Run or Rent

One job, one card, one API

Qwen: Qwen3 Coder 30B A3B Instruct on GeForce RTX 5090

The sums below use the “Chatbot for a small team” example: 4,200 jobs a month, 1,500 tokens in, 400 out. 8 people, 25 questions on each of 21 working days. 1,500 tokens in (the question plus a short history) and 400 tokens out. A worked example, not a log from a real team.

Not enough yet

On running cost, the card is cheaper: £1.10 of electricity against £1.46 for the API, a saving of £0.35 a month. Type what you would pay for the GPU to see how many months it takes to pay for itself.

Upper estimate. 575 W × (1900 tokens ÷ 304.9 tokens/sec ÷ 3,600) ÷ 1,000 × 26.32p/kWh

That expression is pence for one job. Divide by 100 for pounds, then multiply by the jobs in the month. Inputs: 575 W maximum rated power, 1,900 tokens, 304.9 tokens/sec, 26.32p/kWh.

API this month
£1.46
Electricity this month
£1.10 · 4.18 kWh
GPU purchase
Not on file
Break-even
Add a price
  • API cost is input tokens and output tokens, priced per million, times jobs per month. No caching, batch discount, or free tier.
  • Pound figures use the stored exchange rate. They are not a quote from the company that trained the model.
  • Upper estimate. Every token, in and out, is divided by the reported tokens per second. Prompt tokens are usually faster than that, and a card rarely holds its maximum rated power for the whole job, so the real bill is often lower.
  • Formula, in pence per job: 575 W × (1900 tokens ÷ 304.9 tokens/sec ÷ 3,600) ÷ 1,000 × 26.32p/kWh. Divide by 100 for pounds, then multiply by the jobs in the month.
  • The tokens per second are one published run, not a promise for your PC. Context length and software change it.
  • The watts are the card's maximum rated graphics power. Real draw during inference is usually lower.
  • The standing charge on a home bill is not added. A GPU does not create a second one.
  • API price: £0.112 in and £0.447 out per million tokens, from $0.150 and $0.600 at 0.74501 dollars to the pound on 2026-11-10. OpenRouter feed 2026-11-11.
  • Speed: 304.9 reply tokens/sec, 514.5 prompt tokens/sec, Q4_K_M, llama.cpp 0.7.0-dev (commit 1268eda). Run r_vsype4pm1ko, marked Measured by llm-speed. Chat-short. No power draw in the row. Page read 11 November 2026.
  • GeForce RTX 5090: 32 GB GDDR7. Total graphics power: 575 W. Specs read 2026-11-11.
  • Speed is the cited run, not a guess.
  • Upper estimate, based on the official total graphics power of 575 W.

OpenRouter prices · Exchange rate · GPU spec · Speed source · Ofgem unit rate

Change the job size Model page · GPU page