Apple put the Mac Studio on sale on 22 September with "up to 512GB of unified memory" and "1.2TB/s of memory bandwidth" on M5 Ultra. Both sit on the top configuration, so that pairing holds. What Apple measured is far smaller.
Every AI figure on Apple's Mac Studio page rests on one workload: "Time to first token measured with an 8K-token prompt using a 14-billion parameter model with 4-bit quantization." Fourteen billion parameters at four bits works out to about 7GB, 1.4% of the memory sold, and time to first token is prompt processing, which compute sets. The 9.8x headline beats an M1 Ultra; the M3 Ultra beside it already had 512GB. The "4x" in the coverage is the Mac mini's.
Memory bandwidth sets generation. Apple publishes 1.2TB/s. NVIDIA publishes "1,440 GB total, 64 TB/s HBM3e bandwidth" across a DGX B200's eight GPUs, 180GB and 8TB/s each, 6.7 times Apple's figure.
So a Mac Studio holds more in one address space than a Blackwell GPU and reads it 6.7 times slower. Apple's own route to a larger model is four clustered over Thunderbolt 5, footnoted at 72 billion parameters. Neither vendor publishes a tokens-per-second figure on a model genuinely needing 512GB.
The new Mac mini and Mac Studio are available today — 22 September 2026
Capacities, bandwidths and every multiple with its testing footnote are Apple's, from its 22 September release and Mac Studio pages read that day; accelerator figures are NVIDIA's published DGX B200 specification. The per-GPU split, 7GB and 6.7x are ours.
