Mac mini M6 and Mac Studio M5 Ultra were barely announced before the comments section filled with the same question: what exactly is unified memory?
32GB, 128GB, 512GB—these numbers sound similar to those on ordinary computers, but Apple’s memory works completely differently from the way memory works on a Windows PC. Once unified memory’s role is clear, it becomes possible to understand why a desktop Mac can run a 40-billion-parameter AI model locally while an Nvidia graphics card costing nearly US$2,000 cannot.
This is the memory extension of the end-to-end AI hardware supply chain. It starts with unified memory on the consumer side, then looks at its design logic, AI use cases, and the pressure it faces within the supply chain.
What is unified memory? One sentence and one picture
Start with a picture.
In a typical Windows computer, the system memory used by the CPU (DDR5) and the graphics memory on the graphics card (VRAM, usually GDDR7) are two separate pools. Data moves between them over the PCIe bus, a high-speed channel on the motherboard that connects components. When the CPU needs to hand an image to the GPU, it first has to “copy” a version and send it over.
Apple Silicon removes that step. The CPU, GPU, and Neural Engine—the hardware unit in Apple chips dedicated to accelerating AI computation—share the same block of LPDDR5X memory, soldered directly beside the chip, only a few millimeters away. When any processor needs the data, it reads the same block directly instead of making a copy.
In one sentence: Unified memory means “all processors share the same block,” so data does not have to move.
How is it different from memory in a traditional computer?
| Traditional architecture (Windows PC + discrete graphics card) | Apple Silicon unified memory | |
|---|---|---|
| Memory type | DDR5 system memory plus GDDR7 graphics memory, one pool each | One block of LPDDR5X shared by everything |
| How data moves | CPU → PCIe → GPU; a copy is required | No transfer; everything reads the same block directly |
| How much memory the GPU can use | Only the memory on the graphics card (the RTX 5090 has 32GB) | All of it. The M5 Ultra has 512GB, all of which the GPU can use |
| Expansion | System memory can be replaced; graphics-card memory cannot | Soldered at the factory; decided at purchase |
The core difference is “how much memory the GPU can use.”
In a traditional architecture, the Nvidia RTX 5090 has only 32GB of video memory. If the model does not fit, part of it has to be stored in system memory and fed over PCIe in batches, causing a major drop in speed. With unified memory, the pools are not separate: the entire 512GB is within the range the GPU can access directly.
The tradeoff is equally clear: LPDDR5X has lower bandwidth than the GDDR7 used by Nvidia. The RTX 5090 has 1.79TB/s of memory bandwidth, the M5 Ultra has 1.2TB/s, and the M5 Max has only 614GB/s. When the same model fits on both, the Nvidia graphics card computes faster. But once the model is too large for the graphics card’s memory, unified memory’s capacity advantage outweighs the bandwidth gap.
Core data snapshot
The Apple Silicon unified-memory lineup as of August 2026, compared with Nvidia’s flagship graphics cards.
| Chip | Maximum unified memory | Memory bandwidth | Process | Products | Taiwan starting price |
|---|---|---|---|---|---|
| M6 | 32GB | 170 GB/s | 2nm (N2) | Mac mini | NT$29,900 |
| M5 Pro | 64GB | 307 GB/s | 3nm | Mac mini | NT$59,900 |
| M5 Max | 128GB | 614 GB/s | 3nm | Mac Studio, MacBook Pro | NT$84,900 (Studio) |
| M5 Ultra | 512GB | 1.2 TB/s | 3nm (four dies) | Mac Studio | NT$199,900 |
| Nvidia graphics card | VRAM | Memory bandwidth | Suggested retail price | 2026 actual market price |
|---|---|---|---|---|
| RTX 5090 | 32GB GDDR7 | 1.79 TB/s | US$1,999 | Above US$5,000 |
| RTX 5080 | 16GB GDDR7 | 960 GB/s | US$999 | Out of stock at most retailers |
Nvidia wins on bandwidth; Apple wins on capacity. That gap becomes most pronounced when running AI.
Why does unified memory matter for AI?
Running an AI model has two thresholds: memory capacity determines “whether it can run,” while bandwidth determines “how fast it runs.” Capacity comes first; if the model does not fit, even the highest bandwidth will not help.
