From “which model should we use?” to “what must we own to keep our future ours?” — why nations and enterprises are converging on the same answer.

On 1 July 2026, Palantir‘s Alex Karp went on CNBC and said what most enterprise leaders think privately but rarely say on air: on the token-consumption model that OpenAI and Anthropic popularized, “something has gone completely wrong.” The day before, Palantir had published a nine-point manifesto on “AI sovereignty” and paired it with a deal to run NVIDIA’s open Nemotron models inside classified, air-gapped US government environments the customer controls.

Part of this is a competitor’s pitch dressed as principle — Palantir sells into customer-controlled environments, so “own your whole stack” is commercially convenient. But strip away the self-interest, and it lands on a fault line running through every boardroom and every capital in the world. The question of the decade is no longer how good the model is. It is who controls the intelligence your institution runs on — and therefore who controls your future. That is why sovereignty is not an IT-architecture preference. For a growing list of countries and companies, it is about survival.

The Provocation: A War on How AI Makes Money

Palantir’s nine points are a direct attack on the frontier labs’ business logic. Reframed as imperatives, they read like a strategy memo:

  1. Your sovereignty dictates your future. It is the precondition for choice; relinquish it and others choose for you.
  2. Data is your treasure. Transferring it hands over both your current edge and the raw material for the next one.
  3. “Tokenmaxxing” corrodes your value orientation. Optimizing for token usage rewards disposable scripts and a false sense of progress over durable software.
  4. Controlling your weight is controlling your fate. Fine-tuned weights are distilled institutional memory; cede them and your alpha migrates to someone else.
  5. Sovereignty and alpha are not in tension. The right architecture lets you own and compound your tribal knowledge.
  6. Don’t politicize technical sovereignty. Techno-politicization manufactures false sovereignty and quietly surrenders real agency.
  7. Real expertise is existential. Let politics pick your technology, and you reward persuasion over correctness.
  8. Learn from institutions that deliver under pressure — those with skin in the game, not those with the best slides.
  9. Trust track records. A history of being right is the only reliable signal of future correctness.

The economic argument beneath the philosophy is what resonates with every CFO. As each model generation costs more than the last, enterprises are discovering that high token consumption is not high value creation. More prompts, more copilots, more tokens burned — none of it shows up as revenue created or cost eliminated, while proprietary data leaks out one API call at a time. The market agreed with the diagnosis if not the messenger: Palantir rose 8% that day, and the enterprise pivot from tokenmaxxing to ROI is already pushing buyers toward controllable open-weight models — several of them Chinese.

Reframing Sovereignty: Not Autarky, But Control of What Compounds

Here is where I part from the maximalists. Sovereignty is not autarky. No serious country or company will build the entire stack — silicon, compute, models, data, applications, talent — alone. Those claiming to are mostly engaged in sovereignty theatre.

Sovereignty is better understood as control over the layers of the stack that determine your future optionality. Some you can own; most you cannot. The strategic art is knowing the difference.

The layers you will rent — and that is fine:

  • Advanced silicon. Almost no one truly owns this. NVIDIA holds around 85% of AI GPUs, and access is US-governed. The move is to negotiate your access tier, not to pretend you own the fab.
  • Compute and data centers. Ownable with capital and power, but the chips inside are still imported. Here, your real assets are energy and land.
  • Foundation models. Frontier training is out of reach for most, though open-weight models are narrowing the gap fast. Rent the frontier; own a fine-tuned derivative.

The layers you can — and must — own:

  • Proprietary data. Your highest-leverage sovereign asset. It compounds over time, so never transfer it casually.
  • Fine-tuned weights. Encoded institutional memory. Own the weights that carry your IP, and control where they run.
  • Orchestration, context, and governance. Where durable enterprise advantage actually accrues. Build a model-agnostic one with workflows and human-in-the-loop.
  • Talent. The apex, and the scarcest input of all — the one moat money alone cannot buy.

The reconciliation of Palantir’s argument with pragmatism sits here: concede the commodity layers at the bottom (rent frontier models, run on someone else’s chips) and fortify the compounding layers at the top (data, weights, context, talent). Foundation models are becoming interchangeable; your data and workflows are not.

Why is this survival? At two levels. For nations, where intelligence becomes the primary factor of production, importing 100% of it means renting your future from a few foreign labs and one chip vendor — economic survival dressed as industrial policy, sharpened by the fear of a foreign “kill switch.” For firms, your moat was never the model; it is the accumulating stack of proprietary data, fine-tuned weights, and governed workflows. Rent that wholesale, and you are paying to train someone else’s moat while eroding your own.

The Five-Region Map, at a Glance

Every major economy now treats sovereign AI as a budget line, but they play very different games with very different hands.

  • United States — the one economy that owns the full stack, spending toward roughly $700bn in 2026 capex, with Project Stargate alone targeting $500bn, led by OpenAI, Anthropic, Nvidia, and Palantir. Its risk is not dependence but a demand-and-financing bubble concentrated on two loss-making labs and a single chip foundry.
  • China — cut off from top chips, it is winning the distribution race by giving open-weight models away. Alibaba’s Qwen drew 153.6m downloads in February alone, even as training stays partly Nvidia-tethered and Huawei targets 70% chip self-sufficiency by 2028. Champions: Alibaba, DeepSeek, Zhipu, and Huawei.
  • Europe — the world’s most developed rulebook (the AI Act, NIS2), sitting on barely 2% of US compute spend. Its champions, Mistral and Aleph Alpha, still rent GPUs from the hyperscalers they aim to escape. It can regulate itself out of the race.
  • Middle East — the boldest bet, funded by sovereign wealth. The Gulf deployed 43% of all global sovereign capital in 2025 through G42, HUMAIN, Qai, and Falcon. But it is “sovereign” only with US permission — and, since February’s missile strikes, freshly aware of physical sovereignty.
  • India — the most capital-efficient programme ($1.25bn India AI Mission, with roughly $250bn pledged), winning on language and frugality through Sarvam, Krutrim, and BharatGen, and holding Tier-1 unrestricted chip access. The catch: it buys every GPU rather than making one.

