Local LLM vs Cloud API Cost Calculator

Work out how many tokens per month it takes to break even on buying a GPU / Mac to self-host an LLM. Enter hardware and usage to get fixed monthly cost, single-box capacity and the break-even volume, plotted as a cost-vs-volume chart. Unsure of model prices? See the Token Cost Calculator.

🖥️ Local hardware

Apply a hardware preseteditable references; throughput swings with model / quantization / concurrency

Hardware price $ Quantity units Depreciation months Fixed monthly cost hosting / network / ops $ Local throughput tokens / s System power W Electricity $/ kWh

Utilization50%

☁️ Cloud API reference

Reference model

Input : output ratio : Target monthly volume M tokens / mo

Blended cloud price — / M tokens

Fixed monthly cost—F = price÷depreciation + fixed

Monthly capacity—M / mo

Break-even volume—M / mo

—

Local monthly—

API monthly—

Monthly delta—

Utilization at this volume—

Cost-vs-volume chart Where the local line (fixed + marginal power) crosses the cloud-API line (pure usage) is the break-even point

Self-host Cloud API

Guide

How the break-even model works

Self-hosting and cloud API are two cost structures. Local = fixed cost F + marginal m×volume: F is hardware price amortized monthly plus hosting/ops; m is electricity per million tokens (from power, price and throughput). Cloud API = blended price c × volume — pure usage, zero fixed.

Where the two lines cross is the break-even M\* = F ÷ (c − m): above it local wins, below it cloud API wins. The other constraint is single-box capacity M_max = throughput × utilization × seconds per month — one box can only emit so many tokens a month; beyond that you add hardware and fixed cost multiplies.

A worked example

One card: price $2,000, 24-month depreciation, $30/mo fixed → fixed F ≈ $113. With a blended cloud price of $2/M tokens and local marginal $0.1/M tokens — break-even ≈ 113 ÷ (2 − 0.1) ≈ 59 M tokens/mo.
So below ~59M tokens a month, cloud API is actually cheaper.

Throughput (TPS) dominates the result and depends on model size, quantization, batch and concurrency — measure it on your real deployment. To calibrate the cloud price first, see the Token Cost Calculator or LLM price comparison.

FAQ

How is break-even M* calculated?

M* = fixed monthly cost F ÷ (blended cloud price c − local marginal m). It is the monthly volume where local total (F + m×volume) equals cloud API (c×volume). Above M* local is cheaper; below it cloud API is.

Why enter power and electricity price instead of a flat power bill?

The marginal cost of running a model locally is mostly electricity. We derive cost per million tokens from system power × price ÷ throughput — more intuitive and accurate than guessing a per-million-token figure.

Why does utilization matter?

No one runs a GPU at 100% inference 24/7. Utilization sets the single-box monthly capacity M_max = throughput × utilization × seconds per month. Lower utilization makes payback harder.

Is hardware priced in CNY or USD?

It follows the site: CNY on the Chinese site, USD on the global one. The selected reference model's price is converted to the same currency before the calculation.

Are the preset throughput (TPS) values accurate?

They are order-of-magnitude references only. Real throughput depends on model size, quantization, batch size and concurrency, varying several-fold — measure on your own deployment.