Apple's M5 Ultra Mac Studio sells 512GB of memory, and the options list is the risk

The configuration menu is the interesting document here, not the spec sheet. Apple is again offering 512GB of unified memory on the top Mac Studio ^[1], after spending the past year reducing the M3 Ultra machine's memory choices to 96GB and nothing else ^[9]. That restored ceiling is 5.3 times what was actually orderable a day ago ^[19]. On the M5 Max tier, the 128GB maximum is a third more than the 96GB the M4 Max version had been trimmed to ^[22]. No silicon decision produced those two numbers. Supply did.

The chip underneath is two M5 Max dies bonded by a 4.4TB/s UltraFusion layer, which Apple describes as its first quad-die design ^[3]. Externally that shows up as up to 1.2TB/s of memory bandwidth ^[6], roughly double the M5 Max's 614GB/s ^[20], and as up to 80 GPU cores carrying Neural Accelerators for the first time on an Ultra part ^[2]. Apple's headline figure for the tier is LLM prompt processing in LM Studio, up to 4x an M3 Ultra ^[8]; Engadget records a claim of up to 4.5x faster AI compute against the same predecessor ^[24]. Every one of those numbers is Apple's own and untested.

The pricing tells you where the shortage was absorbed. The Ultra starts at $5,499 with 1TB of storage, $200 above the M3 Ultra configuration ^[13], which is 2.2 times the M5 Max entry price for approximately twice the silicon ^[21]. The M5 Max Studio holds at $2,499 ^[12]. The Mac mini, meanwhile, goes to $899, a $100 rise that Engadget counts as $300 against older minis ^[4][5]. So the RAM crisis reaches the cheap machine as a price and the expensive machine as a shorter list of things you are permitted to buy.

Engadget places the M5 Ultra Studio with the likes of Pixar rather than typical creatives ^[17]. That reading sits awkwardly next to the same publisher's account of why Mac minis kept selling out: buyers running local models, sometimes in clusters, to the point that Apple discontinued the cheapest mini in May 2026 ^[7]. Clustering adds throughput, but each M6 node still holds only its own 32GB at 170GB/s ^[4], and the M5 Pro mini stops at 64GB ^[16]. A single 512GB pool is 16 times the M6's ceiling and eight times the M5 Pro's ^[23]. For anyone whose constraint is which weights fit rather than how fast tokens arrive, that is the only comparison on the page that matters.

Which makes this a timing problem rather than a performance-per-dollar problem. The benchmark claims will hold for a year. The order page will not: the 512GB line item is precisely the one that was withdrawn last cycle, and a budget that clears in six months may find itself buying a 96GB machine with an 80-core GPU attached to it.