Cloud GPU Cost Calculator
Pick a GPU and how long you need it. See what the job costs at eight GPU providers, cheapest first, and whether buying would pay off.
Pay by the hour, stop any time.
24 if it runs round the clock.
A training run of 1,000 GPU-hours takes 125 hours on 8 GPUs.
Setting up, debugging, or leaving it on by mistake.
If you serve a model: gives the cost per million tokens.
Models, datasets and checkpoints on persistent storage.
Egress. Charged only where a provider lists a price.
Adds a “Your quote” row to compare.
GPU clouds bill in US dollars. Add today's rate to see another currency.
Every provider for this job
Cheapest total first, in US dollars. Tap a provider to see its numbers above. Prices as listed by each provider; checked on 10 October 2026. AWS and Azure will be added once their prices are confirmed.
Edit prices (GPU-hour, disk and egress)
Got a different price or a discount? Type it in, in US dollars. Disk ($ per GB-month) and egress ($ per GB) are blank where the provider doesn't list one. GPU prices reset when you change the GPU or price type.
Buy or rent?
How many months of renting it takes to pay for your own GPU. Purchase prices are typical, not quotes.
CPU, memory, case, networking.
Leave empty to use your schedule above.
How your GPU bill is worked out
GPU-hours × the price of one GPU-hour, plus disk and downloads, then compared with the dearest option.
- 1
Count GPU-hours
1 GPU × 8 h × 30 d
= 240 GPU-h - 2
Find the price
DeepInfra
$2.20 a GPU-hour - 3
Rent the GPUs
240 GPU-h × $2.20
= $528 - 4
Add the extras
$528 + $0 disk
+ $0 downloads
= $528
$1,392 less than Fireworks AI
Read more: the formula, a worked example and every price used
GPU clouds charge for each GPU by the hour (most bill by the second or minute). Your bill is the number of GPU-hours you pay for times the price of one GPU-hour, plus disk you keep and data you download where those are charged. The same GPU can cost several times more at one provider than another, so it pays to compare. Three kinds of provider are listed: rent a GPU (RunPod, Lambda, Jarvislabs, Nebius) gives you a machine to do anything with; dedicated inference (Together AI, DeepInfra, Fireworks AI) runs GPUs for serving AI models on their platform; and a big cloud (Google Cloud) rents whole virtual machines.
The formula
hours on each GPU = hours a day × days a month × months
(or: GPU-hours for the job ÷ number of GPUs)
hours paid = hours × (1 + idle share)
GPUs paid = number of GPUs rounded up to the machine size
compute = price per GPU-hour × GPUs paid × hours paid
total = compute + disk GB × price per GB-month × months + GB downloaded × price per GB
break-even months = buy price × (1 + rest of machine) ÷ (rent a month − power a month)
cost per million tokens = price per GPU-hour ÷ (tokens a second × 3,600 ÷ 1,000,000)
Worked example
One NVIDIA H100 SXM for 8 hours a day, 30 days, on-demand, is 1 × 8 × 30 = 240 GPU-hours. The cheapest listed price is DeepInfra (dedicated inference) at $2.20 a GPU-hour: 240 × $2.20 = $528. The cheapest place to rent a whole GPU machine is Jarvislabs at $3.49: 240 × $3.49 = $837.60. The most expensive option that rents a single GPU is Fireworks AI at $8: 240 × $8 = $1,920. Against that, DeepInfra saves $1,392 (73%). Google Cloud only rents this GPU in 8-GPU machines, so one GPU there means paying 8 × 240 × $11.06 = $21,235.20; it stays in the table but isn't used for the savings.
Buying one instead: $30,000 + 20% for the rest of the machine = $36,000. Power: 700 W + 30% = 0.91 kW × 240 hours = 218.4 kWh × $0.15 = $32.76 a month. Renting at $2.20 costs $528 a month, so owning saves $495.24 a month and pays for itself after $36,000 ÷ $495.24 = 72.7 months (6.1 years) of this much use.
On-demand, spot and reserved
- On-demand: pay by the hour, stop any time. The prices in the table.
- Spot or preemptible: spare capacity at a big discount, but the provider can stop your machine at short notice. Save checkpoints often. Together AI, Nebius (dynamic “from” prices) and Google Cloud publish these prices; Jarvislabs advertises spot “up to 56% off” without a list price. Providers with no published price are shown at their on-demand price and marked “on-demand only”.
