Cost per 1M generated tokens

What it costs to generate a million tokens of qwen3-8b two ways: the electricity your machine burns doing it, and the hourly rate to rent the same class of card. Local figures use $0.15/kWh — set your own rate on the interactive page.

The honest framing, because the first number is startling: local electricity cost per token is remarkably low. It is also not the whole cost — it excludes the machine itself, 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?

What this ranking rests on: 1 measured of 55 rows. 10 rows are modelled on an architecture whose efficiency constant was fitted from real runs; 44 rest on a constant nobody has calibrated. Every row below says which it is.

Hardware	Speed	Power	Local $/1M at $0.15/kWh	Cloud $/1M	Evidence
1. Apple M5 Max	51 tok/s	35 W SoC package	$0.028	—	ESTIMATED · uncalibrated
2. Apple M4 Max	46 tok/s	35 W SoC package	$0.031	—	ESTIMATED · uncalibrated
3. Apple M1 Max	35 tok/s	30 W SoC package	$0.035	—	ESTIMATED · uncalibrated
4. Apple M3 Ultra	65 tok/s	60 W SoC package	$0.038	—	ESTIMATED · uncalibrated
5. Apple M1 Ultra	64 tok/s	60 W SoC package	$0.039	—	ESTIMATED · uncalibrated
6. Apple M2 Ultra	64 tok/s	60 W SoC package	$0.039	—	ESTIMATED · uncalibrated
7. Apple M3 Max	35 tok/s	35 W SoC package	$0.041	—	ESTIMATED · uncalibrated
8. Apple M5 Pro	28 tok/s	30 W SoC package	$0.045	—	ESTIMATED · uncalibrated
9. Apple M2 Max	35 tok/s	40 W SoC package	$0.047	—	ESTIMATED · uncalibrated
10. Apple M4 Pro	25 tok/s	30 W SoC package	$0.050	—	ESTIMATED · uncalibrated
11. Apple M5	14 tok/s	23 W SoC package	$0.066	—	ESTIMATED · uncalibrated
12. Apple M1 Pro	19 tok/s	30 W SoC package	$0.067	—	ESTIMATED · uncalibrated
13. Apple M2 Pro	19 tok/s	30 W SoC package	$0.067	—	ESTIMATED · uncalibrated
14. Apple M4	11 tok/s	22 W SoC package	$0.080	—	ESTIMATED · uncalibrated
15. NVIDIA H100 80GB	181 tok/s	350 W board	$0.080	$4.43 checked 2026-07	ESTIMATED · uncalibrated
16. NVIDIA A100 80GB	146 tok/s	300 W board	$0.086	$2.65 checked 2026-07	ESTIMATED · provisional fit (n=1)
17. Apple M3 Pro	14 tok/s	30 W SoC package	$0.087	—	ESTIMATED · uncalibrated
18. Apple M2	10 tok/s	20 W SoC package	$0.087	—	ESTIMATED · uncalibrated
19. Apple M3	10 tok/s	22 W SoC package	$0.096	—	ESTIMATED · uncalibrated
20. NVIDIA GeForce RTX 5060	59 tok/s	145 W board	$0.102	—	ESTIMATED · uncalibrated

On the SoC marker: unified-memory systems (Apple Silicon, DGX Spark, Strix Halo) report whole-package power — CPU, GPU and memory controller on one die — where a discrete card reports board power for the card alone. Neither figure is the wall draw of the complete machine. They are not the same measurement, so read the ranking as a real efficiency difference that is partly a difference in what is being counted.

Reference point. A mainstream hosted API charges materially more per million output tokens than any figure above — see the dated per-provider list on the cost calculator, last checked 2026-07-30. That comparison is the one every reader is making anyway, so it is better cited than guessed. It is also not like-for-like: a hosted flagship is a larger and more capable model than anything in this table.

How this is calculated. local $/1M = (watts / 1000) x kwh_rate x 1e6 / (tok_s x 3600) and cloud $/1M = hourly_rate x 1e6 / (tok_s x 3600). Worked checks: 450 W at $0.30/kWh and 100 tok/s is $0.375 per 1M; $0.34/hr at 100 tok/s is $0.94 per 1M. Cloud figures assume the rented card generates for the whole hour — idle time between requests is billed and is not in the number.

Also ranked: tokens per watt, performance vs cost vs power vs VRAM, or the full filterable database.