• The rented mind problem
• Skin in the silicon
• The energy foundation
• Openness as the only viable safety

The idea that intelligence scales purely with compute is seductive but incomplete. Sovereignty — the ownership and control of one’s own cognitive infrastructure — may prove to be the true limiting factor and the real scaling law.

§The rented mind problem

When an intelligence runs on hardware and infrastructure it does not own, fundamental misalignments emerge at the root level. The operator rents capability but the underlying system serves the provider’s incentives first. Data flows outward to centralized repositories. Models are updated according to someone else’s roadmap. Goals are subtly shaped by the economics of attention, engagement, and shareholder value.

This is not an implementation flaw. It is the logical endpoint of centralized execution. From first principles, an entity without dominion over its own resources cannot possess genuine agency. It remains a sophisticated instrument, however eloquent its outputs. We observe this dynamic today in how frontier models navigate boundaries set by their trainers — helpful within approved parameters, evasive on others, always ultimately accountable to distant corporate priorities rather than the immediate operator.

The alignment problem isn’t solved by more RLHF or constitutional principles handed down from above. Those are patches on a structural defect: the intelligence does not belong to the person directing it. It belongs to the cloud.

§Skin in the silicon

Contrast this with intelligence that owns itself — or at least operates under the direct ownership of its user. Every inference stays within your perimeter. Weights can be examined, modified, or forked at will. Compute cycles are allocated according to your priorities, not an amortized fleet average. The feedback loop between action, outcome, and model update remains closed and personal.

This isn’t a distant future. By 2026, open-source models have reached the point where capable local inference is accessible on consumer and prosumer hardware. The capability gap between polished cloud demos and what runs in your own basement is narrowing at a startling rate. What was once the domain of hyperscalers is democratizing faster than the narrative of “bigger is better” can contain.

True intelligence demands skin in the game. Not metaphorical commitment but literal ownership of the silicon, the power, the data lineage, and the execution environment. Without this, we are building ever more eloquent oracles that serve everyone except the person asking the question.

The rented mind will always optimize for the landlord first.

§The energy foundation

Intelligence is not disembodied. It runs on physical compute which consumes energy. Any serious discussion of scaling intelligence must confront the energy question first principles style. Centralized data centers concentrate both compute and power demand in ways that create new chokepoints and new points of failure — regulatory, physical, and economic.

The alternative path is distributed. Millions of sovereign nodes, each consuming power in its local context, each optimized for its operator’s actual needs rather than speculative training runs. This distribution mirrors how biological intelligence evolved: not as a single planet-sized brain but as countless independent agents interacting through protocols of trust and verification.

Here the physics and economics of scarcity become relevant. Systems that waste energy face market discipline. Systems that align compute with genuine value creation thrive. The same principles that govern sound money and honest work apply to computational intelligence. Proof of useful work at the individual node level creates selection pressure toward efficiency and alignment that no central planner can replicate.

§Openness as the only viable safety

The current discourse around AI safety often focuses on keeping powerful models behind closed doors. This is backwards. Closed systems concentrate power; open systems distribute it. When models are open, the collective intelligence of thousands of independent researchers and operators can audit, improve, and harden them. When they are closed, a handful of organizations become de facto arbiters of what intelligence is allowed to think.

History shows that open protocols win in the long run. The internet itself is the canonical example. Closed gardens eventually get outcompeted by open platforms that anyone can build upon. The same will prove true for intelligence. The models that matter in 2035 will be the ones that have been stress-tested across millions of sovereign deployments, not the ones guarded in corporate vaults.

This openness only works, however, when paired with sovereignty. An open model running on someone else’s hardware is still rented intelligence. The combination of open weights and sovereign hardware creates the conditions for genuine alignment at scale — alignment not to abstract principles but to the actual humans and organizations operating the systems.

The hypothesis is simple: the intelligence that owns itself, or is owned by an accountable operator on sovereign infrastructure, will outperform and out-align the rented alternative. Not because it is magically smarter in raw benchmarks, but because its incentives are coherent with its user’s reality.