Skip to content
Understand
Tool

Self-hosting break-even calculator

At what volume does running your own model beat paying per token? Runs in your browser. Nothing is stored or transmitted.

Compare paying per token with running an open-weight model on dedicated GPUs. Showing illustrative defaults. Replace them with your own quotes and measurements.

Workload

Billion tokens a month, input and output together.

Blended across input and output. Use a serverless open-model price or a proprietary API price. Some small proprietary models now list below £0.20, which pushes break-even much further out.

Self-hosted capacity

From a current quote. List prices vary several-fold between providers.

A 70B model needs about two 80 GB GPUs at 16 or 8 bits, one at 4 bits.

Billion tokens a month. Measure it on your own workload. 10 billion a month is about 3,800 tokens a second.

Headroom for peaks. A new copy is added once this is reached.

Engineering, on-call, monitoring, security and upgrades.

Pay per token

£6,000

a month, £0.60 per million tokens

Self-hosted

£7,650

a month, £0.77 per million tokens

Self-hosted: 1 copy of the model at 50% average utilisation, £3,650 GPUs plus £4,000 people.

At this volume, paying per token is cheaper by £1,650 a month. Self-hosting is cheaper at every volume above 25 billion tokens a month. Below that the answer flips: self-hosting first wins from about 18.8 billion, then loses again just after an extra copy of the model is added.

Monthly cost by volume

Self-hosting cheaper£00bn£12,50012.5bn£25,00025bn£37,50037.5bn£50,00050bnBillion tokens a month (dashed line: your volume)

Self-hosted cost rises in steps, because each new copy of the model adds its full GPU cost before it carries any traffic. The comparison covers running cost only. It assumes the open model clears the workload's quality floor, and it leaves out evaluation, migration and any data or sovereignty requirement that rules one option out.

The reasoning behind each input is set out in How Inference Choices Change AI Cost. The wider sourcing decision, including committed capacity and owned infrastructure, is covered in Rent, Reserve or Own Intelligence? Before comparing prices, check that the open model clears the quality floor, as described in The Economics of Model Selection.