Locally Explains

Don’t Buy a Loud GPU Tower for Local AI All Day

Every figure shown in the video, with where it comes from. Pages were read on 5 October 2026.

All yearly costs use the published idle (or measured) draw, 8,760 hours a year and $0.16 per kilowatt hour.

Draw	kWh a year	Cost a year
4 W (Mac mini M6, idle)	35.0	$5.61
6 W (Mac mini M5 Pro, idle)	52.6	$8.41
7 W (Mac Studio M5 Max, idle)	61.3	$9.81
9 W (Mac Studio M5 Ultra, idle)	78.8	$12.61
87 W (Acemagic M1A PRO+, Silent mode at the wall, under load)	762	$121.94
163 W (Acemagic M1A PRO+, Performance mode at the wall, under load)	1,428	$228.47
385 W (Mac Studio M5 Ultra, Apple’s maximum)	3,373	$539.66

Dollars per token per second: $899 / 9.5 tokens per second = $94.6.

Weights at 4 bit, at roughly half a byte per parameter: 27B about 13.5 GB, 70B about 35 GB, 120B about 60 GB, before cache and runtime overhead.

Apple, Mac mini power consumption and thermal output. Mac mini (M6), M6 with 16 GB and 256 GB SSD: 4 W idle, 70 W maximum. Mac mini (M5 Pro), 64 GB and 8 TB: 6 W idle, 145 W maximum. https://support.apple.com/en-nz/103253
Apple, Mac mini (M6 or M5 Pro) tech specs. M6: 16 GB unified memory, configurable to 24 GB or 32 GB (170 GB/s). M5 Pro: 307 GB/s memory bandwidth, configurable to 48 GB or 64 GB. The ECMA-109 acoustic table gives a sound pressure level at the operator position of 4 dB at idle (configuration tested: M5 Pro, 48 GB, 8 TB); the electrical summary line on the same page reads 5 dBA at idle. https://support.apple.com/en-us/128108
Apple Newsroom, 22 September 2026. “Mac mini with M6 starts at $899, while Mac mini with M5 Pro is available at $1,699. Mac Studio with M5 Max starts at $2,499, while Mac Studio with M5 Ultra starts at $5,499.” https://www.apple.com/newsroom/2026/09/the-new-mac-mini-and-mac-studio-are-available-today/
ComputerBase, Apple Mac mini M6 im Test, 21 to 23 September 2026. Idle about 4 W at the wall; up to 60 W under maximum load. Under sustained full load (HandBrake) 35.7 dB(A) at 40 cm. Qwen3.8 27B (Q4_K_M, LM Studio 0.4.24): 9.5 tokens per second; Qwen3 14B (Q4_K_M): 17.8 tokens per second, on the 32 GB model. https://www.computerbase.de/artikel/pc-systeme/apple-mac-mini-m6-test.99469/

Apple, Mac Studio power consumption. M5 Max (36 GB): 7 W idle, 200 W maximum. M5 Ultra (512 GB): 9 W idle, 385 W maximum. https://support.apple.com/en-au/102027
Apple, Mac Studio (M5 Max or M5 Ultra) tech specs. M5 Max: 36 GB, configurable to 48, 64 or 128 GB. M5 Ultra: 96 GB, configurable to 256 GB or 512 GB. ECMA-109 table: 7 dB at the operator position at idle (configuration tested: M5 Ultra, 96 GB). https://support.apple.com/en-euro/128107
Apple Store. 512 GB memory option for M5 Ultra “coming late October”. https://www.apple.com/shop/buy-mac/mac-studio
COMPUTER BILD, Apple Mac Studio 2026 M5 Max: Test des Desktop-PC, 24 September 2026. 21 W in normal operation, 182 W at full load; 0.1 sone in normal operation, 1.0 sone at full load. https://www.computerbild.de/artikel/Tests-PC-Hardware-Apple-Mac-Studio-2026-M5-Max-Test-des-Desktop-PC-3bc9c91-41239213.html
Tom’s Hardware, Apple Mac Studio (M5 Ultra) review, 21 September 2026. Tested with Qwen 3.8-27B-Q4_K_M in llama.cpp; tokens per second “double that of the M4 Max and almost four times higher than the DGX Spark across the board”; “The system stayed largely quiet (though the fan can make itself known)”. The cheapest M5 Ultra is $5,499 with 96 GB; the 36-core version adds $1,300; moving from 96 GB to 256 GB adds $4,000. 512 GB configuration “coming in late October”. https://www.tomshardware.com/desktops/mini-pcs/apple-mac-studio-m5-ultra-review

Framework Desktop. Recommended models measured on Framework Desktop: Ryzen AI Max+ 395 with 64 GB runs Qwen3.5-122B-A10B at Q3_K_S using 52.5 GB; Ryzen AI Max+ 395 with 128 GB runs DeepSeek V4 Flash 0731 at UD-IQ2_XXS using 90.9 GB. https://frame.work/desktop
Framework Desktop DIY configurator. Max+ 395 with 128 GB listed at 3,449 (the configurator served in GBP from this location) and marked out of stock. https://frame.work/products/desktop-diy-amd-aimax300/configuration/new
ServeTheHome, Acemagic M1A PRO+ review. In the 140 W Performance mode the system drew 163 W at the wall with a GPT-OSS model on the GPU, a CPU stress test and a roughly 10 W portable monitor; Silent mode brought that to 87 W. About 43 dBA in Performance mode and 36 to 38 dBA in Silent mode, against a 34 dBA lab noise floor. https://www.servethehome.com/acemagic-m1a-pro-review-an-amd-powered-128gb-ai-mini-pc/4/
PC Watch, Minisforum MS-S1 MAX. Noise at 50 cm: Quiet mode 32 dB idle and 41.6 dB at full load; Performance mode 37 dB idle and 48 dB at full load. https://pc.watch.impress.co.jp/docs/column/hothot/2050899.html

NVIDIA, GeForce RTX 5090. 32 GB of GDDR7. https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5090/
GamersNexus, RTX 5090 Founders Edition review. A two slot cooler “capable of handling 575W or more”; about 32.5 dBA at 1 m on the passive test bench. https://gamersnexus.net/gpus/nvidia-geforce-rtx-5090-founders-edition-review-benchmarks-gaming-thermals-power

The opening’s “ninety billion parameter model” is, in Framework’s own table, a 90.9 GB DeepSeek V4 Flash at about 2 bit; the table gives memory used, not a parameter count.
The 21 W, 182 W and sone figures for the M5 Max Mac Studio are from COMPUTER BILD, not ComputerBase.
Tom’s Hardware prices 256 GB on the M5 Ultra at $4,000 over the $5,499 base; $6,799 is the price of the 36-core M5 Ultra at 96 GB.
“Ninety one gigabytes for a roughly hundred twenty B mixture of experts”: Framework’s 90.9 GB row is DeepSeek V4 Flash; its roughly 120B mixture of experts row (Qwen3.5-122B-A10B) uses 52.5 GB.
Apple’s acoustic figures (4 dB and 7 dB) come from the configurations Apple tested (M5 Pro 48 GB; M5 Ultra 96 GB), not every model.
The Framework 128 GB price was read in GBP; the US dollar price was not re-checked.