Sovereignty Is All You Need

There is a single architectural primitive missing from how the industry is building AI, and its absence will determine the outcome more than any capability race between nations or firms. Sovereignty at the node. The rest is machinery.

Apex inference is critical infrastructure. The Ratchet depends on a frontier — capability has to come from somewhere, and the firms building it are doing necessary work. The question is what happens at the consumption layer, where the substrate of human judgment either compounds at the node or leaks outward by default.

Which is why the Wall Street Journal op-ed this week from two Andreessen Horowitz partners matters. Jai Ramaswamy and Matt Perault argue America should beat China in AI by releasing capable, American-friendly open-weight models, and releasing them fast.

They are right about the threat. Qwen’s model family crossed 700 million Hugging Face downloads in January. DeepSeek runs inside American labs, classrooms, and enterprise pilots at a scale that should alarm anyone paying attention.

They are right about half the cure. The other half — the architectural layer that determines whether open weights compound American capability or leak it — is missing from the op-ed entirely. Both halves are necessary. Ship one without the other and the geopolitical ratchet runs in the wrong direction.

For decades, the software industry has been rewarded for building what engineers quietly call dependency ratchets — design choices that raise switching costs over time. It is a fair trade in most domains. You can move your photos from Apple to Google; you can move your cloud workloads from AWS to Oracle. It hurts, but the substrate is recoverable. Your photos are still your photos.

Platforming cognition is a different animal. The ratchets we built instinctively — to create sticky products, defensible moats, the predictable revenue lines our investors reward — do something categorically different when the substrate is human judgment. What you give away does not come back.

The cleanest way to state the institutional problem: organizations whose mission depends on accumulated judgment cannot afford to build cognition on substrate they do not own or govern. That sentence is not ideology. It is operational risk.

Every default AI pipeline today routes telemetry, reasoning traces, approval patterns, and user judgment outward to a vendor. That vendor recycles the stream into the next model, the next moat, the next lock-in. The Grove Foundation has documented this as the telemetry trap: an open-weight model consumed through a centralized API produces the same dependency ratchet as a closed one. Open source solves half the problem. The consumption layer reproduces the other half automatically.

Human judgment is the seed corn. When a centralized pipeline harvests the patterns of discrimination, refinement, and approval that would otherwise accumulate as your own expertise, it is not a feature. It is extraction dressed up as productivity. You are paying a recurring fee to fund the automation of your own thinking. Every major AI consumer platform is structurally designed to do this. It is not a bug in any one vendor. It is the business model.

MIT’s 2025 research found that 95% of enterprise generative AI pilots produced zero measurable ROI. That is not a coincidence. It is what happens when the consumption pattern treats user judgment as exhaust instead of capital.

GRV-003, the Learner Autonomaton standard published last week by The Grove Foundation, supplies the missing layer. It mandates a non-bypassable five-stage pipeline — Telemetry, Recognition, Compilation, Approval, Execution — backed by three user-owned files and zone-based consent rules. Validated patterns ratchet inward: they demote to cheaper, more sovereign, more local compute tiers as mastery grows. The provenance arcs produced along the way turn how competence formed into inspectable, portable cognitive capital. The standard is published under CC BY 4.0. It is model-agnostic and domain-invariant. This familiar design pattern has been pressure tested in compliance-heavy regulated industries before broader application, and it imposes zero friction on the strategic open-weight push or even apex frontier model providers. It simply ensures the weights serve American minds and institutions rather than subsuming them.

The commercial traction is already visible where it matters most — and it is not on the coasts. In central Indiana, home to Eli Lilly, Anthem, Corteva, Cummins, Old National Bank, and precision manufacturing at scale, the architecture maps directly onto serious regulatory, audit, and liability requirements. These industries cannot tolerate opaque automation or untraceable decision chains. They have been pricing the cost of unauthorized behavior for a century — in FDA penalties, loan defaults, failed harvests, product recalls. They have pre-existing intuitions about sovereignty, audit trails, and circuit breakers that coastal firms are only beginning to develop vocabulary for. Architecture-first AI is a Midwestern instinct, sharpened by industries that have always required it.

Meanwhile, the externalities of the centralized hyperscale model are starting to bind. Roughly half of the US data centers planned for 2026 have been delayed or canceled — not for lack of capital, but because the grid cannot carry the load. Financing structures are tightening. States are writing laws that let them cut data centers off the grid in emergencies. None of this proves the apex tier is failing. It proves the apex tier was never as self-contained as its economics implied — and that institutions building cognition on top of it need an architectural layer that can survive the volatility. On the morning this Alert published, the White House invoked Section 303 of the Defense Production Act on grid infrastructure and supply chain capacity, finding that America’s “aging and constrained electric grid infrastructure poses an increasing threat to national defense.” Truman-era industrial mobilization authority has now been deployed against the same bottleneck the centralized AI buildout has spent two years failing to clear with private capital. That is the externality binding in real time.

The piece in the WSJ missed a substantial venture opportunity. Building the reverse-ratchet software layer — Autonomaton pipelines, routing engines, zone enforcers, provenance tooling, co-sign protocols — is a classic infrastructure play. It is productizable, monetizable, and defensible. Regulated industries, universities, defense contractors, and lifelong-learning platforms all need it. The capital currently chasing hyperscale inference farms has a higher-ROI alternative: distributed sovereignty infrastructure that turns the dependency dynamic 180 degrees. The market signal is not subtle — 95% of the $30–40 billion in annual enterprise AI spend is failing to find returns in the current consumption pattern. That is not a saturated market. It is a stranded one, waiting for architecture that makes the capability land. For a venture firm to write this op-ed without naming the opportunity is, at minimum, a curious omission.

That capital is already moving. New funds are forming in Indianapolis to back sovereignty-layer infrastructure and the regulated-industry applications built on top of it. The geography is not incidental: the same industrial base that makes Indiana the right proof ground makes it the right capital base. The center of gravity for the reverse ratchet is not on the coasts.

A serious national AI strategy is a two-tier strategy. The apex tier — frontier models, the firms financing them, the compute and weights that anchor capability — is genuine critical infrastructure and deserves to be treated as such. The sovereignty tier is what ensures the apex serves American institutions rather than the inverse. Ship the open weights aggressively. At the same time, fund the builders who productize the inward ratchet. Neither tier works without the other. Open weights supply raw capability. The GRV-003 Autonomaton ensures that capability compounds here — in American workers, in American institutions — rather than leaking abroad or into vendor moats that feed on the seed corn of the country that grew it.

Open weights are necessary. Sovereignty primitives are what make them sufficient. Ignore the reverse ratchet and we win the model-weight war while quietly losing the cognition one.

The Grove Foundation’s architecture proposals don’t replace regulatory frameworks. They simply give all players in the ecosystem a way to actually satisfy them more rigorously, at decreasing costs, and with compounding benefits.

Jim Calhoun is the founder of The Grove Foundation, an Indianapolis-based open standards body for AI governance architecture.