Five Trends That Cut Across Every Region

  1. Compute is the new chokepoint — and the world is fracturing into chip-access tiers. Which chips you can buy, how many, in which country, is now a matter of export licenses and bilateral diplomacy. Sovereignty increasingly means negotiating your tier, not owning the fab.
  2. The sovereignty paradox is universal. Nearly every “sovereign” stack on earth still rests on a US chip and often a US cloud. China trains partly on Nvidia; Europe rents from the hyperscalers it wants to escape; the Gulf builds only with US sign-off; India buys every GPU. Silicon-layer sovereignty is, for now, a fiction almost everywhere.
  3. Hedging beats purity. The mature strategy is not independence but managed interdependence — spreading bets across the American, European and Chinese stacks. The UAE investing simultaneously in Stargate and Mistral, and building a model on China’s Qwen with US hardware, is the template.
  4. Open weights are the great equalizer. Open-weight models — Chinese and European — now come within a whisker of the frontier at a fraction of the cost, and can be self-hosted, air-gapped and fine-tuned. This is what makes sovereignty at the data-and-weights layer newly affordable.
  5. Energy is the binding constraint, and regulation is bifurcating. Power, not chips, will cap the buildout by 2028; the Gulf’s cheap energy is a genuine sovereign asset. The world is splitting into a build-first pole (US, Gulf, India) and a rules-first pole (Europe), with China as a controlled third path.

The Non-Obvious Insights

  • Tokenmaxxing and national sovereignty are the same argument at two scales. A CFO worried that token spend yields no ROI and a minister worried about a foreign kill-switch are both asking: am I renting capability, or building an asset that compounds? The answer is identical — own the layer that accumulates.
  • Distribution is the real battleground, not benchmarks. China is losing the frontier race by roughly nine points and winning the substrate race by giving its models away. Whoever owns the default open-weight base owns the next decade of derivative development. Most commentary watches the wrong scoreboard.
  • Sovereignty and speed are in genuine tension — and honesty about it is the differentiator. Europe proves you can regulate yourself out of the compute race. The winners protect what compounds without strangling adoption.
  • “Sovereign-washing” is the coming risk. As the label sells, expect hyperscalers to rebrand dependence as “sovereign cloud.” Test for effective control — where data lives, who can compel access, whether you can exit — not marketing.

The Imperatives

For countries:

  1. Pick your archetype honestly. Full-stack owner, state-coordinated champion, or trusted regulated hub — most are the second or third, and pretending otherwise wastes capital.
  2. Secure compute and power. Energy is the constraint chips cannot fix; treat it as strategic infrastructure.
  3. Own data, language, and culture. These are the most defensible sovereign assets; import them, and you import someone else’s cultural defaults.
  4. Hedge your suppliers deliberately. Diversify across the US, European, and Chinese stacks; single-vendor sovereignty is an oxymoron.
  5. Regulate for adoption, not just restriction. Couple governance with genuine buildout; rules-first has a ceiling.
  6. Invest in talent above all. It is the one input that capital cannot import at scale.

For companies:

  1. Separate the commodity from the sovereign layer. Rent frontier models; own your data, fine-tuned weights, context and orchestration.
  2. Measure ROI, not tokens. Track revenue created, cost eliminated, and cycle-time reduced — never consumption for its own sake.
  3. Architect for model-optionality. Build a model-agnostic so you can switch labs as prices and capabilities shift.
  4. Own the weights that carry your IP. Where fine-tuning encodes proprietary knowledge, control those weights and where they run.
  5. Go hybrid by design. Managed frontier models for general work; sensitive workloads self-hosted or air-gapped.
  6. Govern from day one. As AI turns operational, orchestration, human approval, and auditability become the moat.

Where the Sovereignty Narrative Is Oversold

Balance matters because the maximalist case is partly a sales pitch. Most organizations do not need to own foundation-model weights — a governed, model-agnostic architecture over rented frontier models delivers most of the benefit at a fraction of the cost. Open-source improves so fast that single-vendor lock-in is a diminishing risk. And pushed too far, sovereignty becomes protectionism: duplicated capex, fragmented markets, slower diffusion of the very technology it claims to protect. The honest position is not own everything — it is own the layers that compound, rent the rest, and never confuse consumption with capability.

The Bottom Line

Palantir is directionally right about the trend that matters: the durable advantage of the next decade will not come from access to Claude, GPT, or Gemini — those are becoming interchangeable commodities. It will come from owning your enterprise context: proprietary data, fine-tuned weights, workflows, governance, and accumulated institutional knowledge. The same logic scales to the nation-state, where compute is becoming the successor to oil, and importing all of one’s intelligence means renting one’s future.

For both, the conversation has moved from “which model should we use?” to “how do we build an AI capability that preserves our sovereignty while staying model-agnostic?” The institutions that answer well will compound their advantage. Those that keep burning tokens and shipping their data offshore will wake up to find their alpha has quietly migrated to someone else’s balance sheet.

That is why sovereignty is not an architectural debate. It is about survival — and survival, in the age of intelligence, means owning what compounds.

Sources drawn on include CNBC, Reuters, the Middle East Institute, IISS, CSIS, CNAS, the Atlantic Council, the European Commission, OpenAI, G42, NVIDIA, PwC and Gartner, current as of July 2026. Enterprise examples are illustrative and anonymized at the sector level.