Cloud API spend against your own GPUs.

Enter your monthly tokens and the price you pay per million. Pick the GPUs you would run instead. The calculator shows both costs over 36 months, with every assumption stated. It is an estimate, not a quote.

Cloud vs on-prem calculator

Estimate, not a quote

Tokens per month (millions)

Cloud price per million tokens

Your GPUs1× H100 80GB≈ $27,500 · $25–30k4× H100 node≈ $210,000 · $180–240kCustomyour count and cost

Add LLM Deploy fees$25,000 once + $1,500 per node a month

Cloud API a month$10,000

On-prem a month$1,043

Break-even2.8months

Saved over 36 months$322,451

Cloud $10,000 a month. On-prem $1,043 a month. Break-even in month 2.8. Over 36 months: saved $322,451.

Cumulative cost over 36 months (USD)Cloud APIOn-premBreak-evenSaved

Showing the default inputs. The sliders and options need JavaScript to recalculate. The table view has the same numbers.

Hover or press the left and right arrow keys on the chart to read a month.Tap or drag along the chart to read a month.

Table view

Month	Cloud API	On-prem	Difference
0	$0	$27,500	−$27,500
6	$60,000	$29,175	$30,825
12	$120,000	$30,850	$89,150
18	$180,000	$32,524	$147,476
24	$240,000	$34,199	$205,801
30	$300,000	$35,874	$264,126
36	$360,000	$37,549	$322,451

On-prem monthly, itemised

Hardware, amortised over 36 months	$764
Power and cooling (PUE 1.4)	$143
Ops overhead, 15%	$136
Total a month	$1,043

Assumptions

Cloud monthly = tokens × price per million tokens.
Hardware amortised over 36 months, no resale value. GPU prices are midpoints of the ranges shown.
Power: 700 W per GPU, 24/7 (730 h a month), PUE 1.4, $0.20/kWh.
Ops overhead: 15% of amortised hardware plus power.
With fees: $25,000 deployment spread over 36 months, plus $1,500 per node a month, one node per 8 GPUs.
Throughput is not modelled. The pilot confirms the GPU count for your model and load.

What you would pay us.

We size the hardware and source it at cost, with no markup. The "Add LLM Deploy fees" option above adds the production and managed ops prices to the on-prem side.

Pilot
$4,900fixed
- One open-weight model live on your data in 72 hours. The clock starts after access and prerequisites.
Production deployment
from $25,000one-time, hardware separate
- Model selection and GPU sizing.
Managed ops
from $1,500per GPU node / month
- Model and driver updates, monitoring, on-call, SLA tiers.
Confidential compute
By quote
- Run private AI on GPUs you don't own: NVIDIA confidential computing on H100, H200 and Blackwell.