Consider the task of “running a large language model on your own computer.”
A 70B-parameter model, such as Llama 3 70B, needs around 40GB of memory after quantization. The RTX 5090’s 32GB is not enough, so part of the model has to be moved to system memory and slowly fed over PCIe. Inference speed drops directly from more than 20 tokens per second to single digits. The M5 Max’s 128GB holds it easily, and running it with Apple’s MLX framework—a framework designed specifically for unified memory—can produce more than 10 tokens per second.
A 40B-parameter MoE model, such as Llama 4 Maverick, is even more demanding. It needs more than 200GB after quantization. Even two RTX 5090 cards would provide only 64GB, which is nowhere near enough. The M5 Ultra’s 512GB is the only solution in 2026 that can fully fit this kind of model on a single desktop, without a server or the cloud, at around 12 tokens per second.
The rule is simple: The larger the model, the more obvious unified memory’s capacity advantage becomes. Small models run on both platforms, with Nvidia winning through bandwidth; large models do not fit on Nvidia cards, and Apple wins through capacity.
MLX can perform so well for the same reason: unified memory eliminates the step in the CUDA workflow that moves tensors from CPU memory to GPU memory. The data is already in the same block, so startup is fast and memory utilization is high.
Supply-chain role: the LPDDR5X scramble
The LPDDR5X used for unified memory follows the same memory specification used in phones. There are only three suppliers: Samsung, SK hynix, and Micron.
That became a major problem in 2026.
SK hynix and Micron shifted a large amount of capacity to HBM, the high-bandwidth memory used by GPUs in AI data centers, because it is more profitable and gets production priority. As a result, LPDDR5X capacity was squeezed. In the second quarter of 2026, LPDDR5X prices jumped by nearly 90% in a single quarter, and lead times stretched from the teens of weeks to 40–58 weeks. Samsung filled part of the gap and currently supplies 60–70% of the memory in Apple’s iPhones, but reports say Apple accepted a doubling of Samsung’s quoted price. Tim Cook called this wave of price increases a “once-in-a-century flood” on an earnings call.
The price increases showed up directly in Mac prices. The Mac mini M6’s Taiwan starting price jumped from NT$19,900 for the previous generation to NT$29,900, an increase of more than 50%. The MacBook Air and MacBook Pro also rose by nearly 20% each. In the United States, there was even a federal lawsuit accusing memory manufacturers of price fixing.
The structural problem runs deeper: Nvidia has also begun purchasing LPDDR5X in large quantities for edge AI platforms such as Jetson Thor. Nvidia’s LPDDR purchases are expected to exceed those of traditional smartphone manufacturers, making it the world’s largest source of LPDDR demand. The situation in which Apple and Nvidia compete for the same pool of memory may not ease until memory manufacturers finish expanding capacity in 2028.
Apple tried to find a fourth memory supplier to diversify its risk, but reports say it was unsuccessful. The DRAM market as a whole is not expected to see new entrants before 2027.
The key point
Unified memory does something intuitive: it combines all processors’ memory into one block, with no separation and no copying. This design has taken on new meaning in the AI era. As models grow larger, “memory capacity” becomes the ceiling hit before “computing speed,” making unified memory’s capacity advantage a ticket to running AI locally. The M5 Ultra’s 512GB is currently the largest capacity available in a personal desktop, and it can hold models that previously required a server to run.
But it also ties the Mac’s fate more closely to the memory supply chain. LPDDR5X is squeezed by HBM for capacity and by Nvidia for demand, and soaring prices have directly raised Mac prices across the line. The supply-and-demand struggle around this memory will be a major variable affecting Apple’s product pricing and shipment cadence over the next few years. Now that Nvidia’s edge AI platforms are also starting to consume LPDDR in large quantities, it is still unclear how the memory war between consumers and data centers will divide up.
To continue with another type of memory used in AI data centers, read What is HBM?. To see how the eight stages of the full AI hardware supply chain connect, return to the supply-chain overview.
This article is a knowledge summary of the AI hardware industry. All specifications and prices are based on the time of publication, will be updated over time, and do not constitute investment advice.