The Morning the Models Went Dark

A near-future scenario, and why European leaders should read it as a warning rather than fiction.

It's a Tuesday in late 2027. Your Head of Risk opens her laptop at 7:40 a.m. and the compliance agents that have monitored your transaction flow for the past eighteen months — the ones that quietly caught four AML anomalies last quarter before they became regulatory incidents — return a single line: "Model tier unavailable in your region. Fallback engaged."

Fallback means the previous generation. Six to nine months behind. Slower, blunter, and — crucially — now a full step behind whatever your American and Gulf-based competitors are still running, because their access never lapsed. Nobody hacked you. No contract was breached. A capability you had silently built your operations around simply stopped being something you were allowed to have at full strength.

You spend the morning on the phone with your provider. You are told, politely, that frontier access is now "subject to deployment prioritization." You are number 4,000-something on a list. You are, in the language of the new world, no longer trusted enough, big enough, or close enough.

This scenario is not science fiction. It is the logical endpoint of a shift that is already underway — one that a recent essay names with uncomfortable precision.

The assumption we never questioned

Almost every European AI strategy — corporate and national — rests on one quiet premise: that state-of-the-art American AI will always be broadly available. Buy the tokens, build the wrapper, capture the value. The frontier is a utility; we are all customers.

In "Cut Off", published on his Substack Threading the Needle, Anton Leicht argues this assumption is a fundamental error. Leicht is worth listening to: a fellow with the Carnegie Endowment's Technology and International Affairs Program who writes on the political economy of AI, with a background in German and EU policy and government. He spends his time thinking about exactly this — the geopolitics of frontier models and the fate of the "AI middle powers" caught between the US and China. When someone with that vantage point says the abundance assumption is broken, it's not San Francisco hype.

It's worth being precise about that vantage point, though, because it shapes the conclusions. Leicht writes from inside a leading American institution, and the analysis — for all its sympathy toward middle powers — ultimately reasons from US strategic interest outward. Europe appears in it much the way it tends to appear in Washington's AI discourse: as a recipient of access, a negotiating counterparty, a "middle power" whose best play is to bargain cleverly for a place in someone else's stack. That framing is honest and clarifying. It is also exactly the framing European leaders should refuse to accept as destiny.

His piece points to a turning point he calls the "Mythos moment" — a frontier cybersecurity model whose developer chose to release its capabilities not to the open market, but to a short, US-only list of trusted partners. A competitor's comparable model followed the same restricted path. The era of "good enough, available to everyone" — what Leicht calls the end of AI's "Warhol moment," where everyone got their fifteen minutes of frontier access — may be closing.

Leicht's thesis is that three forces are converging to make frontier AI scarce rather than abundant:

1. Security. The most capable models are the most dangerous ones. Cyber-offense, bio-risk, hostile distillation — these push developers (and governments) toward restricted, vetted, staged release rather than mass deployment.
2. Compute. Unlike software, serving a frontier model has a brutal marginal cost. Every additional institutional user competes for scarce chips. This is a zero-sum game, and it is getting worse, not better.
3. The American state. What begins as legitimate security caution ends, inevitably, in politics. Once Washington holds a hand on the tap, that control becomes strategic leverage — wielded in trade disputes, alliance management, and national-interest calculus that has nothing to do with your quarterly roadmap.

The conclusion is stark: a world split between nations and firms with frontier access and those without — a divide in wealth, security, and capability as consequential as the unequal spread of the industrial revolution, which, the essay reminds us, triggered mass migration, revived dormant conflicts, and destabilized democracies.

The scarcity case is right about the risk — and too quick about the remedy

Leicht is right about the direction of travel, and I think anyone running a European enterprise should take the diagnosis seriously. But I'd push back on how absolute the scarcity becomes — and on where that leaves us.

His picture, drawn from the policy vantage point, can read as near-deterministic: the tap narrows, Washington's hand tightens, and middle powers are left negotiating for scraps. I think that underweights the counter-forces already pushing the other way. Open-weight models keep landing within months of the frontier and are increasingly good enough for the overwhelming majority of real enterprise workloads. Efficiency curves keep collapsing the cost of yesterday's frontier capability — so the gap between "what's gated" and "what's freely runnable" matters less for most operations than the headline suggests. And the supply side is fragmenting, not consolidating: more labs, more sovereign compute initiatives, more credible non-US options every quarter. Scarcity at the bleeding edge is real. Scarcity of useful, deployable capability is far from settled.

To his credit, Leicht anticipates this objection and rejects it head-on. Efficiency curves, he argues, make last year's frontier cheap, but not next year's — and if competition forces you to run "not good enough AI, but the best AI," then cheaper trailing capability buys you nothing. That's the crux, and it's exactly where I disagree. The premise that every serious player needs the bleeding edge is true for a narrow band of use cases — frontier cyber-offense and -defense, where his Mythos example genuinely bites — and false for almost everything a real business actually runs. Reconciliation, claims, compliance monitoring, pricing, workforce planning: these do not need the single best model on earth. They need a reliable, auditable, controllable one, wired into the right process. The "you need the best or you lose" framing is the load-bearing assumption of the whole scarcity-is-destiny argument, and for the enterprise, it simply doesn't hold. Once you drop it, frontier gating stops being an existential threat to your operations and becomes what it actually is: a constraint on one capability tier, routable around by design.

