Following the successful military operation to capture Venezuelan President Nicolás Maduro, Trump has renewed threats to annex Greenland, with the White House stating that “utilizing the U.S. Military is always an option” to acquire the Danish territory. Danish Prime Minister Mette Frederiksen warned that a US attack on Greenland would mean “the end of trans-Atlantic” cooperation and NATO, whilst European leaders including Macron, Starmer, Meloni, and Scholz issued a coordinated and somewhat weak statement affirming that “Greenland belongs to its people”.
The geopolitics are spectacular. The technology implications are also significant.
The Trust Collapse
Geopolitical reliability now trumps (ha!) technical capability. When AI infrastructure can be arbitrarily weaponized, rational actors must diversify regardless of performance cost. Chinese open models suddenly offer what US vendors fundamentally cannot: immunity to American political volatility.
Last year we saw “Liberation Day”, with the unilateral imposition of tariffs. If Trump threatens model access restrictions - and arbitrary trade actions are now a well documented precedent - every business building its AI based on US Frontier Models (primarily ChatGPT, Gemini and Claude) carries existential counterparty risk. The combination of “Liberation Day” and threatening Greenland significantly increases that switch off risk. The question shifts from “which model performs best?” to “which model won’t be switched off?”
Qwen, Kimi, Z and rivals win not through superiority but downloadability. You can deploy them locally and operate indefinitely without external dependencies. US API models, even if technically superior, remain vulnerable to presidential whim.
The Rest of the World’s Calculation
Denmark’s military intelligence service has classified the US as a security risk for the first time. With NATO cohesion fracturing, European governments face acute vulnerability: critical infrastructure - healthcare, finance, defense logistics - increasingly dependent on large language models potentially controlled by a hostile power.
The rest of the world won’t necessarily be classifying the US as a security risk (although Trump’s list of possible targets is long). However, exactly the same calculus applies to them: can we trust the US? Whatever level that was at a year ago, there is a lot less trust now.
This is no longer theoretical risk management. Once developers invest in open model infrastructure (fine-tuning pipelines, deployment systems, integration frameworks), switching costs create lock-in. Many companies are only at the beginning of their AI/LLM journey. They might need a rapid course change to account for this new risk.
Open Weights as Sovereignty
The architectural difference proves decisive. Broadly, Chinese labs release open-weights models - download once, it’s yours permanently. US labs favor API access or controlled releases maintaining ongoing leverage.
That business model worked when US geopolitical stability was assumed. It becomes fatal when that assumption breaks. You can’t remotely disable a locally-deployed model. You absolutely can terminate API access, revoke licenses, impose retrospective restrictions.
China wins by offering digital sovereignty: “Download this, we can’t take it back.”
The One-Way Ratchet
Trump doesn’t need to actually restrict model access. Any game theorist will tell you a credible threat suffices. Risk-averse CEOs outside the US must plan for worst-case scenarios. Even 20% probability of arbitrary restrictions over five years demands redundancy: build a system based on open models now whilst options remain available.
The best truly open source, open weight providers are currently Chinese, but other providers might emerge. Even from the US. The leading US frontier models might even recognise the issue and think about their own offerings, although that seems unlikely given how locked in to their business model they are.
That open infrastructure won’t disappear when political weather improves. Investment is sunk, integration complete, open models will keep improving. If it transpires it will be a permanent market loss through self-inflicted reputational damage making superior products uninsurable.
The Counter-Arguments
So far, so doom laden. There are, of course, counter-arguments: you can’t trust China either; its compute that matters; the Greenland talk is bluster; the EU will never break with the US; open models are of very poor quality. Let’s examine these in turn:
“China will do the same thing - why trust them?” Valid point. But open models can be deployed locally with no ongoing Chinese control. China is currently incentivized to build trust during market-share expansion. China could poison open models with backdoors, although auditable code and active security research communities limit this risk significantly.
“Compute dominance matters more than model access” US retains chip export controls and cloud infrastructure advantage. Even if Chinese models gain adoption, they’re limited to inference on existing architectures, dependent on legacy hardware (Nvidia H100 embargo bites eventually), and vulnerable to US infrastructure choke-points if cloud-deployed.
True for frontier capability. But most use cases don’t require frontier. If 80% of applications run fine on open-weights models deployable on available hardware, the US loses the market that matters commercially. The enterprise software industry isn’t built on cutting-edge research - it’s built on reliable, deployable solutions. And China are pressing very hard to develop their own advanced chip manufacturing capability…
“This overstates Trump’s bluster” Historically, Trump threats don’t always materialize (TACO - Trump Always Chickens Out). Markets may ignore bluster.
This is irrelevant. Risk management doesn’t wait for proof. The risk is there now. One high-profile incident - even accidental or bureaucratic - validates the concern permanently. Prudent CEOs will act on the possibility, not probability. As discussed, the mere threat changes behavior irrevocably.
“EU/allies won’t actually break with US” Decades of intelligence-sharing, interoperability, and institutional inertia argue against rapid decoupling.
Agreed for peacetime. But if US unreliability is becoming a pattern: Liberation Day tariffs, arbitrary policy reversals, threats against allies. Allies and partners will increasinglyt seek hedges. Even short of military conflict, sovereign capability becomes non-negotiable when your primary technology supplier demonstrates contempt for alliance norms.
“Open models are poor quality.” The argument is that recent open model gains are the result of cherry-picked benchmarks, gaming evaluation metrics, distillation from US models (ironic if true), or hitting capability ceilings without access to cutting-edge compute.
Maybe. But if the quality gap is “good enough for most uses” rather than “frontier,” the argument still holds. Most businesses aren’t doing cutting-edge AI research - they’re automating customer service, analyzing documents, and generating routine content.
The Irony
The irony is exquisite. Trump frames Greenland acquisition as a national security imperative - protecting American interests against Chinese and Russian encroachment. Yet the threats themselves accelerate the very outcome they claim to prevent: Chinese dominance in the technology infrastructure that actually matters.
Leaving aside any thoughts of the rule of law, the rights of the inhabitants, the evils of colonialisation and so forth, Greenland offers the US mineral resources and strategic geography. AI models offer control of the global digital economy. One might eventually be worth hundreds of billions. The other is already worth trillions and growing exponentially.
While Trump tweets about territorial expansion, Xi Jinping watches European bosses increasingly download Chinese models. Only one of these leaders appears to truly think strategically.