From Cloud to Laptop: The Decentralisation Arc
Speculative thesis, not a management recommendation. These pages extrapolate from early signals. Each separates what is already happening from what is projected, and says what would prove the thesis wrong.
Outcome-based value management is the approach most likely to survive decentralisation, because it measures the outcome, not the meter.
The Hidden Assumption Breaking
Almost every AI cost and value tool assumes AI runs in someone else's data centre. That assumption is breaking.
The Four-Stage Arc
Stage 1: Frontier Cloud APIs (Now)
Evidence OpenAI, Anthropic, Google. Pay per token. The meter is the API call. Cost measurement is straightforward: count the tokens, multiply by the rate.
Stage 2: Self-Hosted and Open-Weight (Emerging)
Evidence Llama, Mistral, self-hosted inference. The meter moves from API to infrastructure. Cost becomes capacity cost, not consumption cost. Attribution gets harder.
Stage 3: Edge and On-Device (Emerging to Future)
Interpretation Models run on laptops, phones, IoT devices. Marginal cost approaches zero. The meter disappears entirely. Consumption measurement becomes impossible.
Stage 4: Hybrid Everywhere (Future)
Speculation Workloads route dynamically between cloud, edge, and device based on cost, latency, privacy, and sovereignty. The meter is fragmented across infrastructure types.
Cost Structure Morphs, Meter Disappears
Follow the money across the arc. The basis of measurement dissolves.
- Stage 1: Per-token pricing. Clear meter.
- Stage 2: Capacity cost. Meter exists but attribution is harder.
- Stage 3: Sunk cost. No meter.
- Stage 4: Hybrid cost. Fragmented meter.
Interpretation The token meter is a phase-one artefact. Anchoring AI economics to per-token cost is building on a meter scheduled for removal.
Attribution and Governance Degrade as Chokepoint Disappears
The centralised API is a single chokepoint. Every request passes through it. That makes measurement, governance, and control straightforward.
Interpretation Decentralisation removes the chokepoint stage by stage. Into the gap rushes the shadow-AI economy.
The dark AI gap: the proportion of AI work you cannot even see. As decentralisation progresses, this gap widens.
Sovereignty Improves While Governance Fragments
Interpretation The arc pulls two things in opposite directions. You gain control over where data lives and where models run. You lose control over visibility and governance.
This is a genuinely uncomfortable trade-off worth naming.
The Two Readings of the Same Arc
Decentralisation as Liberation
Cheaper, faster, more private, more sovereign. Lock-in falls away. Marginal cost approaches zero. This is the optimist's reading.
Decentralisation as Loss of Control
Ungovernable estate, unmeasurable value, invisible risk. The meter goes dark exactly when AI becomes ubiquitous. This is the sceptic's reading.
Synthesis
Interpretation Both are correct. The resolution depends on your measurement choice.
The Durable Claim
Consumption-based measurement dies with the meter. Outcome-based value management survives because it never depended on the meter.
If you measure tokens, you lose visibility in Stage 3. If you measure outcomes, you hold visibility across all four stages.
If We Get It Right / If We Get It Wrong
Right: A single value lens holds steady while infrastructure churns. Governance travels with workloads. The estate remains visible.
Wrong: The measurement system goes dark exactly when AI becomes ubiquitous. Shadow AI becomes the majority of AI.
Practical Instruction
Invest now in outcome telemetry. Build the instrumentation while the meter still works. When the meter disappears, the outcome telemetry is all you have left.
Interpretation Firms that wait until Stage 3 to build outcome measurement will find themselves governing an invisible estate.
What would change this view
Already happening
- Open-weight models and on-device inference are available alongside cloud APIs.
- Enterprises already run some inference on their own infrastructure.
Projected
- A growing share of inference leaves metered cloud APIs, so cost and usage telemetry fragments.
- Outcome measurement becomes the only consistent view across cloud, self-hosted and device inference.
What would prove it wrong
- Inference stays concentrated in a few metered providers.
- Gateway or endpoint telemetry rebuilds a reliable meter cheaply across all deployment types.
Next review of this thesis: April 2027.
Related reading