So my disagreement isn't with the risk — it's with the implied helplessness, and specifically with what it implies for Europe. Read from Washington, the European move is to negotiate harder for access. Read from here, that's a trap: it accepts dependence as the starting condition and merely haggles over the terms. And here's where I think the policy lens misses the most actionable part of the story. Leicht's remedies are, by the nature of his work, policy remedies: build more data centers, strike compute-for-access deals, improve data-center security. All necessary. All slow. All dependent on governments — most of them not European — behaving well.

But the most powerful lever against frontier scarcity isn't policy at all, and it isn't a better seat at America's table. It's architecture. The question that actually determines whether the scenario above is fatal to your business is not "will Washington keep the tap open?" — it's "did you build something that can survive the tap narrowing?" That's an engineering decision, it sits entirely within European control, and it doesn't require anyone's permission. For Europe, sovereignty isn't a diplomatic concession to extract; it's a property you can design into your own systems, starting now. Which is the part I want to spend the rest of this post on.

Why this should keep European C-suites awake

If you run a European enterprise, here is the part that should sting. In the scarcity world, being a willing, paying customer is no longer enough. Leicht is blunt: neither most governments — unaccustomed to million-dollar AI subscriptions — nor European firms — structurally weaker at monetizing software — are positioned to win a purely economic bidding war for frontier capacity.

So the question stops being "Which model do we buy?" and becomes a governance question:

• Where does the capability physically run, and who can switch it off?
• What happens to a critical, AI-operated process the morning access is downgraded?
• How much of our operational core have we built on a dependency we do not control and cannot insure?

Most organizations have answered the first question implicitly — someone else's data center, under someone else's jurisdiction — and have never asked the other two at all.

The uncomfortable mirror for "AI strategy"

A great deal of what passes for enterprise AI strategy today is, in effect, a bet that the tap stays open at full pressure forever. Chatbots, copilots, thin product layers over a frontier API — useful, but architecturally fragile. They assume abundance.

The scarcity thesis inverts the strategic priorities:

• Resilience over novelty. An AI-operated process that cannot survive a capability downgrade is not an asset; it is a latent liability.
• Sovereignty as architecture, not slogan. Auditability, reversibility, and the ability to run within your own environment stop being compliance theater and become operational insurance.
• Governed autonomy over maximal automation. Knowing exactly what each agent does, why, and under whose control is what lets you swap the engine underneath without grinding to a halt.

This is not an abstraction. The morning the models go dark, the only things that keep your critical processes running are the ones you designed for that morning: human governance by design, sovereignty and control over your own data and infrastructure, and full auditability and reversibility. Not because they look reassuring on a slide — but because they are what lets you swap the engine underneath without the operation grinding to a halt. The enterprises that treated these as architecture rather than compliance theater will be the ones still operating. The ones that bought abundance and optimized for it will be number 4,000-something on a list.

What to do before the scenario becomes a memo

Leicht's essay is not fatalistic, and neither should we be. The scarcity future is not inevitable. Several of the remedies are mundane — more data centers, better data-center security, and compute-for-access deals between the US and its allies, where European partners trade favorable build conditions and energy for contractual guarantees of frontier access. These are policy moves, and Europe should pursue them with urgency rather than wait to be allocated a place in line.

But policy is slow, and your operations are not. For leaders responsible for a balance sheet now, three questions are worth putting on the next board agenda:

1. Map your frontier dependency. Which critical processes would degrade — and by how much — if you dropped to last-generation capability tomorrow? If you can't answer, that is the answer.
2. Insist on reversibility. Every AI-operated process should have a defined fallback that does not depend on the goodwill of a foreign provider or government.
3. Treat sovereignty as a capability, not a constraint. The firms that built for control will be the ones still operating when access becomes political.

The deeper lesson is that this was never really a new problem requiring clever new solutions. It is the old, unglamorous discipline of building infrastructure you can stand on, and a world that can absorb this technology without fracturing. The signal, as he puts it, is already here. Where I part company is on who the signal is for. Leicht writes it as a brief for policymakers, mostly American ones. I'd read it as a brief for European builders: the people who can decide, this quarter, whether their critical systems are designed to outlive a narrowing tap — without waiting for a deal in Washington to save them.

The naïve move is to assume abundance and optimize for it. The strategic move — for Europe especially, and for anyone building on this technology — is to assume the tap can close, and to make sure that when it does, you are still the one running your own processes.

Anton Leicht's essay "Cut Off" was published on his Substack, Threading the Needle. The piece is sharp, and worth reading in full.

If the questions above are live ones in your organization, let's talk — building governed, sovereign, reversible agentic systems is what we do.