The New M5 Ultra Mac Studio is an AI Beast That Holds Up to 512GB of Memory

Posted by Andru Edwards Categories: Apple, PC / Laptop, Artificial Intelligence

Apple announced the machine this morning alongside the new M6 and M5 Pro Mac mini. Pre-orders are open now, shipping starts September 22, and the 512GB configuration doesn't arrive until late October. The M5 Max model starts at $2,499.

M5 Ultra is the most powerful chip Apple has ever made

The Ultra scales to a 36-core CPU (12 super cores, 24 performance cores) and an 80-core GPU that Apple calls its most powerful silicon GPU to date. For the first time, the Ultra gets Neural Accelerators inside every GPU core, which is what drives Apple's claim of up to 4.3x the peak AI compute of M3 Ultra and 9.8x that of M1 Ultra.

Memory is the real story. Up to 512GB of unified memory at 1.2TB/s of bandwidth, 50 percent more bandwidth than M3 Ultra. Bandwidth is what determines how fast tokens come out once a model is loaded, so this is the number that decides whether running a big model locally feels usable or feels like a science project.

The M5 Max is the sane one

At $2,499, the M5 Max version gets an 18-core CPU, up to a 40-core GPU, up to 128GB of unified memory, and 614GB/s of bandwidth. Apple says AI performance is up to 3.9x the previous generation, Magic Mask in DaVinci Resolve Studio runs up to 3x faster than on M4 Max, and Redshift scene rendering is 1.4x faster.

For most people who edit video, mix audio, or build 3D scenes for a living, this is the machine. The Ultra serves a much narrower group: people running or fine-tuning very large models locally, and VFX shops pushing uncompressed 8K.

Clustering, and the ports to do it

Thunderbolt 5 arrives with 120Gb/s of bandwidth and up to six ports, and Apple is explicit about the point. Using Thunderbolt 5 and RDMA, you can cluster multiple Mac Studios into one shared memory pool. Apple says four of them together deliver up to 3x faster AI inference than a single system. Storage moves to a PCIe Gen 6 SSD architecture that's up to twice as fast, which matters mostly for how quickly a huge model loads off disk.

The Apple-designed N1 chip brings Wi-Fi 7 and Bluetooth 6 to Mac Studio for the first time. It will drive up to eight displays, or four Studio Display XDRs at full 5K and 120Hz. Genlock over USB-C is here too, for syncing a display to a pro camera.

About that price

$5,499 is a steep entry point, and memory and storage upgrades on top of it are not cheap. Apple is also selling this partly on cloud-cost math: buy once, run models forever, no per-token bill. That math works if you're running inference constantly. If you fire up a local model twice a week, an API subscription will look a lot smarter for a very long time.

Who should buy it

If your work is bounded by how much model you can hold in memory, nothing else sits on a desk and does this, and the 512GB configuration in October is the reason to wait. Everyone else should look hard at the $2,499 M5 Max first, or even the $1,699 Mac mini with M5 Pro, and be honest about whether they would ever use the headroom.

Apple built a desktop that quietly argues the interesting AI work is moving back onto local machines. Whether that argument holds depends less on the hardware than on whether the models worth running keep fitting in 512GB. I personally can't wait to get one of these on my desk!