What does it cost to run a local LLM per month?
The power is cents. Running Qwen3.8 27B on a Mac Studio M5 Max, 64GB at 500k tokens a day costs $0.26 a month in electricity. The monthly number that matters is the machine: $3,499 once, which is $146 a month if you keep it two years. Renting the same month's work costs $6.94.
Jump to: Where the $0.26 comes from · A month at each level of use · The power is not the cost. The machine is. · If you pay a subscription instead
Where the $0.26 comes from
500k tokens a day at 15:1 is 468,750 you send and 31,250 the model writes back. Only the ones it writes take time: at 24.7 tok/s that is 21 minutes of generating a day. At 145 W that comes to 1.55 kWh over a month, and at $0.17 per kWh (the United States average, US EIA) that is $0.26 a month.
That is the power the machine draws while it is generating. The 145 W is a stand-in: nobody has put a meter on this machine, and its own page says what the figure borrows. Left out, all of which favour the machine less than shown: the time and power that reading your prompt takes, the machine idling while you think, and API prompt caching, which cuts the price of context you send again. Left out in the machine's favour: what it is worth when you sell it, and every other job it does while you own it.
A month at each level of use
Use it harder and the electricity rises with the hours; the rental bill rises with it, from the same tokens, and a great deal faster. The last column is the machine's price divided by what the month leaves over.
| A day's use | Generating | Electricity a month | Rented | Pays the machine back in |
|---|---|---|---|---|
| 50ka few chats a day | 2 min | $0.03 | $0.69 | 436 years |
| 200klight assistant use | 8 min | $0.11 | $2.78 | 109 years |
| 1Ma moderate coding-assistant day | 42 min | $0.53 | $13.89 | 22 years |
| 4Mheavy coding with an agent | 2.8 h | $2.11 | $55.55 | 5.5 years |
| 20Magents running most of the day | 14.0 h | $10.53 | $278 | 13 months |
At 20M tokens a day, agents running most of the day, the power bill is $10.53 a month and the rental bill is $278. That gap is the whole case for buying, and at 500k tokens a day it is $6.68 a month against a $3,499 machine.
The power is not the cost. The machine is.
Across the 30 current machines that run Qwen3.8 27B at 32k context, a month of this work costs between $0.25 and $0.65 in electricity. Divide what they cost to buy over two years and the same month runs from $53.45 to $750. The machine you pick moves the monthly figure by 14 times; the power it draws moves it by cents.
| Machine | Price | Electricity a month | A month over one year | A month over two years | A month over three years |
|---|---|---|---|---|---|
| Framework Desktop, 32GB | $1,269 | $0.58 | $106 | $53.45 | $35.83 |
| Mac mini M6, 32GB | $1,299 | $0.43 | $109 | $54.55 | $36.51 |
| Radeon AI PRO R9700, 32GBcard only | $1,299 | $0.45 | $109 | $54.58 | $36.53 |
| Corsair AI Workstation 300, 64GB | $1,700 | $0.58 | $142 | $71.41 | $47.80 |
| Framework Desktop, 64GB | $1,959 | $0.58 | $164 | $82.20 | $55.00 |
| GeForce RTX 5090, 32GBcard only | $1,999 | $0.39 | $167 | $83.68 | $55.92 |
| GMKtec EVO-X2, 64GB | $2,200 | $0.65 | $184 | $92.32 | $61.76 |
| Mac mini M5 Pro, 48GB | $2,299 | $0.51 | $192 | $96.30 | $64.37 |
| MacBook Pro M5 (14-inch), 32GB | $2,399 | $0.47 | $200 | $100 | $67.11 |
| Mac Studio M5 Max, 36GB | $2,499 | $0.35 | $209 | $104 | $69.77 |
| Mac mini M5 Pro, 64GB | $2,699 | $0.51 | $225 | $113 | $75.48 |
| Mac Studio M5 Max, 48GB | $3,099 | $0.26 | $259 | $129 | $86.35 |
| Framework Desktop, 128GB | $3,449 | $0.58 | $288 | $144 | $96.38 |
| Mac Studio M5 Max, 64GB | $3,499 | $0.26 | $292 | $146 | $97.46 |
| GMKtec EVO-X2, 128GB | $3,500 | $0.65 | $292 | $146 | $97.88 |
| MacBook Pro M5 Pro (16-inch), 48GB | $3,599 | $0.51 | $300 | $150 | $100 |
| GMKtec EVO-X3, 128GB | $3,600 | $0.65 | $301 | $151 | $101 |
| Minisforum MS-S1 Max, 128GB | $3,799 | $0.58 | $317 | $159 | $106 |
| MacBook Pro M5 Pro (16-inch), 64GB | $3,999 | $0.51 | $334 | $167 | $112 |
| Beelink GTR9 Pro, 128GB | $4,349 | $0.58 | $363 | $182 | $121 |
| DGX Spark, 128GB | $4,699 | $0.61 | $392 | $196 | $131 |
| Corsair AI Workstation 300, 128GB | $4,700 | $0.58 | $392 | $196 | $131 |
| MacBook Pro M5 Max (16-inch), 48GB | $4,999 | $0.26 | $417 | $209 | $139 |
| Mac Studio M5 Max, 128GB | $5,099 | $0.26 | $425 | $213 | $142 |
| MacBook Pro M5 Max (16-inch), 64GB | $5,399 | $0.26 | $450 | $225 | $150 |
| Mac Studio M5 Ultra, 96GB | $5,499 | $0.25 | $459 | $229 | $153 |
| HP Z2 Mini G1a, 128GB | $5,544 | $0.58 | $463 | $232 | $155 |
| MacBook Pro M5 Max (16-inch), 128GB | $6,999 | $0.26 | $584 | $292 | $195 |
| Mac Studio M5 Ultra, 256GB | $10,799 | $0.25 | $900 | $450 | $300 |
| RTX PRO 6000 Blackwell, 96GBcard only | $18,000 | $0.35 | $1,500 | $750 | $500 |
List prices, current machines, each running Qwen3.8 27B at 32k context and 500k tokens a day. Graphics cards are priced as the card alone, so add the PC around one before comparing it with a complete computer. All seven cards here are ranked by what each one holds. Watts do not order the electricity: the RTX PRO 6000 Blackwell, 96GB draws 600 W and spends $0.35 a month, where the Mac mini M6, 32GB draws 65 W and spends $0.43. A machine that generates faster is finished sooner.
If you pay a subscription instead
Every rental figure here is priced by the token, from the published rate for Qwen3.8 27B, because that is a price this site can check. A subscription is a flat monthly bill, and the number to set against it is in the table above: $146 a month for 500k tokens a day on a Mac Studio M5 Max, 64GB, or $97.46 if you keep it three years. If your bill is smaller than that, the machine does not pay for itself by replacing it.
The calculator takes the bill directly: enter what you pay each month and it prices the machine against that instead of against per-token rates. Two things go with it, and both are in the small print there. A subscription buys the lab's own model, so the comparison only holds if Qwen3.8 27B can do the work you are paying for. And a bill you pay whatever you use is not a bill that rises with use, so the falling-price assumption is switched off in that mode.
Every figure is Qwen3.8 27B at Q4_K_M unless the row names another machine, at 32k context, 15:1 input to output, $0.17 per kWh, and today's API prices held flat. A month is 30.44 days, the year divided by twelve. The speed above is worked out from memory bandwidth rather than measured, and each machine page says which of the two it has. For the same arithmetic counted in tokens rather than months, read what a million tokens costs each way; for the machine that pays back soonest at each level of use, the best buys.