- Reserved: a lower price for promising to pay for a set time, whether you use it or not. Together AI lists reservations of 3–30, 31–90 and 91–180 days; the 91–180-day price is used here. Nebius offers commitments “up to 35% less” and Google Cloud committed-use discounts of up to 37% (1 year) or 55% (3 years), without list prices, so they are not modelled. Others don't publish one; ask for a quote and type it into Edit prices.
Prices used ($ per GPU-hour, on-demand)
| GPU | RunPod | Lambda | Jarvislabs | Nebius | Together AI | DeepInfra | Fireworks AI | Google Cloud |
|---|---|---|---|---|---|---|---|---|
| NVIDIA B300 | — | — | — | $9.50 | — | $4.89 | — | — |
| NVIDIA B200 | $7.99 | $6.69 | — | $8.50 | $8.19 | $3.69 | $13.00 | $8.06 (×8) |
| NVIDIA H200 | $5.29 | — | $4.59 | $5.40 | $5.99 | $2.69 | $8.00 | $10.60 (×8) |
| NVIDIA H100 SXM | $3.99 | $3.99 | $3.49 | $4.50 | $3.99 | $2.20 | $8.00 | $11.06 (×8) |
| NVIDIA H100 NVL | $3.19 | — | — | — | — | — | — | — |
| NVIDIA H100 PCIe | $2.89 | — | — | — | — | — | — | — |
| NVIDIA A100 80GB | $1.79 | $2.79 | $1.49 | — | — | $0.89 | — | $5.07 |
| NVIDIA A100 40GB | — | $1.99 | $0.89 | — | — | — | — | $3.67 |
| NVIDIA RTX PRO 6000 Blackwell | — | — | $2.19 | — | — | — | — | — |
| NVIDIA L40S | $1.09 | — | — | from $1.55 | — | — | — | — |
| NVIDIA RTX 5090 | $1.19 | — | — | — | — | — | — | — |
| NVIDIA RTX 4090 | $0.89 | — | — | — | — | — | — | — |
| NVIDIA RTX A6000 | $0.59 | — | — | — | — | — | — | — |
| NVIDIA L4 | $0.59 | — | $0.44 | — | — | — | — | $0.71 |
(×8) means the GPU is only rented in 8-GPU machines. DeepInfra and Fireworks AI list “H100 80GB” without saying SXM or PCIe; it is shown as H100 SXM. Prices as listed by each provider; checked on 10 October 2026. Google Cloud prices are the whole VM's price divided by its GPUs, so they include its CPU and memory. AWS and Azure will be added once their prices are confirmed on their own pages.
Providers, discounts and extras
| Provider | Type | Spot / preemptible | Reserved | Disk / GB-month | Egress / GB |
|---|---|---|---|---|---|
| RunPod | Rent a GPU | not listed | not listed | $0.07 | not listed |
| Lambda | Rent a GPU | not listed | not listed | not listed | free |
| Jarvislabs | Rent a GPU | not listed | not listed | $0.10 | not listed |
| Nebius | Rent a GPU | B300 from $0.99, B200 from $0.99, H200 from $0.79, H100 SXM from $0.79 | not listed | $0.0745 | free |
| Together AI | Dedicated inference | B200 $4.09, H200 $2.99, H100 SXM $1.99 | B200 $6.79, H200 $3.99, H100 SXM $3.19 | not listed | not listed |
| DeepInfra | Dedicated inference | not listed | not listed | not listed | not listed |
| Fireworks AI | Dedicated inference | not listed | not listed | not listed | not listed |
| Google Cloud | Big cloud | B200 $4.95, H200 $6.35, H100 SXM $6.29, A100 80GB $3.04, A100 40GB $2.20, L4 $0.40 | not listed | not listed | not listed |
- RunPod (pricing page): On-demand pods. Disk: network storage $0.07/GB-month under 1 TB; container disk $0.10; volume $0.10 running, $0.20 idle. Egress not stated.
- Lambda (pricing page): On-demand instances in 1, 2, 4 or 8 GPUs. No egress fees. B200 is the SXM6 180 GB version.
- Jarvislabs (pricing page): Storage $0.10/GB-month. Spot instances “up to 56% off” have no list price, so they are not modelled.
- Nebius (pricing page): Prices from 1 October 2026 (HGX B300, B200, H200, H100). Preemptible prices are dynamic “from” prices. Commitments “up to 35% less” (no list price). Shared filesystem $0.08 per GiB-month; egress free.
