Is Decentralisation Good or Bad?
Liberation from the chokepoint, or the end of governance?
The Question
AI is moving from hyperscaler data centres to own racks, then to laptops, phones and devices. Cheaper, more private, more sovereign, but invisible to measurement and governance. Liberation or loss of control?
Optimist
The Optimist's Case
Decentralisation is what maturing technology does. Evidence Mainframe to PC, data centre to cloud, now cloud to edge. Self-hosting converts a per-token bill to a controlled capacity cost. Data and prompts stay inside your own walls.
On-device: marginal cost goes to zero, latency goes to nothing. Sovereignty pressure: regulators want to know where models run. On-premise or on-device is the cleanest answer. Lock-in falls away.
Summary: cheaper, faster, more private, more sovereign, less captive.
Sceptic
The Sceptic's Case
The optimist is right about the benefits. The sceptic points at the cost: the chokepoint.
Centralised AI: one place to measure, govern, switch off. Decentralised AI: none. Interpretation The consumption-measurement stack assumes a meter someone else runs. Open-weight on a laptop: no bill, no log, no record.
Shadow IT was painful. Shadow AI is worse. The estate you govern is the estate you can see. That gap widens quarterly.
The central API is where safety filters, rate limits, audit logs, and incident response live. Every control must be rebuilt per device, per team, per model.
EU AI Act assumes an identifiable provider and deployer. A thousand fine-tuned local models strain that assumption. Darkest version: decentralisation dissolves control, including the safety case.
Synthesis
House viewWhat Would Settle It
Watch the tooling. If consumption-measurement extends to self-hosted and edge, the crisis is solvable. If tooling stays meter-dependent while workloads leave the meter, the crisis is real.
Evidence Early signals are mixed. At least one observability tool is entering maintenance mode.
Watch governance experiments: do device-first organisations hold evaluation coverage, audit trails, and incident rates comparable to centralised peers?
If We Get It Right / If We Get It Wrong
Right + prepared: Workloads run wherever they are cheapest, fastest, most sovereign. Governance travels with them.
Right + unprepared: The meter goes dark. Dashboards empty. You lose visibility exactly when the estate scales.
Wrong + centralised anyway: Higher cost, deeper lock-in. You pay the sovereignty and privacy penalty for staying visible.
Wrong + scattered anyway: Ungovernable estate, unmeasurable value. The worst of both worlds.
The Author's Honest Position
Decentralisation is not good or bad. It is coming. The question is which disciplines survive it.
Interpretation Consumption-based measurement does not survive. Outcome-based value management does. Build outcome telemetry now while the meter still works.
Treat every workload's governance as something it carries, not something it inherits from infrastructure.