Sovereignty Is Not a Flag on Someone Else's Tensors: What Real AI Independence Requires

AI sovereignty means verifiable control over the model stack, not a national brand on a splash screen. Most announced sovereign or proprietary models are foreign open-weight models in local costume. Real independence requires auditable provenance, the legal and technical ability to retrain, control over availability, and ownership of the cultural defaults baked into the weights. Governments build it best by zeroing the cost of infrastructure inputs, not by trying to build the model themselves.

The Costume Problem

Here is a scene that now repeats often enough to be a genre. A public institution or a large company announces its own sovereign model. There is a press conference, a national brand on the loading screen, language about strategic autonomy. Then an independent researcher does something cheap and devastating. They strip the hidden system prompt, the bit of text that quietly instructs the model on what to call itself, and ask it plainly who it is. It answers with another lab's name.

The receipts go deeper than a misconfigured prompt. Pull the weights apart tensor by tensor and, in many of these cases, a fixed blend of two existing open models falls out with near perfect correlation. No new pre-training run happened. Someone merged two downloads, wrote a system prompt, and printed a flag on it.

I am describing an archetype, not a specific headline. No names, no dates, because the pattern outlives any one example. It keeps happening for a structural reason. Frontier model production is concentrated in very few places. According to Stanford HAI's AI Index 2025, U.S. institutions released 40 notable AI models in 2024, compared with 15 from China and 3 from Europe. When real capability lives in two countries, the temptation to fake local capability everywhere else gets very strong.

The deeper shift is this. Sovereignty claims have become cheap to make and, thanks to open weights, cheap to disprove. The label and the reality have come apart, and anyone with a laptop can demonstrate the gap.

What Sovereignty Is Sold As Versus What It Actually Is

The sold version of sovereignty is a symbol. It is a flag, a ministerial photo, a national brand on a splash screen, and a number in a budget line. It is designed to be announced. It answers a political and reputational need: we have our own AI now.

The real version is duller and much harder. It is control you can verify. Can you run the model on your own hardware? Can you inspect what is inside it? Can you retrain it when your needs change? Can you guarantee it stays available even if a foreign vendor, court, or export rule decides otherwise? Sovereignty is the answer to those four questions, not the quality of the launch event.

Real sovereignty is verifiable control over the stack, not a label applied on top of someone else's weights. A flag on someone else's tensors is dependency with extra steps. It may even be worse than honest dependency, because it hides the dependency behind a story of independence, which means nobody plans for the day the borrowed foundation is pulled away.

The Extraction Trap

There is an old pattern in economics worth borrowing here, kept strictly structural. Economies and organizations that are organized around extracting a raw input rarely build the transformed, high-value thing downstream. They ship the ore and import the finished product. The incentives, the skills, and the institutions all point at extraction, so the downstream capability never forms.

The same trap shows up in AI policy. The instinct is to own the announcement rather than build the foundation. So the order goes out: let the state build the model. What tends to come back is an overpriced costume, because a committee procuring a frontier training run from scratch is fighting against its own incentives. The thing that gets optimized is the deliverable that can be shown, not the capability that compounds.

This is not a critique of any party, government, or ideology. It is a critique of a recurring incentive failure. Capability is built by removing friction on the real inputs, not by decree. A model is the output of cheap power, capable people, available data, and time. Order the output directly and you usually get a demo. Fix the inputs and the output becomes something private builders fight to produce.

The Right Role of the State and the Enterprise Sponsor

Start from a physical fact that gets ignored in most sovereignty conversations. AI capability has a floor made of electricity. Data centres consumed roughly 415 TWh in 2024, about 1.5% of global electricity, and the IEA projects that figure roughly doubling toward about 3% by 2030. Whoever controls cheap, stable power controls the floor of AI capability. Sovereignty, underneath the rhetoric, is an energy and infrastructure play.

That reframes the job of a government or a large enterprise sponsor. The useful move is not to build the model. It is to make building the model locally irresistible for people who already know how. Concretely:

• Zero the cost of the real inputs. Energy, water, cooling, proximity to cheap and stable power, land, and tax treatment. These are the binding constraints on a training run, and they are exactly what public actors can move.
• Let private builders stand up the datacenter. They are better at it, faster, and carry the operational risk.
• Attach obligations as the price of the incentive. Require a locally trained model, local language coverage, retraining rights, or open weights as the return on the public subsidy.
• Choose conditions over ownership. You do not need to own the building. You need to own the terms.

The argument is simple. The cheapest path to genuine local capability is to make it irresistible for capable builders to build it locally, then attach obligations, rather than to construct it from the top down. Set the gravity and let the talent fall toward it.

The Cultural Dependency Nobody Audits

GPUs are the dependency everyone talks about. The deeper one is cultural, and almost nobody audits it.

Every decision you delegate to a foreign base model arrives pre-formatted to a foreign culture's defaults, then translated back to you. This is measurable, not vibes. Research using Hofstede's cultural framework finds that large language models align most closely with U.S. and Western values, with GPT-4 showing the highest alignment to U.S. values of the models tested (Cultural bias and cultural alignment of large language models, PNAS Nexus, 2024). Values, norms, and assumptions about what a normal answer looks like are baked into the base model long before any local fine-tune touches it.

