Is a Whole Mini PC Really Cheaper Than the RAM Alone? Checking the 2026 Mini PC Market

2026年8月27日

This page contains advertising (affiliate links). See our Privacy Policy for details.

14 August 2026

I was looking up memory prices when something struck me as odd.

Buying a stick of RAM on its own, 64GB now runs past $600. Meanwhile a mini PC with a similar amount of memory already installed was selling for close to that. A whole machine for about the price of the memory alone.

This reversal was being talked about through August 2026. I looked into whether it is real, and why it happens.

This article is a summary of published information, not a record of things I bought and compared myself. Prices move, so I have noted the date I checked (14 August 2026).

Contents

1. What exactly has flipped
2. Why it happens
3. Is this any use if you want to run AI locally?
- 3.1. There is one thing you have to check
- 3.2. What one machine here can show
4. What to look at, if you are buying
5. In summary: this looks usable
6. Sources

What exactly has flipped

Laid out side by side, it looks like this.

Memory on its own, versus a machine with memory in it

DDR5 32GB (memory only)	about $375	$80–120 in mid-2025
DDR5 64GB (memory only)	over $600	$150–200 in 2025
Mini PC: Ryzen 7, 16GB, 512GB SSD (complete machine)	about $424	CPU, SSD, power supply and case included

The capacities differ, so this is not a like-for-like comparison. But there really are cases where buying just the memory costs more than a whole machine.

On the face of it this makes no sense. A complete machine contains a CPU, an SSD and a power supply on top of the memory. All of that landing at roughly the price of the memory alone does not add up.

Why it happens

The reason turned out to be straightforward: the two were bought at different times.

Two different ways a price gets set

Buying memory on its own

You pay today’s market price, directly.

If the market rises, the price rises the next day

Buying a machine with memory in it

Manufacturers contract months or years ahead to secure their memory supply.

The memory inside machines on shelves today was, in many cases, bought under contracts signed before the rise

In other words, the finished machines in shops right now are carrying pre-increase stock. That is the whole of the reversal.

Which means this is not permanent. Once the pre-increase stock runs out, finished machines should rise too. It is more natural to read this as an unusual window than as a new normal.

Is this any use if you want to run AI locally?

This was the part I actually cared about. Cheap is established — but does it help for running models?

There is one thing you have to check

“Mini PC" covers two quite different things inside. Get this wrong and a cheap purchase will not reach what you wanted.

	How the memory is used	What that means for local AI
An ordinary mini PC	Memory for the CPU. The graphics side can only use a small slice	Large models will not load. Running on the CPU alone is too slow to be usable
A unified-memory mini PC	CPU and graphics share the whole pool	Most of the installed memory can hold the model

Unless the machine has unified memory (one pool shared by the CPU and the GPU), it will not get you there. Machines built on AMD’s Ryzen AI Max+ 395 (codename Strix Halo), and Apple’s Macs, are the ones that qualify.

Most cheap mini PCs are the first kind. Something advertised as “32GB for $250" is a fine buy for office work or watching video, but not for holding a large model.

What one machine here can show

I use one machine of the second kind. It shares 128GB between the CPU and graphics, and models from the 70B class up to the 235B class do run on it.

Assembling the same capacity out of graphics cards would mean lining up five or more 24GB boards. The price, the space and the electricity are not comparable. As a way past the capacity wall, this is still a realistic option.

GMKtec EVO-X2 (Ryzen AI Max+ 395 / 128GB / 2TB)128GB unified memory mini PC

Amazon

As an Amazon Associate we earn from qualifying purchases.

Speed is another matter, and here it does not match a graphics card. The rate at which it can move data to and from memory is different. “Large things run" and “things run fast" need to be held apart.

What to look at, if you are buying

Unified memory or not. This is the first fork in the road. You can tell from the chip name (Ryzen AI Max+ 395, Apple M-series, and so on)
How much memory. 32GB and 64GB versions exist. For the 70B class you want 128GB
You cannot add more later. The memory is soldered to the board, so whatever you buy is what you have. This one deserves particular care
Same chip, much the same speed. Several vendors ship Ryzen AI Max+ 395 machines, but AI speed is set by the chip and the memory, so they land close together. The differences are in cooling, expansion, networking and price
Today’s prices may be temporary. When the pre-increase stock is gone, finished machines should follow the market up

In summary: this looks usable

Bought on its own, DDR5 runs about $375 for 32GB and over $600 for 64GB. A finished machine with memory in it can be had for close to that — whole
The reason is when the parts were bought. Manufacturers contract well ahead, so machines on shelves today often carry memory priced before the rise
It is not permanent. Once that stock is used up, finished machines should rise as well
If you are choosing for local AI, check whether it has unified memory. Miss this and a cheap machine still will not hold a large model

Having looked into it, my conclusion is that this is usable. For someone starting out who wants to get hold of a 128GB-class machine, now is if anything an easier time to buy than usual. Set against lining up several graphics cards, it is more reasonable on price, on power and on where you put it.

From the position of already owning the hardware, there is no immediate reason to buy more. But “memory is expensive, so local AI is out of reach" does not seem to hold at the moment.