Spend once, spend correctly
Most “AI laptop” shopping lists are written for people with a budget and a benchmark score. This one is written for an 8 GB machine: what removes a real bottleneck, what is a waste of money, and what to buy used.
The single upgrade that turns “a 3B model, maybe” into “a 7B model with a browser open.” If your laptop has a free slot or replaceable sticks, this is the first purchase.
Check your exact model's slot layout before buying. Single-channel 8 GB → dual-channel 16 GB also improves CPU inference speed.
Check current price →Model files are large. Four quantized models can fill a 256 GB drive. A bigger SSD also cuts model load time versus an old HDD, though not generation speed.
Clone the old drive before swapping. A clean install is usually faster than a clone on a 2018 machine.
Check current price →If your laptop is soldered and slow, a three-year-old business mini PC often costs less than a RAM upgrade plus SSD and runs circles around it. No GPU, but a desktop CPU with a higher power limit.
Look for 8th-gen Intel or newer, 16 GB RAM, and an NVMe slot. Avoid anything with a 35 W+ CPU in a tiny chassis unless you accept fan noise.
Browse refurbished →Not glamorous, but sustained CPU inference throttles a thin laptop within minutes. Elevating the chassis and adding airflow recovered roughly 10–15% of generation speed in testing.
A simple stand that lifts the rear edge is enough. Fancy RGB fan pads add little beyond noise.
See options →Whisper is only as good as the audio. A $50 USB mic with a cardioid pattern beats a laptop's built-in array for meetings and interviews, especially in a room with echo.
Record at 16 kHz mono for Whisper. A 48 kHz stereo file only adds conversion work.
See options →Keeps a growing model library off the system drive. USB 3.0 is fast enough for loading; do not run inference from a spinning USB hard drive.
A 500 GB external SSD holds roughly 100 quantized 7B models. You will not need that many.
See options →External GPUs. A Thunderbolt eGPU enclosure plus a used GPU costs more than a better laptop, and the 2018 ultrabook's 4-lane Thunderbolt link caps the benefit. It is a legitimate path on newer hardware; it is not a sensible one here.
“AI” USB sticks and NPU dongles. Support is inconsistent, model formats are limited, and the price per token/second is poor compared with a used mini PC or a cloud free tier.
32 GB kits for this laptop. The CPU can address it, but the 15 W chip becomes the bottleneck long before 32 GB pays off. Buy 16 GB and put the difference toward an SSD.
RGB anything. It does not make tokens generate faster. We checked.
Every item on this page is here because it removes a bottleneck we hit on the test bench, or because it made a measurable difference to a workflow documented in a field report. Prices are approximate US street prices and change often — the links go to current listings rather than fixed prices.
We do not list a product solely because it has an affiliate program. If a cheaper item does the same job, it gets the recommendation.