Why 16GB Is Now the Real Minimum for a Laptop Built for On-Device AI

A laptop with 8GB of memory is no longer a sensible entry point — it is a machine that is quietly being shut out of the features vendors promote hardest. Windows 11 will still install on 4GB, but the on-device AI layer that sits on top of it now publishes a very different number: 16GB. Understanding that gap between what an operating system needs to start and what its headline features need to work is the most valuable thing you can do before picking a configuration.

• 16GB is the practical floor: Microsoft’s Copilot+ PC programme and Google Chrome’s built-in AI models both set the bar there.
• Apple still sells 8GB Macs, with caveats: Apple Intelligence runs on them, but the memory tiers hint at smaller local models.
• 32GB is for genuine local AI: running language models through LM Studio or Ollama changes the maths, and memory bandwidth matters as much as capacity.

The 16GB floor, written into two spec sheets

Microsoft’s official Windows 11 requirements are modest: a dual-core 1GHz processor, 4GB of RAM, 64GB of storage and TPM 2.0. The Copilot+ PC requirements live in a different league — an NPU capable of more than 40 trillion operations per second (TOPS), 16GB of DDR5 or LPDDR5 memory and a 256GB SSD. As of October 2026, the silicon that qualifies includes AMD’s Ryzen AI 300 and 400 series, Intel’s Core Ultra 200V and 300V families, and Snapdragon X.

The important nuance hides in the footnotes. Microsoft states that some PCs carry a 40+ TOPS NPU yet are not Copilot+ PCs, citing machines with less than 16GB of RAM as the example. A laptop with a modern chip and 8GB is therefore excluded from features such as the Snipping Tool’s Perfect Screenshot.

Hardware already follows that template. The Surface Pro 11 pairs a Snapdragon X Plus with a 45 TOPS Hexagon NPU, 16GB of LPDDR5x and a 512GB SSD; HP’s 27-inch OmniStudio all-in-one uses an Intel Core Ultra 7 355 with a 49 TOPS NPU and the same 16GB of DDR5.

Browsers push the same message. Google’s documentation for Chrome’s built-in AI APIs says that running its local models — Gemini Nano among them — on the processor requires 16GB of RAM or more and at least four cores. The alternative is a graphics chip with strictly more than 4GB of VRAM. Chrome also wants 22GB free on the drive holding its profile, and deletes the model if free space drops below 10GB after download. The APIs work on Windows 10 and 11, macOS 13 or later, Linux and Chromebook Plus — but not on phones.

Apple’s exception, and the fine print behind 8GB

Apple is the outlier worth scrutinising. The MacBook Air with the M5 chip starts at 16GB of unified memory and scales to 24GB or 32GB. The MacBook Pro begins at 16GB with M5, 24GB with M5 Pro and 36GB with M5 Max, topping out at 128GB. Then there is the MacBook Neo, built around an A18 Pro, which ships with 8GB in every configuration and is still compatible with Apple Intelligence — Apple’s requirements page lists it alongside Macs with M1 or newer.

That same page offers a clue about memory pressure. Apple Intelligence uses up to 14GB of storage on Macs with M3 or later and needs at least 12GB of unified memory, while other compatible models get by with as little as 8GB. Apple does not explain the split, but the sensible reading is that machines with more memory receive larger on-device models. If so, the Neo runs Apple Intelligence — just not the same tier as a 16GB Air.

When 32GB stops being overkill

For most buyers, 16GB is enough. The case for 32GB appears once you want to run language models locally with tools such as LM Studio or Ollama. LM Studio’s own guidance recommends 16GB or more on Mac and warns that 8GB only may work with small models and modest context windows. On Windows it asks for at least 16GB plus a GPU with 4GB of dedicated VRAM, alongside a blunt warning that language models can consume a great deal of RAM.

Capacity is only half the story. VRAM determines which models fit, but memory bandwidth determines how fast they answer, and pushing layers into system RAM can slow generation by an order of magnitude — which is why fast DDR5, PCIe 5.0 bandwidth and a capable CPU all count. Creator workloads follow a similar curve: Blackmagic’s Resolve 21.1 asks for 16GB of system memory on Windows, 32GB if you lean on Fusion, and a GPU with at least 4GB of VRAM.

Microsoft has not abandoned low-memory machines. Windows chief Pavan Davuluri has publicly confirmed an 8GB optimisation push, and corporate vice president Marcus Ash described a cross-team effort built around measuring real customer workloads. Microsoft claims reworked memory management keeps up to 2GB more active data in RAM, loads widgets only when opened to free roughly 200MB, and delivers up to 20% better responsiveness — a company-reported figure, not an independently verified one. Testing published before the announcement showed an 8GB laptop idling at 75% memory use against 45% on a 16GB sibling, and hitting 94% with three browser tabs plus Spotify open versus 62%. That is an argument for headroom, not proof that 8GB is comfortable.

Buying advice by profile

• Basic use (browser, email, documents): 16GB on Windows. An 8GB Mac is defensible if price dominates, provided you accept it is an entry-level machine.
• Heavy office work, dozens of tabs, video calls: 16GB minimum, 24GB if the laptop must last five years.
• Local AI, development, photo or video editing: 32GB.

One practical warning: on MacBooks memory is fixed at the moment of purchase, so the decision cannot be deferred. On the second-hand market, check RAM before you check the processor.

The uncomfortable conclusion is that RAM has stopped being a specification for gamers and editors alone. Official minimums stay low because no vendor wants to exclude millions of existing machines, yet the features that fill advertising are designed around 16GB. Expect 8GB Windows laptops to fade from shelves through sheer irrelevance, and expect Apple to explain more precisely what an 8GB Mac cannot do. Until then the rule is simple: buy for the software that is coming, not for the one the installer asks about.