The decision actually gets made in the configurator, with a list of model weights open in another tab. You click up the memory on an M5 Max, watch the total climb faster than it used to because memory and storage upgrades cost more than usual this cycle [5], and start wondering whether the Ultra is the cleaner answer. Apple's memory ladder makes that comparison worse than it looks. The Ultra opens at 96GB [6], which is 32GB less than the most you can order in a Max [21], so the entry premium hands you cores and bandwidth while taking memory away.
Work backwards through the pricing history and the widening is easy to size. MacRumors puts the $5,499 start at $1,500 above the M3 Ultra's launch price and $200 above what the M3 Ultra cost after June's increase [3], which sets the launch price at $3,999 [16] and the post-June price at $5,299 [17]. June added $1,300 to the Ultra tier [18] while the Max tier sat still at $2,499 [2]. The spread between the cheapest and dearest Mac Studio has gone from $2,800 to $3,000 [19], and the Ultra now opens at roughly 2.2 times the price of the Max [20].
Teams tell themselves they are buying headroom for whatever the models look like in eighteen months, but what they actually run is a quantised checkpoint with a resident footprint someone can measure this afternoon. That measurement is the decision. MacRumors' read is that the Ultra only separates itself at 256GB and above, and that Apple restricts the 512GB option to the 80-core GPU model, so the configurations that vindicate the machine sit far above its starting price [9]. Its own guidance sends the overwhelming majority of buyers, most professionals included, to the Max [10], and reserves the Ultra for people who expect to hit a ceiling the Max cannot clear, with whole-model local inference as the obvious case [11].
The Ultra's case weakens further, too. Memory is fixed at purchase, which is why the guide argues for spending on unified memory rather than the higher chip tier [12], the opposite of the usual upgrade instinct. And Apple says four Mac Studio systems clustered over Thunderbolt 5 and RDMA deliver up to 3x faster AI inference than one machine [13], which separates throughput from capacity.
So the forcing question has two axes rather than one. First, does the working set have to be resident all at once, or can it be split across machines? Second, is it bigger than 128GB, the Max ceiling [7]? Resident and under 128GB is a Max with memory maxed, and the money left over goes to storage and displays [23]. Resident and over 128GB is an Ultra at 256GB, priced in the knowledge that the 512GB rung requires the 80-core GPU [9]. Splittable and over 128GB is a cluster [13]. Splittable and under 128GB was never an Ultra question at all.
None of that requires a view on where local models are heading; it comes down to one number, peak resident bytes under a real job, measured before anyone opens the configurator.
