Local LLM TCO Calculator

Estimate hardware, power, throughput, and cloud cost on the same basis.

Rough estimates from your assumptions. Not a quote, not a budget, and not a purchasing decision. The cloud prices and throughput figures on this page are illustrative defaults captured when it was written, and they have already moved. Everything else is arithmetic on numbers you typed. Estimates only, used at your own risk. Open for the full scope limits.

What this captures

Hardware amortization, electricity, estimated token output, and a direct cost-per-million comparison against the selected cloud default. It produces rough financial and throughput estimates from the values you enter and a small set of simplifying assumptions.

What it does not capture

Model quality, latency variance, cooling, networking, setup time, support burden, compliance requirements, credits and negotiated vendor pricing are all outside the model. So is every one of the following:

  • Real electricity tariffs: time-of-use rates, demand charges and taxes.
  • Hardware depreciation curves, resale value and failure replacement.
  • Cooling overhead, networking and software licensing.
  • Regulatory and compliance costs, downtime, and support.
  • The opportunity cost of operator time, which on a self-hosted deployment is usually the largest line item and never appears on this page.
  • The actual sustained throughput of any cloud provider in your region, under your rate limits, at your concurrency.

The cloud figures are stale by construction

Cloud per-token prices and "typical" cloud throughput values shown here are illustrative defaults captured at the time of authoring. Published prices, model availability, rate limits and streaming speeds change frequently, and the direction is not always down. Treat the cloud defaults as placeholders. Verify current pricing and benchmark real performance before making any purchasing, contracting, capacity-planning or budgeting decision.

A comparison between a number you measured and a number this page remembered is not a comparison.

Never use this for

  • A purchase, a lease or a contract commitment of any size.
  • A budget, a capacity plan, or a figure that goes into a forecast someone is held to.
  • Board, investor or customer material, or any published cost-of-ownership claim.
  • A build-against-buy business case, without the operator time, the compliance work and the failure cases costed properly.
  • Vendor selection or negotiation, where the list price you are comparing to is not the price anyone pays.

Before you spend anything

Independently verify all figures with vendor invoices, measured power draw at the wall rather than a datasheet TDP, and current published pricing. Benchmark the model you would actually run, at the quantisation, context length and concurrency you would actually use, on the hardware you would actually buy.

Inputs

Adjust the hardware, duty cycle, power draw, and comparison assumptions used in the estimate.

Hardware Configuration

Total initial hardware expenditure in USD.
How long the hardware will be used before replacement.
Hours per day the computer is powered on.
Percentage of power-on time spent performing inference.

Power Consumption

Power usage during active token generation.
Baseline power usage while idle.
Local cost of electricity per kilowatt-hour.

Performance Benchmarks

Average speed of token generation.

API & Subscription Comparison