Memory
Mac: unified memory of 96, 128, 256 or 512 GB in one case, and a 671-billion-parameter model in 4 bits fits in 512 GB. Graphics cards: 24 GB on the RTX 4090, 32 GB on the RTX 5090, 96 GB on the RTX PRO 6000, 80 GB on the A100 and H100; more only by combining cards. For memory in one device, the Mac wins outright.
Speed
Generation in 4 bits: on a 7-billion-parameter model the Mac M4 Max gives 83.1 tokens per second, the M5 Max 119.9 and the M3 Ultra 92.1; on an 8-billion model the RTX 4090 gives 127.7, the A100 138.3 and the H100 144.5 (different measurements, models of similar size). For one user the difference is small. Processing a long prompt is several times faster on graphics cards; on a Mac Studio, the 671-billion model reads 8,000 tokens in 888 seconds. Dozens of simultaneous users: here, in our estimate, a graphics card with batched request serving is stronger; measure the total speed on your own load before you buy.
Price as of 29 September 2026
Mac Studio M4 Max 128 GB: $11,000. M3 Ultra 96 GB: $7,800. M3 Ultra 512 GB with 4 TB: $38,000. RTX 5090 32 GB: $7,700–$7,900 plus a platform of $3,600–$4,200 in our estimate. RTX 4090 24 GB: $5,900 in July 2026; on 30 September 2026 the seller in the source no longer had it. RTX PRO 6000 96 GB: $13,000. A100 80 GB: $13,000. H100 80 GB: $28,000. Per gigabyte of memory the Mac is cheaper: about $75–$83 per GB against $150 for the RTX PRO 6000 and $250 for the RTX 5090 without a platform. With many users, in our estimate, the graphics card wins thanks to batched request serving.
Electricity, noise, maintenance
Mac Studio M4 Max draws 145 W at most and the M3 Ultra 270 W; it sits on a desk, cannot be heard, and an employee installs the updates. A workstation with an RTX 5090 draws about 825 W by our calculation (the card’s 575 W plus the platform) and can be heard in the office. Server cards with blower fans live in a data center, from $90 a month per server of up to 300 W, and in our estimate need an engineer for 10–20% of a full-time role. If a card fails out of warranty, it costs $7,200–$28,000; a Mac is repaired as a whole.
The answer by task
In our estimate: one to three users, models up to 70 billion parameters, silence and simplicity: a Mac Studio with 128 GB. Very large models for one user, 235–671 billion: a Mac Studio with 512 GB at $38,000; a 235-billion model in 4 bits also fits on two 96 GB cards (about $30,000 with a platform), while 671 billion needs more than 400 GB of video memory. A department of 10–20 people and agents with a queue: a server with one or two 96 GB cards and vLLM. Hundreds of users: server cards in a data center, or rental. Development and tests before buying: a laptop with 36–64 GB.