Cost per 1M generated tokens: local electricity vs cloud rental
What it costs to generate a million tokens of Qwen 3 8B, two ways: the electricity your own machine burns doing it, and the hourly rate to rent the same class of card.
The honest framing, because the first number is startling: local electricity cost per token is remarkably low — small fractions of a cent per million tokens on efficient hardware. It is also not the whole cost. It excludes the machine itself, its amortisation, idle draw, power-supply inefficiency and every other component in the box. The cost calculator answers the question this page does not: when does buying the hardware pay for itself?
| Hardware | Speed | Power | Local $/1M at $0.15/kWh | Cloud $/1M | Evidence |
|---|---|---|---|---|---|
| 51 tok/s | 35 W SoC package | $0.028 | — | ||
| 46 tok/s | 35 W SoC package | $0.031 | — | ||
| 35 tok/s | 30 W SoC package | $0.035 | — | ||
| 40 tok/s | 35 W SoC package | $0.037 | — | ||
| 65 tok/s | 60 W SoC package | $0.038 | — | ||
| 64 tok/s | 60 W SoC package | $0.039 | — | ||
| 64 tok/s | 60 W SoC package | $0.039 | — | ||
| 36 tok/s | 35 W SoC package | $0.040 | — | ||
| 35 tok/s | 35 W SoC package | $0.041 | — | ||
| 28 tok/s | 30 W SoC package | $0.045 | — | ||
| 35 tok/s | 40 W SoC package | $0.047 | — | ||
| 25 tok/s | 30 W SoC package | $0.050 | — | ||
| 27 tok/s | 35 W SoC package | $0.054 | — | ||
| 89 tok/s | 120 W SoC package | $0.056 | — | ||
| 84 tok/s | 115 W board | $0.057 | — | ||
| 105 tok/s | 150 W board | $0.060 | — | ||
| 105 tok/s | 150 W board | $0.060 | — | ||
18Apple M6 | 16 tok/s | 25 W SoC package | $0.065 | — | |
19Apple M5 | 14 tok/s | 23 W SoC package | $0.066 | — | |
| 45 tok/s | 72 W board | $0.067 | — | ||
| 19 tok/s | 30 W SoC package | $0.067 | — | ||
| 19 tok/s | 30 W SoC package | $0.067 | — | ||
| 40 tok/s | 70 W board | $0.073 | — | ||
| 79 tok/s | 150 W board | $0.079 | — | ||
| 105 tok/s | 200 W board | $0.079 | — | ||
26Apple M4 | 11 tok/s | 22 W SoC package | $0.080 | — | |
| 52 tok/s | 100 W board | $0.080 | — | ||
| 181 tok/s | 350 W board | $0.080 | $4.43checked 2026-07 · stale | ||
| 146 tok/s | 300 W board | $0.086 | $2.65checked 2026-07 · stale | ||
| 14 tok/s | 30 W SoC package | $0.087 | — | ||
31Apple M2 | 10 tok/s | 20 W SoC package | $0.087 | — | |
| 141 tok/s | 300 W board | $0.089 | — | ||
| 141 tok/s | 300 W board | $0.089 | — | ||
34Apple M3 | 10 tok/s | 22 W SoC package | $0.096 | — | |
| 59 tok/s | 145 W board | $0.102 | — | ||
| 118 tok/s | 300 W board | $0.106 | — | ||
| 58 tok/s | 150 W board | $0.107 | — | ||
| 109 tok/s | 300 W board | $0.115 | — | ||
| 41 tok/s | 115 W board | $0.116 | — | ||
| 71 tok/s | 200 W board | $0.118 | — | ||
| 105 tok/s | 300 W board | $0.119 | — | ||
| 238 tok/s | 700 W board | $0.122 | — | ||
| 44 tok/s | 130 W board | $0.123 | — | ||
| 39 tok/s | 115 W board | $0.123 | — | ||
| 84 tok/s | 250 W board | $0.125 | — | ||
| 59 tok/s | 180 W board | $0.126 | — | ||
| 59 tok/s | 180 W board | $0.126 | — | ||
48Apple M1 | 7 tok/s | 20 W SoC package | $0.127 | — | |
| 71 tok/s | 220 W board | $0.130 | — | ||
| 79 tok/s | 250 W board | $0.132 | — | ||
| 22 tok/s | 70 W board | $0.132 | — | ||
| 90 tok/s | 285 W board | $0.133 | — | ||
| 141 tok/s | 450 W board | $0.133 | $0.67checked 2026-07 · stale | ||
| 72 tok/s | 230 W board | $0.133 | — | ||
| 109 tok/s | 350 W board | $0.134 | $2.19checked 2026-08 · stale | ||
| 110 tok/s | 360 W board | $0.136 | — | ||
