The Day AI Became a Strategic Asset
A U.S. order to take two of Anthropic’s most powerful models offline wasn’t really about Anthropic. It was the moment frontier AI started behaving like a controlled export — and that quietly rewrites the questions every leader should be asking.
I was in a hotel above Athens, reworking the notes for a keynote to a Fortune 150 leadership team, when the story broke. The irony arranged itself without my help. Athens is the city we credit with democracy, philosophy, and the free movement of ideas — a whole civilization built on the premise that thinking should travel. And there, on my screen, was a headline about ideas being locked down.
The U.S. government had ordered Anthropic to pull two of its most advanced AI models offline.
At first it read like every other dispatch from the AI beat. Another launch, another benchmark, another flare-up in a news cycle that has trained all of us to keep scrolling. The industry manufactures milestones so relentlessly that even the genuinely historic now arrives pre-discounted.
This one didn’t scroll past.
The longer I sat with it, the clearer it became that Anthropic wasn’t the story. Anthropic was the occasion. The move to pull the model for certain individuals was the precedent.
Here is what happened. On Friday, June 12, 2026, U.S. officials ordered Anthropic to immediately cut off access to its two most capable models — Claude Fable 5 and Claude Mythos 5 — on national-security grounds. The trigger, according to reporting from NBC News, Bloomberg, and TechCrunch, was the discovery that the safety guardrails on the newer model could be “jailbroken.” Because the Commerce Department directive barred access for any foreign national, Anthropic said the only way to comply was to disable the models outright. The company pushed back hard, calling the response disproportionate and warning that, applied across the industry, the same standard could freeze new model releases for every frontier lab. The order drew sharp criticism from across the tech and policy world — including, pointedly, from authors of the administration’s own AI strategy.
The details are still moving, and some will surely change. But the precedent doesn’t depend on the details.
For the first time, a frontier AI model was treated not as a commercial product but as a matter of national security.
Whether that judgment proves right or wrong almost doesn’t matter. The category shifted. Intelligence took a step toward becoming a strategic export — something governments meter the way they meter advanced semiconductors, encryption, and aerospace.
For everyone building on top of AI — executives, investors, founders, policymakers — that is a different conversation than the one we’ve been having. For three years we’ve asked how AI will change work. The newer, sharper question is who gets access to advanced intelligence in the first place.
From Product to Infrastructure
Every transformative technology travels the same arc: from novelty, to necessity, to something a state feels compelled to guard.
Electricity arrived in the late nineteenth century as a marvel and became, within a few decades, the substrate of modern life — which is why no government today leaves the grid unwatched. The internet followed the same path, mutating from a research curiosity into critical infrastructure for commerce, defense, medicine, and finance.
Artificial intelligence is running that arc at a sprint.
A few years ago, large language models were lab experiments. Today they write production code, accelerate scientific discovery, run customer operations, and sit inside strategic decisions at the highest levels of companies and governments. McKinsey estimates generative AI could add between $2.6 trillion and $4.4 trillion to the global economy every year — a figure, on its own, comparable to the entire annual output of a G7 nation.
As the capability curve steepened, so did the unease. Policymakers have stopped seeing frontier systems as mere productivity software and started seeing them as capability multipliers. The model that helps an engineer find a bug can help an adversary find a vulnerability. The system that speeds legitimate biology can, in the wrong hands, speed the dangerous kind. That dual-use quality is exactly what pulls a technology out of the commercial world and into the security one.
And here AI breaks the old playbook. Governments know how to control physical things — machine tools, centrifuges, chips. You can stop a shipment at a port. But a model is software. It deploys globally in seconds and can be reached from anywhere with a connection. The result is a head-on collision between twentieth-century export controls and a twenty-first-century technology that ignores borders by design.
Why the Order Matters More Than Anthropic
Strip away the specifics and one durable thing remains: a new line item on the corporate risk register.
