256GB and 512GB AI Systems
Apple Mac Studio (M5 Ultra) leads this page with an AIPC Score of 57.7, giving a local model about 384GB of usable AI memory, though it has no verified price yet. 2 systems on this page, 0 with a verified price.
| # | System | Type | AI Memory | GPU/NPU | AI Score | Price | Verified | View |
|---|---|---|---|---|---|---|---|---|
| 1 |  Apple Mac Studio (M5 Ultra) | Desktop | 384GB | Apple M5 Ultra 80-core GPU | 57.7 | No verified price | — | View details → |
| 2 |  NVIDIA DGX Station (GB300) | Workstation | 252GB | NVIDIA Blackwell Ultra | 51.4 | No verified price | — | View details → |
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply to the purchase of this product.
The models nothing else will hold
Above 200GB of usable AI memory you are into territory that effectively one product line occupies. A 512GB Mac Studio gives a model around 384GB, which is enough for the very largest open-weight models at Q4, including mixture-of-experts models whose full parameter count must be resident even though only some experts are active per token.
This is the only band where that last point stops being academic. A 235B or 671B parameter mixture-of-experts model needs room for every expert, and that requirement rules out every other machine on this site.
Who this is genuinely for
Very few people need this, and the price makes that clear. It is for running the largest open models locally for reasons of privacy, cost at volume or simple independence from an API. If your largest planned model is 70B, the 96GB band does that job for a fraction of the money and you should buy there instead.