The formula
Per month: the hardware price divided by the depreciation period in months, plus the power in kilowatts times 720 hours, times the share of time under load, times the rate, plus data center hosting, plus maintenance as a yearly share of the hardware price divided by 12, plus the engineer’s share of a salary. Defaults: 36 months, 50% of time under load, a rate of about $0.10 per kWh (the flat household rate in Moscow for the first half of 2026), maintenance at 10% a year, and an engineer cost of 0 for a laptop and a Mac Studio, $300 for a workstation and $600 for a server; maintenance and the engineer share are our estimates.
Examples at prices as of 29 September 2026
A MacBook Pro M5 Max with 128 GB for $8,100: depreciation $200, electricity about $3 at 100 W by our estimate, maintenance $67, total about $300 a month without an engineer. A Mac Studio M4 Max with 128 GB for $11,000: $300 of depreciation, $5 of electricity at 145 W, $89 of maintenance, total about $400; with 10% of an engineer’s salary about $700. A workstation with an RTX 5090 for $13,000 including the platform: $350 of depreciation, $28 of electricity at 825 W, $100 of maintenance, $300 for the engineer, total about $800. A server with two RTX PRO 6000 cards for $30,000: $800 of depreciation, $51 of electricity, $400 of hosting, $250 of maintenance, $600 for the engineer, total about $2,100. A server with four H100 cards for $120,000 including the platform: about $6,100 a month with hosting and an engineer.
What to adjust for your case
Rate: for businesses the price is not regulated and depends on the region and category, so take it from your contract. Depreciation period: 36 months by default; by our estimate 24 for gaming cards under round-the-clock load and 48–60 for a Mac. Engineer share: 0 for a single user, 10–20% of a salary for a department server by our estimate, a full salary for a cluster; the median DevOps engineer salary in 2026 is $2,700 to $3,000 a month. Hosting: 0 in the office, from $90 a month for a server up to 300 W in a data center, and $24 for each extra 100 W.
What the calculation leaves out
Implementation cost: model setup, connection to your systems, tests; with us this is a pilot from $1,499 in 14 days. A spare card or a rental contract in case of failure: from about $0.50 an hour for an RTX 4090 in a Russian cloud. Replacing a card out of warranty: $7,200 to $28,000 at prices from the RTX 4090 to the H100. Staff downtime if the service fails without a plan. Add these lines to the Other field if you want to see them in the total.
How to read the result
Compare the monthly total with the cloud API calculation at the same token volume and with the GPU rental price: an H100 from $1,700 a month, an RTX 4090 from $300. If the local total is lower than the cloud bill, the break-even page shows after how many months the hardware pays for itself. If it is higher, the local option makes sense only because of data requirements, and that is a normal answer: many companies in this section’s cases pay for a closed setup, not for savings.