If you're planning to buy an M5 MacBook and upgrade the RAM later, stop: it's not possible. On every Mac with Apple Silicon chips from M1 through M5, the memory is soldered directly onto the chip itself, making it a permanent part of the hardware. Unlike older MacBooks or Windows laptops where you could pop open a panel and swap out RAM modules, modern Macs force you to choose your memory configuration at checkout, and that decision is final.

Why Can't You Upgrade M5 MacBook RAM?

The reason comes down to Apple's unified memory architecture, a design choice that actually makes these machines faster but comes with a significant trade-off. Since the M1 chip launched in 2020, Apple stopped treating RAM as a separate component. Instead, the memory is built into the same package as the CPU, GPU, and Neural Engine, all sharing one pool of memory bandwidth. This is genuinely smart engineering for performance: the GPU and CPU aren't shuffling data back and forth between separate memory pools like on a PC with a discrete graphics card.

But this unified design means the RAM amount gets locked in the moment you complete your purchase on Apple's website. The memory is soldered directly onto the logic board, not plugged into a slot. Even the storage on most post-2018 Macs is soldered, which is why you can't upgrade either component after purchase.

Apple prioritizes thinness and battery life in its design, and soldered memory helps achieve both. Soldered chips sit closer to the processor, use shorter electrical traces, and need less board space than a socketed module. That tight integration is part of why M-series Macs get genuinely excellent battery life. But it's a one-way door: once you buy, you're stuck with what you chose.

What Does This Mean for Running AI Models on M5 Macs?

The unified memory architecture actually becomes an advantage when you're running local AI models. Unlike PCs with discrete graphics cards that have separate VRAM pools, M5 Macs make all system RAM available to AI inference workloads. An M5 MacBook Pro with 16GB of unified memory has roughly 16GB available for model weights, minus what macOS consumes at idle (typically 4-6GB). There's no PCIe bottleneck slowing things down when a model exceeds GPU memory.

The M5 Max offers up to 819 GB/s of memory bandwidth, which means AI models can access that shared memory pool at full speed. A 16GB MacBook Air running a 7-billion-parameter model actually outperforms most PC setups with discrete 8GB graphics cards on the same workload, despite the PC's nominally higher GPU specifications.

How to Choose the Right RAM Before You Buy

• Treat RAM selection as a 5-6 year decision: Since you can't upgrade later, buy the RAM tier you'll need in year 4 or 5, not just what you need today. Going from 8GB to 16GB is often the single best money you can spend on a new MacBook, sometimes more impactful than a CPU upgrade.
• Understand your actual memory needs: An 8GB M5 Mac can run AI models between 1.7 billion and 8 billion parameters cleanly. A 16GB Mac handles up to 14 billion-parameter models. If you're running larger models like Llama 3.3 70B, you'll need 64GB or more of unified memory.
• Monitor memory pressure on current machines: Open Activity Monitor and check the Memory Pressure gauge under the Memory tab. If it sits in yellow or red, you're experiencing memory constraints. Close background apps like Chrome tabs, Slack, or Discord, which consume disproportionate amounts of RAM compared to native Mac apps.
• Use quantization to reduce memory footprint: If you're running AI models, setting OLLAMA_KV_CACHE_TYPE=q8_0 can free 2-4GB of unified memory with negligible quality impact, allowing you to run larger models on smaller RAM configurations.
• Keep at least 15-20% of your SSD free: A nearly full drive forces macOS to rely more heavily on swap memory (virtual memory stored on the SSD), which slows performance and wears on the drive over time.

The unified memory architecture means that when you're running local AI models like Llama or Qwen on your M5 Mac, all available system RAM becomes model memory. Metal GPU acceleration activates automatically on Apple Silicon, with no configuration required. You can verify GPU usage is active by running "ollama ps" in the terminal; the Processor column should read "GPU," not "CPU." A 7-billion-parameter model generating 6 tokens per second is running on CPU, while the same model at 35-55 tokens per second is running on the Metal GPU.

What About Older MacBooks?

If you own a pre-2012 MacBook or MacBook Pro (non-Retina), you're in a different situation. These machines had actual SO-DIMM slots and could be upgraded with standard RAM modules. But every MacBook Air ever made, every MacBook Pro Retina or Touch Bar model from 2012 onward, and all M-series Macs have soldered memory with no upgrade path.

Technically, a repair shop could swap an entire logic board for one with more built-in RAM, since the RAM lives on that board. But this isn't really an upgrade; it's closer to buying a new Mac's internal components and stuffing them in your old case. The labor cost plus board cost usually approaches or exceeds buying a new machine outright, so it's not a practical option unless the logic board already needs replacing for another reason.

The bottom line: M5 MacBooks are exceptional machines for running local AI models thanks to their unified memory architecture and Metal GPU acceleration. But that same architecture means you must make your RAM choice carefully at purchase. Don't assume you'll "add more later if needed." That option simply doesn't exist on modern Apple Silicon Macs.