- Together AI (pricing page): GPU Clusters. Reserved for 3–30, 31–90 or 91–180 days; the 91–180-day price is used here.
- DeepInfra (pricing page): Dedicated GPUs for serving models. Lists “H100 80GB” without saying SXM or PCIe.
- Fireworks AI (pricing page): On-demand deployments for serving models. Lists “H100 80GB” without saying SXM or PCIe.
- Google Cloud (pricing page): us-central1, whole-VM price ÷ GPUs (includes CPU and memory). B200 (a4-highgpu-8g), H200 (a3-ultragpu-8g) and H100 (a3-highgpu-8g) only in 8-GPU VMs; A100 in 1, 2, 4 or 8; L4 is g2-standard-4. Committed use up to 37% off (1 year) or 55% (3 years).
Typical prices to buy (used for “Buy or rent?”)
| GPU | Memory | Power | Typical price |
|---|---|---|---|
| NVIDIA B300 | 270 GB | 1100 W | $50,000 |
| NVIDIA B200 | 180 GB | 1000 W | $40,000 |
| NVIDIA H200 | 141 GB | 700 W | $35,000 |
| NVIDIA H100 SXM | 80 GB | 700 W | $30,000 |
| NVIDIA H100 NVL | 94 GB | 400 W | $28,000 |
| NVIDIA H100 PCIe | 80 GB | 350 W | $25,000 |
| NVIDIA A100 80GB | 80 GB | 400 W | $12,000 |
| NVIDIA A100 40GB | 40 GB | 400 W | $8,000 |
| NVIDIA RTX PRO 6000 Blackwell | 96 GB | 600 W | $8,500 |
| NVIDIA L40S | 48 GB | 350 W | $8,000 |
| NVIDIA RTX 5090 | 32 GB | 575 W | $2,500 |
| NVIDIA RTX 4090 | 24 GB | 450 W | $1,800 |
| NVIDIA RTX A6000 | 48 GB | 300 W | $4,500 |
| NVIDIA L4 | 24 GB | 72 W | $2,500 |
Purchase prices are rough street prices for one card; data-centre GPUs (B200, H200, H100 SXM) are normally sold only in 8-GPU servers, shown here as a share of one. The break-even leaves out resale value, repairs, rack space, your time, and the fast networking and storage a cloud includes.
Ways to pay less
- Use spot or preemptible GPUs for work that can restart from a checkpoint.
- Pick the smallest GPU that fits your model in memory: an L40S or RTX 4090 is far cheaper than an H100 for small models.
- Turn machines off when idle. Most bills grow from GPUs left running overnight.
- Keep data where the GPUs are: where egress is charged, downloading terabytes adds up fast.
Common questions
How much does it cost to rent a GPU?
GPU clouds charge for each GPU by the hour, and most bill by the second or minute. Your bill is the GPU-hours you pay for times the price of one GPU-hour, plus disk you keep and data you download where those are charged. The same GPU can cost several times more at one provider than another, so it pays to compare.
What is the cheapest cloud GPU provider?
It depends on the GPU and how you rent it. The calculator lists every provider for your GPU and job, cheapest first. Some providers only rent certain GPUs in 8-GPU machines, so a single-GPU job there costs far more than the hourly price suggests.
What is the difference between on-demand, spot and reserved GPUs?
On-demand means you pay by the hour and can stop any time. Spot or preemptible GPUs are spare capacity at a big discount, but the provider can stop your machine at short notice, so save checkpoints often. Reserved gives a lower price for promising to pay for a set time, whether you use it or not.
Is it cheaper to buy or rent a GPU?
It depends on how much you use it. The calculator takes the buy price plus 20% for the rest of the machine and divides it by what owning saves each month after power. The break-even leaves out resale value, repairs, rack space, your time, and the fast networking and storage a cloud includes.
Do I need an H100 for my AI model?
Not always. Pick the smallest GPU that fits your model in memory: an L40S or RTX 4090 is far cheaper than an H100 for small models. Our LLM VRAM calculator shows how much memory a model needs to run or fine-tune.
How can I cut my cloud GPU bill?
Use spot or preemptible GPUs for work that can restart from a checkpoint, and turn machines off when idle, since most bills grow from GPUs left running overnight. Keep data where the GPUs are too, because where egress is charged, downloading terabytes adds up fast.
Related calculators
Made by Zulu Calculator, part of AlphaZulu LLC. Formulas are shown in full and rule figures are checked against official sources; see how we build calculators.