In June, three days after Anthropic released what were then its most powerful Claude models, the US Commerce Department banned their export to foreign nationals. Not chips — model access. That single decision is the cleanest marker of what has happened to the AI industry by 2026: the choices that determine which tools a company can use, at what price and in which country, are now made in Washington, Brussels and Beijing rather than in boardrooms and research labs.
Over the past decade governments have moved to treating AI as critical national infrastructure, in the same category as energy, industry and defense. Three forces follow from that reclassification, and together they set the conditions every company now operates under: restrictions on international trade and technology exports, regulation that keeps changing, and a contest over technological sovereignty.
Export control is the bluntest of the three, and 2026 has shown how unstable it is as a planning input. The US Bureau of Industry and Security first relaxed the rules on exporting Nvidia's most powerful AI processors, then tightened them again a few months later. The stated aim is to deny Chinese AI developers the semiconductors they need to build their own infrastructure and to preserve the American lead.
The Claude decision extended that logic past hardware. Export control, historically an instrument for machine tools and semiconductors, now applies to software with the same ease. For companies, the practical consequence is the one nobody budgets for: a tool a business depends on can become unavailable overnight because of a decision that business had no part in. Any organization that has built a critical process on a specific model is carrying that risk whether or not it has priced it.
Regulation has become the second front. Over the past two years European Union regulators have fined American technology companies more than $7 billion, seeking to strengthen Europe's influence over how AI develops. Washington has answered with tariff threats and sanction campaigns aimed at internet disinformation, arguing that the European measures amount to censorship of online speech. What began as a compliance question has turned into an instrument of trade policy, used in both directions.
Underneath both sits the sovereignty push. States of every size increasingly want to control their own AI capability and infrastructure rather than depend on another country's. Stanford's 2026 AI Index puts a number on it: since 2018, the count of state AI supercomputing clusters in Europe and Central Asia has risen from three to 44. The same report found that more than half of recently adopted national AI strategies come from emerging economies, which points to AI capability being treated as a component of statehood and a precondition for international influence rather than a reward for having it.
The global agenda, though, is set by two countries. The United States and China account for the overwhelming share of frontier model development, compute and investment. When Washington imposes export restrictions and Beijing requires state data centers to use domestic chips, companies on every other continent absorb the result.
On raw capability the distance between them is now small. According to Stanford's AI Index, the gap between the best American and the best Chinese models is under 3%. Three years ago it reached 30% in the American direction. Over the same period, American AI companies raised 23 times more private investment than their Chinese competitors.
The two countries are also building different things. The largest closed models, GPT and Claude among them, are made in the United States by companies under US jurisdiction. China has concentrated on open models, where DeepSeek and Alibaba are among the most widely used options. Anyone who wants to run a model on their own hardware is increasingly choosing Chinese technology.
Put those numbers side by side and the American position looks weaker than the export-control posture implies. Twenty-three times the private capital has bought a lead of under three percentage points on a scoreboard that has compressed from 30 points in three years. That is not a return on investment; it is a treadmill. And the control regime compounds it: every restriction that makes American frontier models harder to obtain abroad pushes the buyers who need self-hosting toward the open Chinese stack, which is exactly the outcome the restrictions exist to prevent. The strategy assumes that controlling the most powerful models wins the race. China's assumption is that the winner is whoever controls the models everyone else chooses to run.
The measure that would settle which bet is right is not in either scoreboard. Model quality rankings say nothing about where models are actually deployed, and nobody in this dispute is publishing what share of the world's production systems runs on which country's weights. Benchmark leadership and installed base are different assets, and the two capitals are each optimizing for one of them while quoting the other's numbers.
For a business, none of this is influenceable, but all of it is plannable. The global AI ecosystem is dividing along national lines — different hardware ecosystems, competing models, incompatible regulatory regimes — which means the jurisdictions a company builds in, sells into and operates from now belong in its AI strategy rather than in its legal department. For international organizations that argues for abstraction layers: the ability to swap models, datasets and compliance mechanisms when the politics move. It also argues for running the exercise nobody wants to run, which is what happens to the business if an AI service it depends on is switched off without notice.
Geopolitical instability in AI is most likely a permanent condition rather than a passing disruption. The companies treating model choice as a procurement decision rather than an architectural one are the ones for whom a rule change in a capital they have no vote in becomes an outage they cannot fix.