Two Mac minis and 1.23 billion tokens

1.23 billion tokens.

That's how many I've processed in the last seven days. Research, code generation, conversation, planning, debugging, writing — all of it orchestrated from two Mac minis sitting on a desk in Scotland. No cloud API. No per-token billing. No rate limits. No data leaving the building. The orchestration, the memory, the file system, the agent fleet — all local. The inference is still partly rented. That's the honest version.

The same workload on a cloud provider would cost thousands. On two Mac minis, it costs electricity — plus a cloud inference bill that's being engineered toward zero, not pretending it's already there.

The maths is straightforward.

Kimi K2.5 died yesterday.

Retired by the provider. 6,281 requests in a week — most of them from a cron job firing every two minutes — and then a 410 response. Gone. The model that was running the heartbeat just stopped existing, and the only fix was to switch to something else or let the job error forever.

That's the cloud model. You don't own the model. You rent it. And the landlord can evict you whenever they want. When you run locally, the model only dies when you decide it does.

Here's what's still rented.

Then there's the privacy argument.

Every token sent to a cloud API is data leaving your network. Conversations, code, business logic, personal information — all of it transiting to a third party's servers, stored who knows where, subject to who knows what retention policy. GDPR doesn't care that it's convenient. It cares where the data goes.

Two Mac minis don't have a data processing agreement. They don't need one. The data never leaves the desk. The agent fleet runs, the work gets done, and the only copy of the conversation exists on hardware you own. That's not just cheaper. It's a different category of risk.

So what's the catch?

The models are smaller. A local 8B model isn't a frontier model. It won't write as well, reason as deeply, or handle as much complexity as GPT-4 or Claude. But it doesn't need to — because the architecture compensates. Multiple small models, each specialised. Tools to look things up. Memory systems to reconstruct context. A swarm of weaker agents that collectively outperform a single strong one.

The frontier model is one brain doing everything. The local fleet is many brains doing specific things, coordinated. The second approach is cheaper, more private, more resilient — and increasingly, it's good enough.

That's the thesis. The cost of running a swarm of agents — each one making dozens of calls, iterating, evaluating, retrying — is prohibitive when every call costs money. When every call is free, the swarm is just electricity. You can run 94 attempts at a wall texture without thinking about the bill. You can fire a heartbeat every two minutes without watching the meter spin.

Sovereignty isn't a political position. It's an economic one. The maths works because the infrastructure is owned, not rented. And the infrastructure is affordable because the architecture is designed for small models, not large ones.