jump to...

When the frontier can be switched off, plan for the floor, not the ceiling.

Everyone is benchmarking the ceiling.

The number that will actually run your business is the floor. And you don’t set it.

Governments and funding do.

For three years, one question seemed to be the only one worth asking. Which model is smartest? Every release got scored against the last. We argued about benchmarks the way we used to argue about clock speeds.

In June 2026, a quieter question took over.

On June 2, the White House issued an executive order inviting frontier developers to give the government up to thirty days of pre-release access to their most capable models, and to “collaborate” on choosing the trusted partners who would see them first. It was written as voluntary. Ten days later, the Commerce Department ordered Anthropic to cut off all access to two of its newest models, Mythos 5 and Fable 5, for every foreign national on earth, including its own employees. Two weeks after that, OpenAI launched its most powerful model, GPT-5.6 Sol, to roughly twenty partners whose names had been individually cleared by the US government. No ChatGPT. No waitlist.

Both companies objected. Both complied.

Look at what actually changed. The most capable intelligence on the planet did not get less capable. It got less available. And availability is the part that touches your work.

The ceiling and the floor

Hold two numbers in your head.

The ceiling is the smartest model that exists. It’s what the labs announce and what the headlines track.

The floor is the smartest model you are actually permitted and able to run. Not in theory. In your jurisdiction, at your company size, with your passport, on hardware you can buy.

For most of this technology’s short history, the two numbers moved together. The best model was a sign-up away, so the floor was the ceiling. June 2026 pried them apart. When the frontier ships to twenty cleared partners and your access turns on your citizenship, the ceiling becomes a spectator sport. The floor is the number that runs your business.

And you don’t set the floor. Two forces do. Policy decides what you’re allowed to use. Funding decides what exists below the restricted line. Neither one reports to you.

“Consumer” was always the wrong word

When we say a restriction hurts “consumers,” we picture a person who can’t open a chatbot. That’s the smallest piece of it.

The entity locked out of a governed frontier tier is just as often a company. A mid-sized firm that never makes the cleared-partner list. A startup incorporated in the wrong country. A multinational whose engineers hold the wrong passports, which is exactly whom Anthropic’s directive severed. The access question is fractal. It reaches from the individual up through the small business, the enterprise, and the nation-state.

We already have the map for the top of that ladder. US compute policy sorts the world into tiers: close allies with open access, most of the world on licensed access with caps, and a banned tier. The Gulf’s AI build-out runs on government-to-government arrangements with US oversight written in. What June added was the same logic pushed down one level, to the model itself, and pushed in one level, to the individual, by nationality.

That is the real subject. Not whether the model is smart, but who stands on which rung of the access ladder, and who controls the rung.

The reciprocity problem

Here I have to be careful, because what follows is a forecast, and forecasts are where strategists embarrass themselves.

If access becomes a lever, levers get pulled from both ends. A government that can switch a model off for foreign nationals has just shown every other government what’s possible, and what’s now necessary. Europe read the message inside the same news cycle. Its leaders started warning, out loud, against building a national tech stack on access that a foreign government can switch off overnight. Brussels’ proposed Cloud and AI Development Act is, in plain terms, a hedge against that switch.

I won’t tell you this hardens into permanent fragmentation. It might not, and I’ll come back to why. But the pressure is real and it runs both ways. Once intelligence gets treated as strategic infrastructure, every capable nation inherits the same two incentives: wall off a piece of it, and stop depending on anyone else’s wall. That isn’t doom. It’s the physics of the position.

Watch the floor, and ask who keeps building it

If the ceiling can be gated, the question that matters becomes how high the floor sits, and who keeps raising it.

The floor is high and still climbing. As of mid-2026, open-weight models you can simply download, including DeepSeek V4, Alibaba’s Qwen, and Moonshot’s Kimi, clear 80% on the coding benchmarks that defined the frontier a year ago. The pattern has held for two years. Open weights trail the frontier by six to twelve months, then close the gap on exactly the capabilities that were hardest last year. The floor is not the basement. It sits about half a year under the ceiling, and it’s the half you can actually own.

Then comes the uncomfortable question: who funds it? The commons gets built by interested parties. Meta opens its models to commoditize the layer, so no rival builds a walled garden around it. Nvidia has put tens of billions behind open weights because if open models win, inference still runs on Nvidia silicon, so it wins either way. Mistral sells sovereignty to Europe. DeepSeek and Qwen extend Chinese soft power and industrial reach. The floor exists because powerful players benefit from a capable commons today. It is not a public good with a public guarantor. It rises on strategy, and strategy turns.

Even a free model isn’t free to run. A capable open model fits on hardware a serious person can buy: a used 24GB graphics card for a mid-size model, a Mac Studio for the large ones. The true frontier still doesn’t fit in your office. It lives in a datacenter, behind an API, behind a policy. So the floor is real, ownable, and bounded. You can hold capable intelligence in your own hands. You cannot buy the best intelligence with a graphics card. The access frontier is, underneath, a compute-and-capital frontier.

The leadership decision underneath

Now it stops being macro and becomes your Monday.

The whole promise of agentic work is delegation. You hand real tasks to systems and trust the output enough to build process on top of it. You cannot durably delegate to a capability that someone else’s government can switch off. A workflow anchored to a model that disappears for your foreign-national team, or your jurisdiction, or your company tier, was never a workflow. It was a dependency in a workflow’s clothes.

So access sovereignty becomes an organizational design decision, the same class of decision as where your data lives or who can sign a contract. The grown-up version asks three quieter questions:

• What floor can I guarantee my people and my agents under any policy weather?
• Which of my critical workflows are anchored to a tap I don’t control?
• For each tap, do I know whose hand is on it?

Most organizations have never answered these, because for three years they never had to. The model was always there. June 2026 is the end of that assumption.

The case I might be wrong

You’re owed the other side. The June order was, on paper, voluntary, not a licensing regime. OpenAI said plainly that government-gated access should not become the long-term default, and promised general availability within weeks. The restrictions could turn out to be a preview-stage speed bump rather than a wall. The tiered system could settle into something closer to export licensing than a Berlin Wall for intelligence, managed and negotiated and open for most allies most of the time. I’d put real weight on that outcome.

But even in the optimistic case, the discipline survives. If access is licensed instead of free, you still want to know your guaranteed floor. If the frontier ships on a delay, you still build on what you can rely on and treat the ceiling as upside. The fragmentation thesis might be wrong about how far this goes. The operating advice stays the same.

The one question

We spent three years asking the wrong question with great precision. How smart is the best model turned out to be a benchmark for spectators. The question that runs your organization is quieter and harder.

If your most important model went dark tomorrow, switched off by someone else’s government in someone else’s interest, what would still run?

Everyone is benchmarking the ceiling. Build on the floor. And always know whose hand is on the tap.

Sources: White House executive order, June 2, 2026 · Anthropic, Fable 5 / Mythos 5 access · OpenAI, previewing GPT-5.6 Sol · TechCrunch on the OpenAI limit · IAPP on the Anthropic suspension · CEPA on US AI export controls · USCC, “Two Loops” report on China’s open-AI strategy

Daron Yondem advises senior technology leaders on AI-driven organizational transformation. Learn more →