Prosumer.one / Manifesto / Rev 2026.10.06
I'm not a sysadmin, an ML engineer or a homelab hobbyist. I want serious AI capability that I can own, move, grow and actually use — without accidentally accepting a second career in infrastructure.
Not a demographic. A relationship to the machine.
More than AI as an app. Less than AI infrastructure as a profession.
Local models, private documents, agents, images, code, long jobs. Enough compute for work that matters.
Turn it on. Understand what fits. Choose a model. Get a result. No initiation ceremony required.
My machine, my data and my keys — with the cloud available when it is genuinely the better tool.
The menu problem
Fair question. Unfortunately, I don't know. If I could predict exactly what I'll do with AI three years from now, hardware procurement would be the least interesting use of that gift.
Today: a local model. Tomorrow: 4,000 PDFs. Next month: images. Then an agent. Then video. Then a new model appears on Tuesday and ruins the whole plan.
A little room around today's needs: enough openness, memory, expansion and software compatibility to ask “what if?” and try the answer before the curiosity evaporates.
Specs are ingredients. I am asking about dinner.
| On the box | In prosumer | |
|---|---|---|
| 120 TOPS | → | Which real workloads benefit? Which models run locally? What changes for me? |
| 96 GB memory | → | Which model classes fit comfortably, which fit with compromises, and what remains for my normal desktop? |
| PCIe / OCuLink / TB | → | Which eGPU or AI box works, and where does the connection become the bottleneck? |
| “Runs 120B” | → | Show the exact model, quantisation, context, speed, power and software stack. “Runs” is not a user experience. |
| Benchmark score | → | Show me a document, an image, an agent or a coding task I recognise. A washing-machine motor has horsepower too. I still want to know whether it washes my trousers. |
Not necessarily battery-powered. Definitely not married to a room.
Count the box, power brick, cables, dock and adapter — not just the pretty chassis.
A machine may fit in a bag and still be ridiculous on a library table. Noise, heat and cables are features too.
Sometimes the most portable compute stays at home. I should be able to reach it from a laptop or tablet without a networking apprenticeship.
Optionality wins
Private drafts, long-running jobs, offline work, predictable availability, experiments, and models nobody has wrapped into a SaaS yet.
Sometimes the biggest model is exactly what I need. Sometimes renting twenty minutes of someone else's GPU is smarter than buying one.
True cost accounting
Then come memory, storage, docks, the power supply that was somehow separate, electricity, subscriptions, shipping and adapters for the adapters.
A £900 machine that eats twelve hours of troubleshooting can be more expensive than a £1,500 machine that works.
What is the probability I will still be using this thing six months from now? A magnificent workstation I stopped switching on produces exactly zero useful results per second.
Prosumer evidence language
A number we actually measured. Setup, method and date stated.
It happened in real use. Useful evidence even without lab instruments.
The manufacturer says so. Important — and still their claim.
A reasonable conclusion. Architecture or other evidence, not a direct test.
Plausible, interesting, not trusted yet. Put the question mark where people can see it.
We tried it; it didn't work as expected. Failure belongs in the data, not the small print.
One philosophy, three surfaces
Manifesto, interesting devices, brands, Field Guides and human translations of hardware.
Current configurations, prices, availability, evidence, repeatable tests and ranking. The moving picture.
A dated snapshot of the same evidence and formula — market state, winners and awards. No separate award maths.
The prosumer test: does this machine help me do more things — or give me more things to maintain?
If it expands my capabilities, it belongs here. If it mainly expands my maintenance burden, it may still be an excellent computer. It is just the wrong one.
Same language, now applied to actual boxes
The rest of the network records the answers.
A no-checkout window of machines worth knowing about. Interesting is enough; endorsement is not required.
Claims, evidence status, capability hypotheses, prices and field results that can change as the market changes.
A dated annual freeze of the same evidence and formula — no special award maths invented after the shortlist.