Apple's quad-die M5 Ultra puts 512GB of inference memory under the desk

Somebody at your firm wants an agent triaging tickets overnight, on a machine nobody logs into. The number that decides whether that runs on a desk or on somebody's API is memory, not cores. Apple has now printed it: 512GB on the M5 Ultra ^[11], sixteen times the 32GB on the base M6 ^[15]. The working set either fits under that ceiling or it does not, and the 80-core GPU does not move the line ^[9].

The clustering claim is the part worth doing the arithmetic on. Thunderbolt 5 runs at 120Gbps, which is 15GB per second ^[12]. UltraFusion moves more than 4.4TB per second between dies inside a single M5 Ultra ^[8]. The fabric between machines is therefore about 293 times slower than the fabric inside one machine ^[16], and roughly 80 times slower than that machine's own memory bandwidth ^[21]. Four maxed Mac Studios would present 2TB of pooled memory ^[18], and Apple's figure for four networked desktops is up to three times faster distributed inference, not four ^[12]. That missing fourth is the price of the seam, stated by the vendor.

One more derived number, because the comparison Apple chose is with itself: if 1.2TB/s is 50% up on M3 Ultra ^[11], the older chip was at 800GB/s ^[17].

What the announcement does not contain is a price ^[22]. That is the whole problem with treating this as a buy-versus-rent reset. Rent is billed per token as consumed; buy is billed once, whole, in advance, and nobody can compare the two with one side blank. The performance claims that are available are Apple's own and baselined on M1: up to 4x AI performance, up to 13.5x LLM prompt processing ^[6]. M6 is the sixth generation of Apple silicon, so that baseline sits five generations back ^[19]. It is a good argument for replacing an old Mac and a weak argument against an API bill.

Teams imagine a fleet of desktops quietly absorbing agentic work, with the cloud invoice going to zero. The spec sheet supports something narrower. Two things decide it, and neither is peak throughput. First, whether the working set fits in the memory of one box ^[11]. Second, whether the duty cycle is continuous or bursty. Owning wins on arithmetic you can check only when the work fits and runs continuously, because the box is busy and the marginal token is free once bought. When the work fits but runs in bursts, you have bought idle silicon and should have rented. When the work does not fit and runs continuously, you are across the Thunderbolt seam ^[16] or into a data centre. When the work does not fit and runs in bursts, you are paying for an API call with extra steps.

There is a gate that overrides all four cells, and it is the one PCMag says Apple is aiming at: data that is not permitted to leave the building, which is the privacy-first, on-device case ^[13]. If that applies, the memory ceiling is your architecture, and the decision rests on capacity rather than cost per token.

Until prices appear, the defensible version of the local-agent case is the memory case rather than the speed case. If the working set fits in 512GB and the machine would be busy most hours of most days, the purchase argues itself. If it runs two hours a day, you have bought a very quiet space heater with excellent memory bandwidth.