Apple sold the 2025 Mac Studio M3 Ultra as the kind of desktop that could keep massive AI models in memory. Now, the machine’s highest-memory configurations appear harder to buy.
Reports indicate buyers configuring the M3 Ultra Mac Studio may now see 96GB as the only available memory option after Apple reportedly removed higher-memory tiers from the purchase flow. For developers, researchers, and IT teams evaluating local AI hardware, the change narrows the Mac Studio’s role from a high-capacity AI workstation to a more specific tool for local testing, privacy-sensitive inference, and lower-concurrency work.
Mac Studio memory options shrink
Apple introduced the 2025 Mac Studio with M4 Max and M3 Ultra chips in March 2025, and said the M3 Ultra model could be configured with up to 512GB of unified memory, according to Apple’s launch announcement.
In March 2026, Tom’s Hardware reported that Apple had removed the 512GB upgrade from the M3 Ultra Mac Studio, leaving 256GB as the highest available configuration at the time. The report tied the change to the broader memory shortage driven by AI infrastructure demand, the same pressure point showing up as memory costs rise across hardware.
The ceiling appeared to shrink again in May 2026, when TechRadar reported that the 256GB M3 Ultra Mac Studio option had also been taken off sale, leaving that model available only with 96GB of memory in Apple’s store. Apple has not publicly explained whether the configuration changes are temporary or permanent, and its Mac Studio technical specifications still list a 256GB option for the M3 Ultra model.
That mismatch makes the safest framing a purchase-availability problem, not an official reset of the product’s long-term specification. Local AI has also made high-memory desktops more relevant as organizations weigh workstations, cloud platforms, and regional data center capacity for AI workloads.
What 96GB means for AI workloads
For local AI, memory capacity decides whether a model can fit. Memory bandwidth, software support, and concurrency determine how well it runs, while Apple silicon’s unified memory design lets the CPU and GPU share one pool of memory.
A 96GB Mac Studio can still support smaller models, quantized larger models, coding assistants, private local testing, and experimentation by solo developers or small teams. The limit becomes more important with long context windows, multimodal inputs, repeated internal requests, or shared-workspace AI agents.
Production serving needs separate testing. Frameworks such as vLLM are built around high-throughput features, including continuous batching and key-value cache management, while Apple’s own MLX framework remains better suited to local development and experimentation than a drop-in replacement for established GPU serving stacks.
For procurement teams, Mac Studio remains a strong candidate when privacy, power efficiency, physical footprint, or local development matter more than high request volume. It is less convincing for shared inference services, high-concurrency internal tools, or deployments already built around NVIDIA-oriented infrastructure.
Teams should benchmark the model they plan to run with realistic prompt lengths and simultaneous users. A single clean-prompt test will not show whether 96GB can support day-to-day internal AI use.
Mac Studio still has a place in local AI work, but current memory availability changes the purchase decision. The key question is whether 96GB is enough for the workload, not whether Apple silicon can run serious models locally.
Read more: For more context on how Apple’s supply chain is shifting toward AI hardware, see how an Apple supplier is investing in robots and AI servers through a $1.1 billion Hong Kong listing.