The formula
Months to break even: the hardware price divided by the gap between the cloud bill and the monthly cost of the local option without depreciation. If the local option costs more than the cloud bill, there is no payback: the cloud is cheaper at any time frame. The calculator takes the price and costs from the local AI calculator and the bill from the API calculator, and you can enter any field by hand.
Examples at prices as of 29 September 2026
A Mac Studio M4 Max for $11,000 with running costs of $94 a month, against a scenario of 60 million tokens: on Sonnet 5.5 the bill is $250 and payback takes 72 months, longer than the hardware lives; on a Russian bank’s Pro model $350, 40 months; on a large Russian tech company’s Pro model $900, 14 months; on Fable 5.1, $1,200, 10 months. A workstation with an RTX 5090 for $13,000 with costs of $450 including the engineer, against Fable 5.1: 17 months. A server with two RTX PRO 6000 cards for $30,000 with costs of $1,300 pays back in 36 months only with a cloud bill above $2,100 a month, which means from 105 million tokens a month on Fable 5.1 at the same input-to-output ratio.
What this means
In money, the local option wins in three cases: a volume of hundreds of millions of tokens a month, expensive cloud models or Russian cloud models at $6 to $14 per million tokens, and a task that an open model of up to 70 billion parameters can solve. If you use cheap models at a small volume, your own hardware makes sense only for data, latency or independence from the provider, and the calculator shows this with a negative gap.
What the calculator does not know
Quality: in a retail chain’s test, a local model with 9 billion parameters did as well as cloud models on chains of 2–3 steps, but on a 15-step chain none of the three models gave a fully verifiable result; test your own task before you buy. Volume growth: the cloud bill grows in line with tokens, the local one does not while the card has capacity. Exchange rate: a bill in dollars changes with the Bank of Russia rate, so enter your own forecast. Hardware life: in the calculation the cards serve 3 years, the same as the depreciation period, after which a next-generation model may not fit.
A third option
GPU rental in a Russian cloud: an H100 from $1,700 a month, an RTX 4090 from $300. The data stays in Russia, the hardware belongs to someone else, and there is no payback to calculate. The calculator shows rental as a third line next to buying and the cloud API.