| 96 tok/s | 320 W board | $0.139 | — | ||
| 170 tok/s | 575 W board | $0.141 | — | ||
| 94 tok/s | 320 W board | $0.141 | — | ||
| 63 tok/s | 220 W board | $0.146 | — | ||
| 170 tok/s | 600 W board | $0.147 | — | ||
| 38 tok/s | 140 W SoC package | $0.153 | — | ||
| 43 tok/s | 160 W board | $0.154 | — | ||
| 43 tok/s | 165 W board | $0.158 | — | ||
| 52 tok/s | 200 W board | $0.160 | — | ||
| 38 tok/s | 150 W SoC package | $0.164 | — | ||
| 75 tok/s | 295 W board | $0.165 | — | ||
| 71 tok/s | 285 W board | $0.168 | — | ||
| 13 tok/s | 54 W SoC package | $0.168 | — | ||
| 85 tok/s | 350 W board | $0.172 | $1.51checked 2026-08 · stale | ||
| 72 tok/s | 300 W board | $0.173 | — | ||
| 83 tok/s | 350 W board | $0.175 | — | ||
| 83 tok/s | 350 W board | $0.175 | — | ||
| 35 tok/s | 150 W board | $0.177 | — | ||
| 35 tok/s | 150 W board | $0.177 | — | ||
| 52 tok/s | 225 W board | $0.180 | — | ||
| 81 tok/s | 355 W board | $0.183 | — | ||
| 27 tok/s | 120 W SoC package | $0.184 | — | ||
| 45 tok/s | 200 W board | $0.185 | — | ||
| 71 tok/s | 315 W board | $0.186 | — | ||
| 71 tok/s | 320 W board | $0.187 | — | ||
| 58 tok/s | 263 W board | $0.188 | — | ||
| 35 tok/s | 160 W board | $0.189 | — | ||
| 37 tok/s | 170 W board | $0.192 | — | ||
| 41 tok/s | 190 W board | $0.192 | — | ||
| 55 tok/s | 260 W board | $0.199 | — | ||
| 63 tok/s | 300 W board | $0.199 | — | ||
| 46 tok/s | 220 W board | $0.200 | — | ||
| 63 tok/s | 304 W board | $0.201 | — | ||
| 41 tok/s | 200 W board | $0.203 | — | ||
| 45 tok/s | 220 W board | $0.203 | — | ||
| 59 tok/s | 290 W board | $0.205 | — | ||
| 90 tok/s | 450 W board | $0.208 | — | ||
| 30 tok/s | 165 W board | $0.228 | — | ||
| 43 tok/s | 245 W board | $0.238 | — | ||
| 30 tok/s | 190 W board | $0.263 | — | ||
| 51 tok/s | no power data | — | — | ||
| 62 tok/s | no power data | — | — | ||
| 52 tok/s | no power data | — | — | ||
| 84 tok/s | no power data | — | — | ||
| 105 tok/s | no power data | — | — | ||
| 170 tok/s | no power data | — | — |
SoC marks the 27 unified-memory systems in this ranking. Their power figure is whole-package power — CPU, GPU and memory controller on one die — where a discrete card reports board power for the card alone. Neither is the wall draw of the complete machine. They are not the same measurement, so read the ranking as a real difference in efficiency that is partly a difference in what is being counted.
How this is calculated
Decode speed comes from the same engine as the rest of the site — a three-term latency model with efficiency fitted per GPU architecture. Full methodology
local $/1M = (watts / 1000) x kwh_rate x 1e6 / (tok_s x 3600) cloud $/1M = hourly_rate x 1e6 / (tok_s x 3600) worked check, local: 450 W, $0.30/kWh, 100 tok/s -> $0.375 / 1M worked check, cloud: $0.34/hr, 100 tok/s -> $0.94 / 1M at your rate of $0.15/kWh, a 450 W card at 100 tok/s costs $0.188 per 1M generated tokens in electricity.
Watts is the heat, and the heat is the noise — but the mapping between them is the cooler's job, and we hold no cooler data. A 450 W card in a well-designed case can be quieter than a 250 W card with a bad fan curve, so nothing here should be read as a loudness ranking. If you own a sound meter, we would like your reading.
lub porównaj w Vast.ai od $0.35/godz. (typical low · varies)
Jako partner Amazon zarabiamy na kwalifikujących się zakupach. Linki do chmurowych GPU są linkami polecającymi — możemy otrzymać prowizję bez dodatkowych kosztów dla Ciebie.