For more than a decade, 8GB of RAM was the “good enough” standard. It was the safe baseline for students, office workers, and even light creatives. Apple sold it as the default on MacBooks from 2012 to 2023. Windows laptops lived comfortably in the 8GB zone for browsing and spreadsheets.
That era is over. It didn’t end because web pages got bigger (though they did). It ended because the fundamental way we use computers is shifting from retrieving data to generating it.
The rise of the “AI PC” and local Large Language Models (LLMs) has introduced a new hard floor for hardware. Microsoft’s Copilot+ standard mandates 16GB of RAM. Apple’s M4 Macs finally start at 16GB. This isn’t just upsell marketing—it’s a matter of physics. If you try to run modern local AI on an 8GB machine, it doesn’t just run slowly; it effectively breaks.
Here is the technical reality of why your 8GB laptop is becoming a brick.
The Math of the “Brick”
To understand why 8GB is dead, we have to look at how Large Language Models (LLMs) use memory. Unlike a browser tab that can be lazily written to the SSD (swap file) when not in use, an LLM needs its “weights”—the parameters that define its intelligence—residing in active, high-speed RAM to function.
Let’s look at the numbers for a standard “small” local model, like Llama 3 8B or Microsoft’s Phi-3, running at a usable 4-bit quantization.
That seems fine, right? You have 8GB. You verify the math: . You have 4GB left.
Wrong.
You forgot the operating system. Windows 11, when idling with a few background services, easily consumes 4GB to 5GB of RAM.
You are now at 112% memory usage before you have even opened a web browser.
The Swap Death Spiral
When your required memory exceeds your physical RAM, the operating system moves data to the “Swap” or “Page File” on your SSD.
On a modern NVMe SSD, read speeds are fast—maybe 5GB/s or 7GB/s. But RAM is an order of magnitude faster, often exceeding 60GB/s to 100GB/s. More importantly, latency (the time to find the data) is the killer.
When an LLM generates text, it scans its weights repeatedly for every single token generated. If those weights are in the swap file, your CPU/NPU has to pause and wait for the SSD to fetch the data.
- RAM Access: Naposeconds.
- SSD Access: Microseconds (1000x slower).
The result is “Swap Thrashing.” The system spends more time moving data between RAM and SSD than it does actually computing. Your tokens-per-second count drops from a conversational 20 T/s to a painful 0.2 T/s. Your cursor freezes. The audio stutters. The machine is “bricked” until you kill the process.
Contextual History: The Long Sunset of 8GB
The 8GB standard has been remarkably resilient.
- 2012: The MacBook Pro Retina launches with 8GB as a premium baseline. It was overkill for most.
- 2016-2020: 8GB becomes the standard. Chrome gets hungrier, Electron apps (Slack, Discord) proliferate, but 8GB holds the line for general productivity.
- 2023: Apple releases the M3 MacBook Pro with 8GB of RAM for $1,599. Reviewers and professionals revolt, calling it “unacceptable,” but Apple defends it, claiming “8GB on Mac is like 16GB on Windows.”
That claim, while debatably true for video editing due to unified memory efficiency, falls apart for AI. An 8 Billion parameter model doesn’t care if it’s on macOS or Windows; it requires mathematically fixed space. 8GB is 8GB.
In 2024, the narrative shifted rapidly. Microsoft announced the “Copilot+ PC” specification, a hardware standard for the next generation of Windows laptops. The requirements were strict:
- NPU: 40+ TOPS (Trillions of Operations Per Second).
- RAM: 16GB Minimum.
There is no Copilot+ PC with 8GB of RAM. It doesn’t exist. Microsoft, knowing the telemetry of their own OS and AI models, drew a line in the sand. They signaled to the entire industry that 8GB is no longer capable of delivering the intended experience.
The NPU Bottleneck
There is another layer to this: Unified Memory.
In modern “AI PCs” (Snapdragon X Elite, Apple Silicon, Intel Core Ultra), the NPU (Neural Processing Unit) shares memory with the CPU and GPU. There is no dedicated VRAM for the AI chip.
This means if you want to use that shiny new NPU to blur your background in 4K, generate an image, or run a live translation agent, that memory comes directly out of your system RAM.
Video cards (GPUs) have had dedicated VRAM for decades for this very reason. A gaming PC might have 16GB of System RAM plus 8GB of Video RAM. A modern laptop has 16GB total. If the AI takes 4GB and the GPU takes 2GB for the display buffer and UI, the CPU is left fighting for scraps.
On an 8GB machine with shared graphics and NPU, the “usable” memory for applications might be as low as 3GB or 4GB. That is barely enough to run a modern web browser with five tabs open, let alone a productivity workflow.
Forward-Looking Analysis: 32GB is the New Logic
If 16GB is the widely accepted minimum for running AI today, what should you buy to future-proof yourself?
The answer is increasingly 32GB.
As local models get smarter, they get bigger. We are already seeing “Quantization-Aware Training” and “Mixture of Experts” (MoE) models that try to be efficient, but the demand for intelligence is infinite. Users will want to run:
- A text summarization agent (Background).
- A coding assistant (In-IDE).
- A voice transcription service (Meeting).
Running multiple models concurrently with your OS and browser will eat 16GB for breakfast.
We are entering a cycle of hardware obsolescence similar to the early 2000s, where a computer from three years ago couldn’t run the latest software. For the last decade, year-over-year gains were marginal. A 2018 laptop is still fine for web browsing today.
But a 2018 laptop—or even a 2023 laptop with 8GB of RAM—cannot participate in the local AI revolution. It is functionally cut off from the defining features of the next OS updates.
The Verdict: If you are buying a laptop today, do not buy 8GB. Do not let friends buy 8GB. It is not a budget option; it is a dead end. For the first time in a decade, buying the base model isn’t just frugal—it’s a mistake.
🦋 Discussion on Bluesky
Discuss on Bluesky