What the shape says
Height is what you feel, and money buys surprisingly little of it. A £1,700 refurbished workstation with two mid-range cards sits about as high as boxes costing three times more. Past roughly £2,000, extra money buys capacity rather than speed, until you reach the professional tier at ten times the price.
Ball size is the bill that never stops. Idle watts, because a home server is idle 90 to 95 per cent of the time. The Apple and unified-memory machines are small dots. A dual-3090 tower is a boulder: 130 W at idle is roughly £285 a year before you ask it a single question.
Depth is what you can hold. Memory a model can actually use, video memory plus system memory, because sparse mixture-of-experts models spill into system RAM by design. This is the axis that moves cheaply, with memory sticks rather than a new machine.
Three things that surprised me
Context fills the card, not the model. A 27B model at 4-bit weighs 17 to 20 GB, which fits a 24 GB card with room to spare. Load a useful context and owners measure it settling at 28 to 30 GB. So 24 GB holds the model and then falls over on exactly the long-prompt work you wanted it for. 32 GB is the floor, not 24.
The biggest models are often the fastest. A sparse model activates a fraction of itself per token. Qwen3.8-Flash-Next is 125B total but only 6B active, and on one machine it finished a reasoning prompt 2.6 times quicker than a dense 27B, despite being nearly five times larger. The same effect shows up on GPT-OSS 120B: owners report 17 tokens a second on a single RTX 3090 with 128 GB of system RAM, and 21 on a 4070 Ti with only 32 GB, streaming the rest off an NVMe drive.
Tuning beats silicon more often than it should. Two owners with the same pair of cards measured 17 and 60 tokens a second. The difference was PCIe lanes on the motherboard. Read this chart for shape and ordering, never for precision.
Every machine, flat
| Machine | Price | Memory GB | Best t/s | Idle W | £/yr |
|---|
Where the numbers come from, and what to distrust
The speeds are owner reports. Gathered from public forum threads in September 2026, each with the hardware, quantisation and context stated by the person who ran it. Click any machine to see them with their caveats. They are not controlled benchmarks and nobody should treat them as such.
The prices are mixed. The GMKtec figure came from the vendor's own product record on 7 September 2026. The rest are UK street prices read the same day, and hardware prices move fast, memory especially.
Most of the idle-watt figures are estimates from the class of hardware rather than measurements at the wall, and the detail card says so for each machine. Idle power decides the running cost of an always-on box, so it deserves a plug-in meter and better data than this.
This is a snapshot. Model names in this field change monthly and will date faster than any of the hardware conclusions.