Mac Studio M5 Ultra 512GB Reportedly $22,299: Claimed vs Verified for Local AI

Apple appears to have turned on the biggest Mac Studio configuration yet. MacRumors reports that the Mac Studio with M5 Ultra now offers a 512GB RAM option starting at $22,299, while the M5 Max model stays capped at 128GB, according to a Techmeme summary of the report. We could not open the MacRumors article itself, and the price does not appear on Apple's public shop page as we read it, so this post separates what Apple said in August from what is newly claimed.

The short version: the 512GB option is real and was announced, the late-October timing matches Apple's plan, and the $22,299 figure is plausible but unconfirmed here.

TL;DR: claimed vs verified

What did Apple announce in August?

Apple unveiled the new Mac Studio on August 25, 2026, with preorders that day in 30 countries and regions and shipping from September 22, per iClarified's launch report and Macworld. Our own launch analysis covered the chips. The key specs from that coverage:

The two outlets disagree on the M5 Max bandwidth, so check Apple's tech specs before relying on either number. Apple also says a four-system cluster over Thunderbolt 5 with RDMA can run distributed inference up to 3x faster than a single unit, and that the M5 Ultra uses a quad-die design, which it describes as a first for an M-series chip.

The 512GB tier was not orderable at launch. Macworld noted the most expensive M5 Ultra then available was the 256GB, 16TB model at $18,299, and the page for the 512GB configuration carried no price.

Why would the 512GB option matter for local AI?

Unified memory is the reason people buy Mac Studios for large models. One pool of RAM is shared by CPU and GPU, so a single machine can hold weights that would need several discrete GPUs on a PC. Apple's own research shows the practical ceiling. In its MLX distributed-inference post, the 816B-parameter Muse Spark Open needs about 493GB in 4-bit NVFP4 form before runtime overhead. That is why it ran on four M5 Ultras rather than one: 512GB leaves almost no room for the key-value cache and the operating system.

So what does 512GB actually buy?

1. Large mixture-of-experts models in 4-bit form. Models in the several-hundred-billion range fit, but with tight headroom.
2. Long contexts on mid-size models. A 70B to 130B model plus a very large context window can need more memory than the weights alone.
3. Several models resident at once. An agent setup with a coder, a reranker and a vision model loaded together stays out of swap.
4. Cluster building blocks. Four 512GB machines give 2TB of pooled memory for models that no single box can hold.

What it does not buy is speed. Apple's data shows decode speed tracks memory bandwidth, so more capacity at the same 1.2TB/s lets you load bigger models, not generate faster. If your model already fits in 256GB, the upgrade changes headroom, not tokens per second.

How much memory does a model need? A sizing rule of thumb

A quick rule helps decide which tier you need. In 4-bit form a model takes roughly half a byte per parameter, so weights alone need about 0.5GB per billion parameters. Add headroom for the key-value cache, which grows with context length, and for macOS and your other apps. This is our own back-of-envelope arithmetic, not an Apple figure, and real files vary by quantization format. Apple's measured example is consistent with it: 816B parameters at 4-bit lands near the 493GB Apple reports, a little above the naive 408GB because of format overhead.

Two practical notes follow. First, the context window is the quiet memory hog: an agent that keeps a whole codebase in context can double the footprint of a mid-size model. Second, mixture-of-experts models hold every expert in memory even though only a few run per token, so a model advertised as "42B active" still needs the full 816B parameters resident, as Apple's Muse Spark Open test shows.

If you are deciding between tiers, the 256GB M5 Ultra covers most open models people actually run today. The 512GB tier is for the narrow band of people who want 400B-class models on one desk, or who plan to chain several machines and want each node to carry a full share.

Is $22,299 plausible?

The claimed price sits $4,000 above the $18,299 Macworld listed for the 256GB, 16TB configuration. That gap is consistent with a memory upgrade on top of a maxed storage option, though we cannot tell which storage size the $22,299 figure assumes, since the headline does not say. Memory pricing context also fits: DRAM supply has been tight all year, as our coverage of SK Hynix's record quarter and memory shortage explains, and Apple removed the 512GB option from the previous generation in March 2026 amid the shortage, per MacRumors' earlier reporting.

We would not treat the figure as final until it appears on Apple's configurator. Regional prices will differ, and storage choice changes the total. The headline number is the "starting at" price for that RAM tier, which is not necessarily the cheapest machine you can order with 512GB.

How does it compare with other local AI boxes?

Apple is not the only option in this class. Our coverage of the Gigabyte W775-V10-L01 GB300 deskside workstation and the Surface RTX Spark dev box shows the alternative: Nvidia silicon with faster compute and software ecosystem, but separate memory pools or smaller capacity. The trade is roughly this:

Apple has also been building the software side of this story, including its Muse Spark Open partnership with Meta and a hands-on Verge local-AI diary that shows what living with a Mac Studio for agents is actually like.

Should you buy the 512GB config?

Ask these questions before you spend five figures:

• What is the largest model you run, in 4-bit form? If it is under about 200GB, you do not need 512GB.
• Do you need it in one box? Two or four 256GB machines may serve a cluster-friendly workload better, if you accept the setup effort.
• Is tokens per second or capacity your bottleneck? More RAM will not fix slow decode.
• Are you waiting for the next generation? Apple's own M6 iMac and MacBook Pro local-AI details suggest the chip roadmap is moving quickly.
• What does cloud cost instead? Renting large-memory GPUs by the hour can beat a $22,299 purchase if usage is occasional.

What to verify yourself

1. Open Apple's Mac Studio configurator and select M5 Ultra to see whether 512GB is orderable in your region.
2. Compare the quoted price with the storage you choose; the headline figure may assume a specific SSD size.
3. Check Apple's tech specs for M5 Max memory bandwidth, given the 460GB/s versus 614GB/s discrepancy between outlets.
4. Read the original MacRumors article for delivery estimates, which we could not see.

Accurate as of October 28, 2026. The $22,299 figure comes from a Techmeme summary of a MacRumors report that we could not open; all other specifications come from launch-day coverage and Apple's shop page.

Related reading

• Mac Studio M5 Max and Ultra for local AI (August launch)
• Apple MLX plus RDMA on four M5 Ultras
• Running a 1-trillion-parameter model on Mac Studio
• Apple and Meta: Muse Spark Open on Mac Studio
• Gigabyte GB300 deskside workstation vs DGX Spark
• Surface RTX Spark dev box
• Official: Apple Mac Studio shop page