512GB of Memory Changes What You Can Run at Home

The number that decides what you can run at home just moved by a factor of sixteen. Apple’s new Mac Studio scales to 512GB of unified memory. The largest consumer GeForce card holds 32GB.

Apple is unusually direct about why, saying in its own announcement that the configuration exists for “loading the largest and most demanding frontier-class open-weight models available today”.

The two numbers that are not the same

	Memory a model can occupy	Bandwidth
RTX 5090	32 GB	the fastest of these, per gigabyte
Mac Studio, M5 Max	up to 128 GB	up to 614 GB/s
Mac Studio, M5 Ultra	up to 512 GB	1.2 TB/s

Apple’s and NVIDIA’s own published figures. Mac Studio starts at $2,499 with M5 Max and $5,499 with M5 Ultra, with availability from 22 September. A maximum-memory machine costs a great deal more than either starting price.

Capacity is a wall. Bandwidth is a speed limit.

This is the distinction that makes the comparison useful rather than a spec contest. A model that does not fit in memory does not run slowly. It does not run.

Once it fits, how fast it answers is governed largely by how quickly its weights can be read, which is bandwidth. So the two figures answer different questions, and which one binds depends entirely on the size of what you are trying to run.

Below about 32GB the graphics card wins, because everything at that size already fits and the only remaining question is speed. Above it the card is not slower, it is absent, and a machine with the memory is the only machine in the conversation.

What this does not fix

Three honest caveats, because the headline number invites more enthusiasm than the situation supports.

Fitting is not the same as being fast. A very large model on a machine with generous memory and moderate bandwidth will produce text at a pace that makes it useful for batch work and frustrating for conversation.

Software has to exist for the platform. A substantial amount of local AI tooling is written against a specific vendor’s stack first and everything else later, and “runs” and “runs well” are different claims.

The price is not the starting price. These are the figures for the base configurations, and memory is the option that moves the total hardest.

Who this actually changes things for

Not most people. The honest reading is that this is a workstation answer to a workstation problem, and the number of households that need to run a frontier-class open model at home is small.

It matters anyway, for two reasons. It sets a ceiling that consumer graphics cards have not approached, and it makes running a large model privately a purchasable option rather than a research-lab one. Whether that is worth $5,499 before you have added the memory is a question about your work, not about the hardware.

What to actually do

• Size the model before you size the machine. Our VRAM explainer has the arithmetic, and it will usually tell you the answer is smaller than you feared.
• Buy capacity for what you cannot otherwise run, bandwidth for what you use daily. They are different purchases and conflating them is expensive.
• Check the tooling before the hardware. The model you want has to be supported on the platform you buy, not merely theoretically runnable.
• Do not treat a large context window as free. It consumes the same memory the weights do, which is why chatbots forget and why a big context is a hardware decision as much as a software one.