Call it geopolitical model risk — the chance that access to a foundational AI capability is curtailed not by a price change or an outage, but by government policy.
Boards have learned to reason about cybersecurity risk, operational risk, regulatory risk, reputational risk. Almost none have a framework for what happens when a model they depend on is declared strategically sensitive overnight. Picture building core workflows on a single frontier model, training thousands of employees on it, wiring it into products and P&Ls — and then waking up to find it restricted. Eighteen months ago that was a thought experiment. Last Friday it became a case study.
Geopolitical model risk may be the most underpriced risk in enterprise AI today.
The Semiconductor Precedent
If you want to see where this goes, look at chips.
There is no modern AI without advanced semiconductors, and governments grasped the strategic weight of those chips years ago. Beginning in October 2022, the United States imposed sweeping export controls on advanced computing chips and chipmaking equipment, explicitly to slow rivals’ access to the compute that powers AI. The logic was blunt: control the chips and you control access to intelligence.
Now follow that logic one step further. If chips are strategic because they enable AI, then the most capable AI models are strategic in their own right. The model becomes as consequential as the hardware that trained it.
That is a profound break with how software has always worked. For a generation, software companies assumed their products could cross borders frictionlessly. The makers of frontier AI are discovering that the assumption no longer holds at the top of the capability curve — and the implications run well past any single company. OpenAI, Google DeepMind, xAI, Meta, and every lab chasing the frontier now operate inside the same reality. The question was never whether governments would get involved. It’s how far.
An AI Cold War, Not a Space Race
It’s become fashionable to call this an AI space race. I think the better analogy is the early stage of a technological cold war.
History is, in part, a record of technologies redrawing the map of power. Naval power built empires. Industrialization relocated the world’s economic gravity. Nuclear capability rewrote diplomacy itself. AI belongs in that lineage. Nations increasingly read leadership in AI as a proxy for economic competitiveness, military readiness, scientific edge, and national resilience — which is precisely why AI is being folded into national strategy.
That creates a structural tension that won’t resolve quickly. Technology advances through openness, collaboration, and speed. Security advances through control, secrecy, and advantage. Those instincts pull in opposite directions, and the friction we’re watching now is likely the opening scene of a much longer drama.
What Enterprise Leaders Should Do Now
The takeaway for executives isn’t to panic. It’s to grow up.
Too many organizations still run AI as a technology project — something to implement and check off. AI has become a strategic capability, and strategic capabilities demand governance, not just deployment. Projects optimize for rollout. Capabilities require resilience.
Practically, that means putting questions on the executive agenda that weren’t there in 2024:
- What happens to the business if access to our primary AI provider changes overnight?
- Do we have a real multi-model strategy, or a single point of failure?
- How many critical processes depend on one vendor’s model?
- Which regulatory or export-control moves could reshape our AI roadmap?
- Who, specifically, owns monitoring the geopolitics of AI access inside our company?
None of these were boardroom questions eighteen months ago. All of them belong on the agenda now.
Intelligence as Infrastructure
The most quietly radical part of this story is what it forces us to admit about what AI actually is. We got comfortable calling it software. It behaves more like infrastructure — and infrastructure shapes economies, concentrates power, confers advantage, and, inevitably, attracts the state.
Whether Anthropic’s models return unchanged is almost beside the point. The event is a signal flare: the rules governing frontier AI are being rewritten in real time, and access to intelligence itself is becoming a strategic variable. A few years ago that idea would have sounded exotic. Today it reads as almost obvious.
The next chapter of AI won’t be written by capability alone. It will be written at the intersection of technology, governance, national security, and global competition — and the organizations that thrive will be the ones fluent in all four.
That, in the end, was the thought I carried back into the keynote in Athens. The future of AI is no longer only about building smarter models. It is about who controls access to them.
Susan Sly is a global AI expert, entrepreneur, and keynote speaker who helps organizations navigate the future of artificial intelligence, leadership, and innovation. Inquire about Susan speaking at your next conference or executive event.
