Apple gives Mac Studio the M5 Ultra and a $5,499 local AI pitch

Why it matters

Apple is positioning the M5 Ultra as a local AI workstation, using up to 512GB of unified memory and clustering to challenge CUDA-centric desktop systems at a sharply higher Mac price.

Apple announced a new Mac Studio on August 25th, pairing its existing M5 Max with the new M5 Ultra in a compact desktop aimed squarely at developers, researchers and creative professionals running demanding work locally. The machines start at $2,499 and $5,499, respectively, with preorders open ahead of deliveries on September 22nd, according to Apple's announcement and The Verge.

The human behind that pitch is Johny Srouji, Apple's chief hardware officer. According to Apple's leadership biography, Srouji joined Apple in 2008 to lead development of the A4, Apple's first internally designed system-on-chip, after senior processor engineering roles at Intel and IBM. The M5 Ultra extends the strategy he has spent nearly two decades building: own the processor, memory architecture, operating system and developer frameworks, then sell the resulting integration as an alternative to assembling a workstation around third-party components.

This release also cleans up an awkward product split. Apple's previous Mac Studio lineup combined an M4 Max with the older-named M3 Ultra because Apple never produced an M4 Ultra. Both tiers now sit within the M5 generation, even if the industrial design remains the same enclosure Apple introduced on March 8th, 2022. The rear USB-A ports have survived another chip cycle.

Srouji's silicon strategy reaches four dies

The M5 Ultra is assembled from two M5 Max packages, each of which already contains two dies. 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 specifications list a 36-core CPU, an 80-core GPU, a 32-core Neural Engine and 1.2TB per second of memory bandwidth.

The larger engineering bet sits in unified memory. M5 Ultra configurations scale to 512GB shared across the CPU and GPU, giving local AI programs room to load models that would exceed the memory capacity of many individual workstation graphics cards. 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. Those figures come from Apple's own tests and cover selected workloads and configurations. They establish Apple's intended positioning; independent testing will determine how broadly the gains carry across model inference, training, rendering and software development.

The underlying memory improvement is easier to quantify. Apple's M3 Ultra offered over 800GB per second of bandwidth and the same 512GB ceiling. M5 Ultra raises bandwidth to 1.2TB per second while retaining that maximum capacity. Apple also says storage can run up to twice as fast as in the previous generation.

Apple wants the Mac to become an AI workstation

Apple's announcement spends much of its attention on local models, coding agents and image generation. That focus explains why Srouji's group added Neural Accelerators to every GPU core and why memory, rather than a redesigned enclosure, carries the product story.

Mac Studio can also connect multiple systems over Thunderbolt 5 using remote direct memory access. Apple says a four-machine cluster can deliver up to three times the AI inference performance of one system. Apple is pitching that arrangement as a shared pool for large open-weight models, giving small research groups and engineering teams a path to add capacity one desktop at a time.

Software remains decisive. Apple points developers toward MLX, its machine-learning framework for Apple silicon, along with the new Core AI framework in macOS 27 Golden Gate. Those tools support running, training and fine-tuning models locally, but developers remain tied to Apple's Metal and MLX software stack rather than Nvidia's CUDA. RuntimeWire recently covered one practical corner of that developer base when Ammaar Reshi used Claude and Fable to bring Command & Conquer Generals to Apple silicon.

Nvidia is pursuing the same desk-side buyer through DGX Spark, a $4,699 system with 128GB of coherent memory, one petaflop of claimed FP4 AI performance and the CUDA software stack. Nvidia says DGX Spark can run inference on models with up to 200 billion parameters. Apple's base M5 Ultra configuration costs $800 more and begins with 96GB of memory, while offering a much higher memory ceiling in configurations whose prices remain unspecified.

Those specifications do not support a direct performance verdict. Nvidia's FP4 figure and Apple's workload comparisons measure different things. They do show the commercial contest taking shape: Nvidia sells a purpose-built AI development box around CUDA, while Apple is asking developers to treat the Mac they use for Xcode, video and 3D work as the local AI machine too.

The chip generation is unified; the pricing has moved

The M5 Max Mac Studio starts with 36GB of unified memory and 512GB of storage for $2,499. That matches the price of the outgoing M4 Max configuration after a $500 increase in June, The Verge reported.

The M5 Ultra version starts with 96GB of memory and 1TB of storage for $5,499. The Verge's pricing comparison puts that $200 above the final starting price of the outgoing model and $1,500 above the $3,999 launch price of the comparable M3 Ultra machine in March 2025. Buyers are getting a newer CPU architecture, faster storage, higher memory bandwidth, Wi-Fi 7 and Bluetooth 6. The enclosure and base memory allocation are unchanged.

Apple says the M5 Ultra model supports up to eight displays, including four 5K, 120Hz Studio Display XDR units, and its media engine can play 33 simultaneous streams of 8K ProRes 422 video at 30 frames per second. Apple's specifications list support for up to five external displays on the M5 Max model.

The M5 Ultra gives Srouji's organization a coherent high-end Mac lineup again and a credible way to pitch local AI without turning the Mac Studio into a single-purpose appliance. Apple has supplied the memory and the marketing numbers. The buyers considering a $5,499 starting point will want application-level results next.