If you’re interested in running text-to-image models and large language models locally but buying a desktop isn’t an option then consider purchasing a used gaming laptop instead.

The newer RTX 20, 30, and 40 series GPUs are great cards that provide big performance boosts for games that take advantage of their newer features such as hardware ray tracing (in fact, the “R” in “RTX” stands for ray tracing) and DLSS. If you intend to play recent games at medium to high settings with high FPS, then by all means go for these modern cards.

But when it comes to AI, VRAM is king, and newer GPUs by Nvidia have been very lacking in the VRAM department. Basically, the AI boom caused Nvidia to become extremely stingy with the amount of VRAM that they put in their consumer products because they don’t want to cannibalize sales of their much more profitable enterprise products designed for corporate clients and data centers.

Case in point, the mobile GTX 1070 and RTX 4070 were mid-range cards released in 2016 and 2023 respectively. Yet, they both have the same amount of VRAM – 8 GB. If you want more VRAM in a portable form factor, Nvidia forces you to splurge on the much more expensive mobile RTX 4080 or 4090. And until AMD can prove itself to be a viable competitor in the mobile space, this situation is unlikely to improve.

Storage is another reason why these older laptops are great for AI tasks. As laptops switched from having both a hard drive and a solid state drive (SSD) to having just an SSD, their storage capacities actually decreased. Most gaming laptops from 2016 to 2019 should be fitted with both an SSD and a mechanical hard drive. The idea was that you would have your operating system and commonly used applications installed on the SSD while things like your photos, videos, music, and documents that don’t benefit from the speed advantages of the SSD can be stored on the hard drive where storage is slower but cheap and plentiful.

For generative AI, any applications you run will likewise be limited by the GPU rather than the speed of the hard drive so the performance benefits of the SSD would be wasted. Furthermore, you will most likely need to keep large amounts of datasets and media files on hand which will benefit from large amounts of cheap storage space where speed does not matter.

And finally, if you’re wondering whether the new laptops marketed with “AI” branding are any better for generative AI then unfortunately, they’re not. Compared to GPUs, the neural processing units (NPUs) inside these laptops are still fairly useless for generative AI tasks. A decent NPU might achieve 40 trillion operations per second (TOPS) while a dedicated GPU can achieve hundreds of TOPS to over 1,200 TOPS depending on the model. Instead, these NPUs are designed to assist in office tasks that might benefit from a light sprinkling of AI such as face detection and reducing background noise during video and voice calls while consuming very little energy in the process and thus reserving the GPU for more demanding tasks.

Depending on where you live, you should be able to score a gaming laptop from 2016 to 2019 for about $250-500 in the second-hand market. Try to aim for at least 16 GB RAM with a 256 GB SSD and a 1 TB mechanical hard drive. The CPU is unlikely to be a bottleneck because gaming laptops are unlikely to be fitted with crappy Celerons and pre-Ryzen AMD APUs. That being said, try to aim for an Intel Core i5 or i7. When it comes to graphics cards, some excellent choices are the 6 GB version of the GTX 1060 (avoid the 3 GB version), the 8 GB GTX 1070, and the 8 GB GTX 1080. AMD GPUs aren’t recommended due of their lack of CUDA support.