do not store or transmit your data.
Limitations
Indicative results. For critical decisions, consult a professional.
📌 How to cite this calculator
Rodríguez, M. (2026). GPU Rental Cost Calculator. Hacé Cuentas. https://hacecuentas.com/gpu-rental-cost-vast-runpod-calculator
Contenido bajo licencia CC-BY 4.0 — reutilizable citando la fuente con enlace a Hacé Cuentas.