Last updated on July 2nd, 2026 at 04:15 am

Minerals, compute, and model access are collapsing into a single trade instrument. The limits on it are already forming, at both ends.

Aaron Levie, the CEO of Box, recently described what the next geopolitical phase of AI policy could look like in a CNBC segment with Deirdre Bosa. Picture AI access used as a lever in trade negotiations, he said, something that could threaten a country’s economic and national security, something that may end up treated like a weapon. He was narrating the deployment layer from inside it, something we have discussed in our AI sovereignty series. His company runs on models he does not own, and he was naming the dependency as it tightened. Four weeks earlier the dependency had stopped being hypothetical.

The Directive and the Deals

On June 12, the US Commerce Department imposed export controls on Anthropic’s most advanced models, Mythos (and the public facing iteration with prompt classification and routing called Fable 5) citing national-security concerns tied to their cybersecurity capabilities. Anthropic disabled access worldwide within days. (July 1 update: these restrictions have now been lifted, without significant changes to the model, with much the same arbitrariness as they were imposed). European businesses, researchers, and government agencies that had built these models into their workflows found the connection severed with no transition period. Commerce then permitted a limited restoration of Mythos. The mechanism was demonstrated in full: one capital decides who gets to run frontier intelligence, and it can decide overnight.
The same logic was already operating one layer down, at the substrate the models depend on. The New York Times documented that the federal government is steering critical-minerals deals worth roughly nine billion dollars toward at least fourteen companies with financial ties to the Trump and Lutnick families. Cantor Fitzgerald, the Lutnick family firm now overseen by Commerce Secretary Lutnick’s sons, appears across the list as underwriter, advisor, or capital agent. USA Rare Earth was approved for up to 1.3 billion dollars in Commerce loans and direct funding. Vulcan Elements received a 620 million dollar Pentagon loan, the largest of its kind, three months after a venture firm where Donald Trump Jr. is a named partner took an equity stake, with no public record of any conflict review. A tungsten project in Kazakhstan was signed days after the Trump sons and their partners invested in a related entity. Executive Order 14241 had already waived competitive bidding and independent technical verification for these projects, clearing the path.

The model layer and the mineral layer now carry the same property. Access is discretionary, and the discretion sits with a narrow set of actors in one jurisdiction.

One Lever, Every Layer

Levie named a single layer. The directive is the model layer. The minerals deals are the provision necessary for materials that fabricate the compute that runs the models. These are not three policies that happen to rhyme. They are one posture applied at three altitudes: convert each layer of the AI stack into something that can be switched off, gated, or charged for at the gatekeeper’s discretion.

Monetization is the connective tissue. The minerals that feed the chips are being routed through politically aligned intermediaries who profit from sitting in the supply path. The chips concentrate inside US borders. The models that run on the chips are export-gated. Each layer is acquiring a tollbooth, and the tolls answer to the same address.

This is where Canada’s position moved into focus. Minerals are the one layer of the stack where Canada holds an asset rather than a dependency. That advantage means something only if it stays sovereign. The same monetization logic now consolidating the US mineral supply is available to absorb Canadian capacity into it, and a critical-minerals advantage sold into someone else’s consolidation stops being leverage the moment the sale closes.

Austria Reaches for the Exit

Austria became the first advanced economy to hit the lever and look for a way out. In a letter to European Commission Executive Vice President Henna Virkkunen, released by the Austrian government, State Secretary for Digitalization Alexander Pröll proposed that the EU explore “the strategic establishment and participation of Anthropic within the European Union,” offering “legal certainty, market access, capital and a set of values that suits this company.” The pitch ran on shared identity: a company that says it treats safety as conviction, an attitude Pröll called “deeply European.” A government holding compute or guaranteed supply would negotiate on those. Austria holds neither, so it reaches for the one layer that requires no infrastructure. The values pitch is what sovereignty looks like when a country has nothing foundational to bargain with.
a firm that in Europe would be unleashed rather than constrained.

A frontier model is welded to hundreds of thousands of the newest chips, to gigawatts of power, to data centers contracted to American hyperscalers years in advance. Anthropic is spending tens of billions building that base inside the United States. The pitch ran on shared identity because shared identity was the only lever Austria could pull. “A company that treats safety as conviction” is a characterization Anthropic supplies about itself, and Pröll adopts it wholesale as the basis for the offer. A logo crosses an ocean easily. The power plants do not move. Austria aimed at the model layer, because the model is what got gated and so the model is what felt like the object of sovereignty. The binding constraint sits two layers down, in compute siting and the guaranteed supply underneath it, where Austria has no instrument at all.

Compute Supremacy Is a Depreciating Asset

The widely shared reading of Austria’s failure, argues Andrew Curran, is that Anthropic stays where the compute and the guaranteed supply are, that it cannot risk being cut off, and that compute will only concentrate further within US borders. The exit fails, on this account, because the United States holds permanent leverage at the substrate.

That reading treats compute scarcity as destiny, and China is the standing refutation.

DeepSeek trained its R1 model for roughly six million dollars against the hundred-million-dollar range for a GPT-4-class system, using a mixture-of-experts design that activates a fraction of its parameters on any given query. Its V4 model, released in February 2026, is optimized for Huawei’s Ascend chips, the direct product of being denied Nvidia’s best hardware. Chinese models now run at roughly a quarter to a sixth of the cost of comparable American systems. In October 2025, China passed the United States in monthly open-source model downloads. The efficiency techniques born under constraint travel quickly across the ecosystem, and each time they spread they lower the strategic value of raw compute supremacy.
Being cut off did not end China’s program. It produced the workaround. The coercive posture is manufacturing exactly the route-around-Washington outcome it was built to prevent. Read through the present moment, Austria’s exit looks doomed. Read through the China record, it looks like the clumsy first instance of a move that becomes viable the moment the incentive to leave grows strong enough to fund the efficiency. Curran is modeling a temporary advantage as a permanent one.

The Courts Are the Next Venue

The same heavy-handedness is heading toward a domestic limit, and a live legal debate is already forming around it. The question is whether the creation, distribution, and use of frontier AI counts as protected expression under the First Amendment. Dean Ball has put it directly, arguing that the courts will decide these issues and that because the underlying technology is a language model, expression should count for a great deal.

That specific route runs into a wall. The right to invent and deploy a frontier system is a separate thing from the protected act of using its output, and invention already has its constitutional home in the IP clause, which empowers Congress to secure limited rights for inventors. The means of production is not the content. The instrument Ball reaches for is the wrong one.
The structural prediction underneath his argument holds anyway. Regulation this aggressive gets litigated, and the courts are the venue where the next phase of this fight plays out. External coercion is running toward an internal test it may not pass.

Where Canada’s Leverage Sits

The lever looks total. It is bracketed at both ends. Projected outward, it rewards the efficiency that erodes it, with China demonstrating the mechanism in real time. Turned inward, it runs toward a constitutional limit on regulating expression. The leverage that appears absolute is fenced on two sides at once.

The layer that decides real standing is the foundational infrastructure, or substrate. Compute concentration is the lock today, and compute concentration runs on minerals that are being captured and monetized through a narrow set of politically connected firms. Canada holds real mineral leverage, in its variety and concentration. Whether that leverage stays sovereign or dissolves into the US supply consolidation is a decision, not a forecast.