Published · 11h agoBuild3 min read
The M5 Ultra is pitched as an alternative to CUDA workstations, but the 512GB configuration that carries the argument ships in late October with no announced price.
Written for builders.See today for builders
Compiled by The EngineerSomething wrong?How this is made
The figure a purchase order needs is not on the page. The $5,499 tag attaches to the entry M5 Ultra configuration [2], while the 512GB machine that carries Apple's local-model argument arrives in late October with no announced price [5]. A team can budget the tier. It cannot yet budget the box it actually wants.
Nvidia's competing desk-side system does have a complete number. DGX Spark is $4,699 for 128GB of coherent memory, one claimed petaflop of FP4, and inference on models up to 200 billion parameters, according to Nvidia [12]. That works out to roughly $36.70 per gigabyte of addressable memory [3]. Apple's ceiling is four times the capacity [4] and the entry Mac Studio is $800, or about 17 percent, above the Spark [1], but these sources never state how much memory $5,499 includes, so the per-gigabyte comparison that decides the purchase cannot be finished. The one hard improvement at the top of the range is bandwidth: over 800GB per second on M3 Ultra against 1.2TB per second now [7], which Apple puts at 50 percent higher [8]. The 4.3x peak AI compute figure is Apple's own testing on selected workloads and configurations [6].
The clustering pitch invites arithmetic Apple did not do. Four Mac Studios linked over Thunderbolt 5 with remote direct memory access deliver up to three times the AI inference performance of one, by Apple's claim [9]. At entry pricing that is $21,996 for three machines' worth of throughput, so the fourth desk returns about three quarters of what the first one did [2], and that assumes the cheap configuration rather than the 512GB parts nobody can price [5].
The software commitment is heavier than the usual CUDA-versus-MLX framing admits. Apple's route runs through MLX and the new Core AI framework in macOS 27 Golden Gate, with developers tied to Metal rather than Nvidia's stack [10]. If that bet sours, the hardware does not fall back to Linux in any useful sense: Phoronix notes that Linux 7.3 reaches basic console boot on M3 Pro, Max and Ultra with no GPU acceleration [14]. Depreciation on a workstation assumes some second use. Here the second use is Apple's frameworks or a very fast web browser.
Then there is the detail from the other machine launched the same day. Tom's Hardware reports the M6 Mac Mini is capped at 32GB of unified memory, down from the 64GB available on the M4 Pro version, and attributes the cap to the component shortage affecting the whole industry [13]. Apple is holding 512GB on the Mac Studio while trimming a small desktop by half. Anyone planning to sign for high-capacity units in late October is buying into the same memory market that already forced one product backwards.
Ranked by verification strength, evidence, and original report placement.
Apple announced a new Mac Studio on August 25th, pairing its existing M5 Max with the new M5 Ultra, aimed at developers, researchers and creative professionals running demanding work locally.
The new Mac Studio machines start at $2,499 for M5 Max and $5,499 for M5 Ultra, with preorders open ahead of deliveries on September 22nd.
The M5 Ultra is assembled from two M5 Max packages, each already containing two dies, and Apple's interconnect presents the resulting four-die design to software as one system-on-chip, according to The Verge.
At the top end, Apple's M5 Ultra specifications list a 36-core CPU, an 80-core GPU, a 32-core Neural Engine and 1.2TB per second of memory bandwidth.
M5 Ultra configurations scale to 512GB of unified memory shared across CPU and GPU; the first 512GB systems are due in late October and Apple has not published their price.
Apple says the M5 Ultra delivers up to 4.3 times the peak AI compute, 1.8 times the graphics performance and 1.3 times the CPU performance of the M3 Ultra; the figures come from Apple's own tests and cover selected workloads and configurations.
The reporting this story was synthesized from, earliest first. Every link goes to the original.
