Compare
The honest comparison.
Everyone in this category publishes the number that flatters them. Here is the whole picture, including the row where we lose, because you will find out eventually and it may as well be from us.
What nobody tells you
On a 128 GB machine, token speed is basically a tie. Prompt processing is not.
38.6 vs 34.1
Tokens per second, 120B model
DGX Spark against a Strix Halo machine. A 13% difference, you would struggle to feel it. This is the number the whole category advertises.
1,723 vs 340
Prompt processing, tokens per second
The same two machines. Five times the difference, and this is the number that decides how long you wait after pasting a 60-page contract.
So the honest summary: if you generate text, almost any 128 GB machine will do. If you push very long documents through it all day, NVIDIA's silicon is meaningfully faster, and we will build yours on it. We are not tied to a chipmaker.
Side by side
Every row, including the ones we lose
| Specification | cato8 ProfessionalAMD build | NVIDIA DGX Spark | Strix Halo mini PC | Mac mini M4 Pro | Gaming PC |
|---|---|---|---|---|---|
| Unified memory | 128 GB | 128 GB | 128 GB | 24–64 GB | 24 GB VRAM |
| Generation, 120B model near-parity | ≈ 34 tok/s | 38.6 tok/s | 34.1 tok/s | Won't fit | Won't fit |
| Prompt processing the real gap | ≈ 340 tok/s | 1,723 tok/s | 340 tok/s | Mid | High |
| Indicative price, excl. VAT | £2,899 | $4,699 | ≈ $2,000 | from ≈ $1,400 | ≈ $2,000 |
| Models pre-installed | Yes | No | No | No | No |
| Works with no internet | Yes | After setup | After setup | After setup | After setup |
| Setup required | None | Substantial | Substantial | Substantial | Substantial |
| Someone accountable | Yes | No | No | No | No |
These figures are targets, the machine is pre-launch and the standard build uses the same silicon class as the Strix Halo column, so expect comparable throughput. The DGX Spark column is also what a cato8 NVIDIA build is based on. We will publish measured numbers on shipping units, not estimates.
Growing into it
The comparison that matters most is against next year
Every machine on this page links to another. Two units pool their memory, so a model that fits on neither will run across both. It means the honest answer to "what if we outgrow it" is "add another one", rather than "buy the expensive one now, just in case".
One Professional
128 GB. Models to about 120B. Comfortable for one person or a small team.
Two, pooled
256 GB as a single space. Opens up the largest open models available.
Two, independent
Twice the people served, each at full speed. Usually what a growing team needs.
The part other people leave out: pooling memory does not pool speed. Split one model across two machines and the link between them carries a fraction of what internal memory does, so the pair answers at roughly the speed of one. You gain capacity and headroom, not a faster single answer. If anyone quotes you double the throughput, they are selling.
Straight answer
When you shouldn't buy from us
There are four situations where something else is the better purchase. We would rather say so now than have you find out after it arrives.
Buy a Mac mini instead
You want one machine for everything, work, AI, everyday use, and 32B-class models are enough. It is cheaper, silent, beautiful, and Apple will support it for years.
Choose an NVIDIA build
You are an AI developer, you need CUDA specifically, or you process long documents constantly. NVIDIA prompt processing is about 5× the AMD build. We build on NVIDIA Grace Blackwell (the DGX Spark chip) and RTX PRO Blackwell too, with the same OS and installation, priced on request.
Build it yourself instead
You enjoy this. You already run Linux, you know what quantisation is, and a weekend of configuration sounds like a good weekend. You'll spend less and learn more.
Stay on cloud AI instead
Your data is not sensitive, you want the newest frontier model the week it ships, and you would rather not own an asset. Cloud is genuinely better at both of those.
So what are you paying for
Not the silicon. The fact that it works.
The chips in this category are broadly comparable and we are not going to pretend otherwise. What you are buying is a machine that arrives with the models already on it, an operating system built for running them, and, if you want it, an engineer who sets the whole thing up in your office.
A Strix Halo mini PC and a cato8 Professional contain similar silicon. One is a box of parts and a browser full of tutorials. The other is a working private AI on the day it arrives. That difference is the product.
Sources
- Token generation and prompt processing, DGX Spark vs Strix Halo, Compute Market and AIMultiple
- Apple Silicon throughput and memory bandwidth, LLMCheck and Compute Market
- DGX Spark pricing, NVIDIA Developer Forums
- Strix Halo mini PC pricing, TerminalBytes
Third-party figures, gathered August 2026. Benchmarks vary with model, quantisation and thermals, treat them as class indicators, not guarantees.