Why I Want to Own My AI

Chapter 01

The first step is small

I want to build an AI.

Not another wrapper around somebody else’s API, drifting somewhere in a hyperscale data centre I’ll never see. I want the machine on my land. I want to deploy it, manage it, measure it, train it, power it increasingly from energy produced here, and understand enough of the stack that I can repair and improve it.

Chapter 01

One machine, deliberately owned

The first step is small: one computer. In my case, it’s a single NVIDIA DGX Spark — a 150 mm desktop box built around the Grace Blackwell superchip, with 20 Arm cores, a Blackwell GPU and 128 GB of unified LPDDR5x memory. It draws less than 240 watts even under load, yet delivers up to a petaFLOP of FP4 compute: datacentre-class capability condensed into something that fits on a shelf.

Owning that first Spark is not about status. It’s about testing a claim: can a household progressively acquire sovereign computational capability in the same way we talk about acquiring energy, food, water and fabrication capability?

Chapter 01

Cloud is powerful. It is also centralised.

Cloud AI is astonishing. It also centralises power. The largest models live behind rate limits, dashboards and SDKs, on hardware controlled by multinational platforms. When I use those systems, I rent intelligence for a few milliseconds at a time. I don’t know where my data sits, I can’t see the power draw, and I can’t do anything with the waste heat.

A Spark is not a frontier training cluster. It will never replace an H100 pod. But it is enough: enough to host serious language models locally, run private RAG over the Solystopia archive, and orchestrate agents that read documents, write code and interact with the other systems on this property.

Chapter 01

Computation has a physical metabolism

Owning my AI means treating computation as part of the physical metabolism of a place. Sparks, storage servers and switches draw power from the same microgrid as the fridge and the heat pump. Heavy workloads can be scheduled when the panels are producing and the batteries are full. Waste heat stops being an annoyance to vent; it becomes a low-grade resource to experiment with, just like compost or greywater.

Chapter 01

Computational subsidiarity

Most importantly, it means shifting the architecture of intelligence. Instead of everything going up to the cloud, I want personal AI on devices, household AI in the Hearthmind cluster, neighbourhood and town compute where several households share capacity, and regional or national compute for the things that genuinely require scale.

Tasks should only move upward when the lower layer cannot perform them competently. That’s computational subsidiarity, and it belongs in Solystopia as clearly as energy subsidiarity or water subsidiarity.

Chapter 01

Where it ends

The question isn’t whether hyperscale AI should exist. It does, and it will. The question is whether ordinary people should also be capable of owning meaningful portions of the computational stack themselves.

Hearthmind is my attempt at an answer: one Spark on private British land, gradually growing into a sustainably powered AI datacontainer measured, documented and shared from the first machine onward.

The first step is small: one computer. Where it ends is considerably more interesting.