Configurations and prices as of 29 September 2026
M4 Max, 16 CPU cores and 40 GPU cores, 128 GB, 1 TB: $11,000. M3 Ultra, 28 CPU cores and 60 GPU cores, 96 GB, 1 TB: $7,800. M3 Ultra, 32 CPU cores and 80 GPU cores, 512 GB, 4 TB: $38,000. These are prices of Russian sellers without their VAT terms; check them on the day you buy.
Power draw
According to the maker, the Mac Studio M4 Max draws 6 W at idle and 145 W at most, and the M3 Ultra 9 W and 270 W. A single 32 GB graphics card draws up to 575 W, a server card up to 700 W. Running under load around the clock, the M3 Ultra uses about 194 kWh a month, which is about $18 at the Moscow household rate of about $0.10; for business the price is not regulated and depends on the contract.
Speed
The M3 Ultra with 80 GPU cores runs a 7-billion-parameter model in 4 bits at 92.1 tokens per second, and the M4 Max with 40 cores at 83.1. A 671-billion-parameter model in 4 bits on the M3 Ultra 512 GB gives 6.21 tokens per second, with prompt processing at 9 tokens per second, so a prompt of 8,000 tokens takes 14.8 minutes. The conclusion from the measurement: 512 GB let you run a 671-billion-parameter model, but not at a speed for a team.
Who it suits
In our estimate, it suits a team of up to 5–10 people with 30–70-billion-parameter models, a department that processes documents in the background, and a company with no server room and no hardware engineer. It does not suit a service for hundreds of simultaneous users or fine-tuning of large models: those need server cards.