CentsPerToken
Open calculator

Calculator

Local vs cloud LLM cost comparer

Pick a cloud API model and a local or rented GPU profile. CentsPerToken charts both cost curves, marks the break-even request volume, folds engineering time into the fixed line, and checks whether your fleet can actually serve the traffic.

Open calculatorRead the guide

Verdict at your volume

Cloud API wins

$3,017.73saved / month

96% cheaper at 150,000 requests per month

Cloud API$137.85
GPT-5 mini
Local GPU$3,155.58
A100 80GB
Cloud / request$0.000919
Local / request$0.021
Cloud $ / 1M tokens$0.5929
Local $ / 1M tokens$13.57
Break-even volumeNever
GPU hours / month1242.0

Crossover analysis

Monthly cost as request volume grows

Cloud APILocal GPU
$0$1.35K$2.7K$4.04K$5.39KYou0300,000 requests / month

Local never overtakes cloud with these settings—variable GPU cost per request is already higher than the API rate.

Step 1

Workload

Step 2

Hardware profile

Capacity headroom170% used

This volume needs 1242 GPU-hours but 1 GPU only supply 730. Add hardware or raise throughput.

Step 3

True cost of ownership

Self-hosting is rarely just GPU hours. Engineering time folds into the fixed monthly line so break-even stays honest.

Infra fixed$200.00
People (8h × $90.00)$720.00
Variable GPU$2,235.58
Local total / month$3,155.58

29% of local spend is fixed overhead — the part that does not shrink when traffic drops.

Sensitivity

Who wins at each volume

10,000
$9.19
$1.07K
Cloud
50,000
$45.95
$1.67K
Cloud
100,000
$91.9
$2.41K
Cloud
250,000
$229.75
$4.65K
Cloud
500,000
$459.5
$8.37K
Cloud
1,000,000
$919
$15.82K
Cloud

Local throughput and GPU rates are rough planning inputs—not MLPerf results. Capacity headroom assumes 730 hours per GPU per month. Self-hosting also adds reliability, security, and on-call risk not priced here. Updated 2026-07-31. Approximate static list prices for planning only. Always verify against each provider’s official pricing page before production budgeting.

Guide

How to get value from this calculator

Compare hosted LLM API spend against self-hosted or rented GPU economics for your workload. Cloud APIs win at low and variable volume; owned or dedicated GPUs can win at high stable throughput—but only after accounting for hardware, ops, and idle capacity. Use this for infrastructure decisions, not precise TCO audits.

Cost formula

Cloud cost = monthly tokens priced at API rates; Local cost = (GPU hourly rate × hours run) + power + ops overhead; Break-even = cloud monthly cost ÷ local monthly cost at given utilization

Why it matters

Local inference looks cheap on GPU-hour spreadsheets but loses when utilization is low or models update frequently. A structured comparison prevents premature infrastructure investment or overpaying cloud at scale.

How to use it

  1. Pick a workload preset, then tune input tokens, output tokens, and monthly requests to match your traffic.
  2. Choose a hardware profile, or override GPU hourly rate and effective tokens per second with your own benchmarks.
  3. Drag utilization down to a realistic figure—GPUs lose their advantage the moment they sit idle.
  4. Add engineering hours and a blended hourly rate so people cost lands in the fixed monthly line.
  5. Set your fleet size and check capacity headroom; if it reads over 100%, the hardware cannot serve the volume.
  6. Read the crossover chart to find the request volume where local overtakes the API, then confirm against the sensitivity table.

Planning tips

  • Variable and spiky traffic favors cloud APIs—you pay only for tokens used.
  • Stable high-volume batch inference is the strongest case for dedicated GPUs.
  • Model updates and quantization require ops time—cloud APIs offload that burden.
  • Include failover and multi-GPU redundancy in local cost, not single-GPU best case.
  • Hybrid architectures (local for bulk, cloud for peak) often beat pure either-or.
  • Compare effective dollars per million tokens on both sides—it normalizes away request shape and makes the gap obvious.
  • Watch the fixed-share bar: when most of local spend is overhead, a traffic dip flips the decision back to cloud.
  • Copy the decision receipt or export the sensitivity CSV so the assumptions travel with the recommendation.

Frequently asked questions

Does local inference include model licensing?

Open-weight models may have license constraints. This calculator focuses on hardware and API cost comparison—verify license terms separately.

Are GPU hourly rates exact?

CentsPerToken uses approximate rental rates for planning. Your cloud GPU provider, spot pricing, and utilization will differ.

What utilization should I assume?

Be conservative. Production clusters rarely sustain 100% utilization. Model at 30–60% unless you have proven steady load.

Are cloud API prices official here?

No. API side uses approximate list prices. Confirm both API and GPU rates before major infrastructure decisions.

Back to tools