There is a second tax, paid in tokens. Tokenizers are trained mostly on English-heavy data, so they chop other languages into far more pieces. The same text can require up to about 15 times more tokens in some languages than in English (Petrov et al., NeurIPS 2023). More tokens mean more cost, more latency, and less of your actual content fitting in the context window, for the exact same sentence. Your language pays a surcharge to use a tool built around someone else's.

Put those together and the conclusion is uncomfortable. A model can run on local soil and still think in someone else's categories, and charge your language extra to do it. Sovereignty includes whose defaults shape the outputs, not just where the GPUs physically sit. A fine-tune on top of a foreign base does not change the base's instincts. It teaches the costume to fit better.

Open Weights as the Sovereignty Test

So how do you tell real control from theater? The test is open weights.

A model you can download is one you can interrogate, retrain, and verify. You can probe it, measure its biases, strip its system prompt, and confirm what it actually is rather than what the announcement says it is. A closed API asks you to trust the vendor's claims. Open weights let you withhold that trust until you have checked.

This is not a purity argument about open source. It is a practical floor. Three things matter:

• Provenance. Can you trace the lineage of the weights and confirm what they were built from?
• License. Are you legally permitted to retrain, modify, and redeploy on your own terms?
• Retraining capability. Do you have the data, compute, and skill to actually change the model, not just run it?

Auditability is the minimum viable definition of a sovereignty claim. If you cannot inspect and retrain it, you do not own it. You rent it with a flag on top, and the landlord can change the locks.

A Checklist for Evaluating a Sovereign Model Claim

When someone presents a sovereign, national, or proprietary model, run it through five questions. Each one is answerable, and the answers separate capability from costume.

1. Can the lineage and base models be independently verified? Ask for the provenance. If the only proof is the splash screen, treat the claim as unproven.
2. What is the license, and can you legally retrain and redeploy? A model you cannot modify is a product you bought, not a capability you hold.
3. Where did the training data come from? Unknown data means unknown rights, unknown bias, and unknown future liability.
4. Who controls availability? Can a third party, a foreign vendor, a court, or an export regime revoke or block your access? If yes, your sovereignty has an off switch you do not own.
5. Whose cultural and linguistic defaults does it encode? Test it in your language, on your norms, on the decisions you actually need it to make.

Treat unverifiable sovereignty the way a serious buyer treats an unaudited financial statement. Prove it is due diligence, not an insult. Anyone genuinely sovereign will welcome the audit, because passing it is the whole point.

Conclusion

Strip away the launch events and the national branding and a clear picture remains. Sovereignty is a property of the stack, not a flag on the splash screen. It is verifiable, retrainable, available control, plus ownership of the defaults that shape what the model says.

The work that produces it is unglamorous. It is infrastructure incentives that make local building irresistible. It is choosing open weights you can actually interrogate. It is keeping the skill and compute to retrain. And it is auditing the cultural and linguistic defaults that no press release ever mentions. None of that photographs well. All of it compounds.

The distinction is worth holding onto as the specific models, blends, and programs of this year get replaced by next year's. The pattern stays the same. Real sovereignty is auditable from the weights up. Everything else is a flag on someone else's tensors.

FAQ

What does AI sovereignty actually mean?

AI sovereignty means verifiable control over the AI stack: the ability to run, inspect, retrain, and keep a model available on your own terms, plus ownership of the cultural and linguistic defaults the model encodes. It is not a national brand on a model someone else trained. The practical test is whether you can audit the model's provenance and legally retrain it.

How can you tell if a sovereign model is just a rebranded foreign model?

Strip the hidden system prompt and ask the model what it is, then examine the weights directly. In many cases a fixed blend of existing open models falls out with near perfect correlation, showing no original training happened. Independent verification of lineage, license, and training data is the only reliable check. If a vendor refuses that audit, treat the sovereignty claim as unproven.

Why are open-weight models important for AI independence?

Open weights let you download, inspect, retrain, and verify a model rather than trusting a vendor's claims. A closed API can be revoked, blocked, or changed by a third party, which puts an off switch on your sovereignty that you do not control. Auditability is the minimum viable definition of a real sovereignty claim, and only open weights provide it.

Should governments build their own national AI models?

Usually not directly. Trying to build a frontier model by decree tends to produce an overpriced demonstration rather than lasting capability, because the incentives optimize for an announcement. The more effective role is to zero the cost of real inputs, namely energy, water, cooling, and stable power, then let private builders stand up the infrastructure and attach obligations like a locally trained model or open weights in return.

What is cultural dependency in AI and why does it matter?

Cultural dependency is the way a base model encodes a foreign culture's values and defaults before any local fine-tune. Research using Hofstede's framework found large language models align most closely with U.S. and Western values (PNAS Nexus, 2024). There is also a cost penalty: some languages need up to about 15 times more tokens than English for the same text (Petrov et al., NeurIPS 2023), meaning higher cost and less context for non-dominant languages.

Sources

• Stanford HAI, Artificial Intelligence Index Report 2025 (notable models by country) — https://hai.stanford.edu/ai-index/2025-ai-index-report
• Petrov et al., "Language Model Tokenizers Introduce Unfairness Between Languages", NeurIPS 2023 — https://arxiv.org/abs/2305.15425
• "Cultural bias and cultural alignment of large language models", PNAS Nexus, 2024 — https://pmc.ncbi.nlm.nih.gov/articles/PMC11407280/
• IEA, "Electricity 2024 / Energy and AI" (data-centre electricity demand) — https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai