LLM cost calculator: cloud API vs local GPU

For a given workload, estimate monthly cost on a cloud API offering and the cost of running an equivalent open-weight model on your own GPU hardware, with amortisation, energy and labour made explicit. Not a benchmark: local throughput is estimated from hardware specifications with efficiency factors you can override below. All amounts in USD. EUR display

Fairness controls

Quality impact of fp8 quantisation is not modelled; this comparison is cost only.

Amortisation horizon
Workload
Or compute from requests/day
Cloud

$22.71 / month ($2.27 / Mtok effective)

$0.64 / 1M chars (o200k, blended slice)

Breakdown
Input$4.62
Output$17.60
Cache read$0.50
Cache write$0.00
Batch saving-$0.00
Per-request fees$0.00
Other hosts of this model
Host$/month$/Mtok
o4-mini (OpenRouter)$22.71$2.27
o4-mini (Azure AI Foundry)$22.71$2.27
o4-mini (Azure AI Foundry)$24.99$2.50
o4-mini (Azure AI Foundry)$24.99$2.50
Local hardware
Custom hardware (used only when Hardware = Custom)
Cost assumptions
Advanced: measured throughput overrides
Cloud$22.71 / mo
Local, workload$– / Mtok
Local, capacity$– / Mtok
Cheapercloud

Breakeven and sensitivity

Breakeven volume: – tokens/month. Breakeven utilisation: –.

Monthly cost vs. tokens/month

Cloud cost scales linearly with volume; local cost steps up each time another box is needed. Log scale on tokens/month.

Assumptions

KeyValueSourceOverridden
hoursPerMonth730Units: a month is 730 hours (8,760 hours per year divided by 12)no
daysPerMonth30.4Workload: 30.4 days per month when converting requests per dayno
weightsOverheadFactor1.05Memory model: weights take 5% more memory than their raw sizeno
fixedOverheadGb2Memory model: 2 GB reserved for runtime and buffersno
tpEffPcie0.85Memory model: tensor-parallel efficiency over PCIeno
tpEffNvlink0.95Memory model: tensor-parallel efficiency over NVLinkno
tpEffUnified1Memory model: no tensor-parallel loss on a single unified-memory machineno
defaultContextTokens81928192 tokens; typical chat context, editableno
decodeEff0.7Throughput model: decode reaches 70% of memory-bandwidth speedyes
prefillEff0.5Throughput model: prefill reaches 50% of peak computeyes
batchEffK0.04Throughput model: batching efficiency 1 / (1 + 0.04 (n - 1))yes
utilisation0.3Throughput model: hardware busy 30% of the month, editableno
maxUtilisation0.95Throughput model: a box is never loaded above 95%no
sparesPct0.1Monthly cost: 10% added to hardware cost for sparesno
residualPct0.2Monthly cost: 20% of the GPU price recovered at the end of the horizonno
loadFactor0.8Monthly cost: GPUs draw 80% of rated power under loadyes
horizonMonths36Monthly cost: hardware written off over 36 months (24 and 48 offered)no
labourHours2Monthly cost: 2 hours of upkeep per month, editableno
labourRate60Monthly cost: $60 per hour of upkeep, editableno
hostingMonth0Monthly cost: no hosting or colocation fee unless you enter oneno
softwareMonth0Monthly cost: no software or support fee unless you enter oneno
electricityUsdKwh0.1453US commercial average retail price, EIA Electric Power Monthly table 5.6.A, July 2026 (14.53 cents per kWh)no
latencyTargetTps20Workload: an interactive stream should reach 20 tokens per secondno
tokenizerSliceblendedTokenizer normalisation: blended average over the measured corpus slicesyes