A week that said a lot about where AI is going (and who controls it)

A lot happened in AI last week. Some of it will fade. What won't: a shift in who controls the most powerful technology in the world, and who gets left out.

TL;DR: Frontier models matter more than ever, and so does who controls them. Anthropic builds the best, OpenAI is catching up, and the week showed that access to the most capable AI can disappear overnight for political reasons. How companies integrate AI, what they do with their data, and whether they can switch providers will define the next few years. Open source is becoming less of an ideological choice and more of a survival strategy. Europe is an ocean behind, with no clear plan to catch up, despite some promising initiatives.

---

Some weeks in tech are just noise. Last week wasn’t one of them.

Three things happened in quick succession that, taken together, say something worth paying attention to — not just about AI companies, but about who controls them, who benefits from them, and what the real race might actually be about.

First: the Fable 5 and Mythos 5 episode

Anthropic launched its two most capable models, Fable 5 and Mythos 5, and within days, the US government issued an export control directive barring access by any foreign national. And that’s whether inside or outside the United States, including Anthropic’s own employees.

Anthropic complied but disputes the directive, arguing the standard would effectively halt all new frontier model deployments across the AI industry. The company said it believed the order was triggered by a jailbreak technique. But that models from other providers, including GPT-5.5, possess the same capability.

The result: a model presented as the future of AI became unavailable because of a political decision made on a Friday afternoon.

For anyone outside the United States, this was a vivid reminder of something that has often felt abstract. The world’s most advanced AI systems are increasingly strategic assets. Access to them is controlled by a handful of American companies and shaped by American political decisions. Europe has spent years discussing technological sovereignty. This week, it saw what dependence actually looks like in practice.

Then came Satya Nadella (Microsoft)

On Sunday, Microsoft CEO Satya Nadella published a long essay (rare for him, and interesting) arguing that most companies are focused on the wrong question. The title said it plainly: A frontier without an ecosystem is not stable. It reached 51 million impressions in under 24 hours.

The real opportunity, he wrote, is not in picking the best model but in building a learning loop on top of models, where human capital and token capital compound together. A company should be able to swap out a generalist model without losing the expertise built into their learning system.

He pushed back on fears that AI erodes human value: “Human capital does not become less valuable as token capital grows. It only becomes more valuable.”

It’s a compelling vision. It’s also worth noting the context in which it arrives: Microsoft is not obviously winning the frontier model race — and has been dependent on OpenAI for its own. If models become commodities and ecosystems become the primary source of value, Microsoft’s position suddenly looks considerably stronger. That doesn’t make Nadella wrong. But it’s worth keeping in mind when evaluating the argument.

There are also harder questions his essay raises without quite answering. The idea that every organisation possesses unique institutional knowledge that AI will amplify sounds reassuring. But a surprising amount of what companies call tribal knowledge is scattered across spreadsheets, procedures and industry practices — much of it not particularly unique, and increasingly the kind of thing frontier models are becoming good at absorbing.

And there’s a subtler tension. When companies bring in AI labs to help build workflows, evaluation systems and business processes, the AI company doesn’t just deliver a service. It builds a cognitive loop, and learns how real organisations function in the process. Traditional consultants leave with slide decks. AI platforms leave with something more durable. Whether that becomes a long-term problem is still unclear. But it’s a question worth asking.

And then there’s Fusion (mix models)

Maybe not the biggest story of the week, but it points somewhere real, and it matches how many people are already working (I’ve been using both Codex and Claude Code). OpenRouter launched Fusion — a system that orchestrates multiple models into a single workflow. Early results suggested that carefully combined mixtures of models could match or exceed frontier performance on some tasks, at lower cost.

The significance goes beyond one benchmark. For years the assumption was simple: the smartest model wins. Fusion suggests something different. For many real-world problems, the optimal solution may not be a single model at all. One extracts information, another validates, another reasons through exceptions, another reviews. The value shifts from intelligence itself to the orchestration of intelligence.

The “sovereign AI” question

The Fable 5 episode also makes certain claims feel more fragile than they did a week ago.

Rio 3.5, a model that briefly made the “X/Twitter” news as a supposed Brazilian AI breakthrough, turned out to be a weighted blend of an existing open-source model and Qwen 3.5, with no original training. It was downloaded over 110,000 times before the authors acknowledged the issue, claiming they had uploaded the wrong file. A cautionary tale about how easily “sovereign AI” announcements can outrun the reality.

Mistral, Europe’s most visible AI startup and the company French President Macron has publicly championed as Europe’s answer to OpenAI, is a genuinely impressive company. But the gap between Mistral and the models that just got restricted is significant. When the most capable systems can be switched off by a government directive, being a credible second-tier alternative and being a true frontier competitor are very different things — and second-tier here makes all the difference.

One of the most striking things to surface this week was Europe 2031 (europe2031.ai) — a detailed scenario project by Judith Dada, Daan Juijn and colleagues, months in the making, that maps exactly where Europe’s current AI trajectory leads. It’s told through fictional characters but grounded in real trends: a French policy worker in Brussels watching, year by year, as Europe debates sovereignty while the gap widens. The compute numbers alone are striking, the US is projected to hold a 12x advantage by 2031.

The scenario doesn’t predict catastrophe as inevitable. It argues the window to change course is open, but closing fast. And it was published this week, as the EU Commission was already scrambling to assess the practical consequences of the Fable 5 directive. Several of the things the scenario imagined as future risks (access rationing, inference caps, European exclusion from key model launches) are already happening, ahead of schedule. Worth reading in full.

A harder question about Anthropic

The week also raised something that goes beyond geopolitics: a question about Anthropic itself, and what it believes it’s entitled to decide on everyone else’s behalf.

Ben Thompson at Stratechery published a sharp analysis today that is worth reading in full. His argument centres on a series of decisions Anthropic made around the Fable 5 launch that received less attention than the government directive. Anthropic initially implemented hidden safeguards that would silently degrade Fable’s performance if used for frontier LLM development, without disclosing this to users. They later walked this back, replacing it with a visible handoff to a less capable model. Separately, the company changed its data retention policy at launch, announcing it would retain all user data for 30 days — even for enterprise customers who previously had zero data retention guarantees — without clarifying whether that data would eventually be used for training. This isn’t a theoretical concern: companies like Figma have seen how quickly Anthropic can move into adjacent territory, sometimes without warning, even after early partnerships.

Thompson’s conclusion from these decisions is direct:

“Anthropic believes that they are the ones who should have final say over how Anthropic is used; given that they think only they should be developing leading edge AI, they by extension think that only they should have final say over AI generally. When you further combine this realization with the company’s pronouncements about AI’s ability to conduct all economic activity, you realize that Anthropic’s leadership effectively wants to have power over everything and everyone.”

He adds something that cuts both ways:

“I respect this alignment, and I fear it. The closest analogy is probably Apple, which has always framed every self-serving action in the guise of doing right by users — and often they were. So it is with Anthropic.”

I’d push back slightly on the Apple comparison. Apple’s approach was never quite this openly contradictory, never this reliant on fear as a marketing lever, and never this quick to reverse its own stated positions within the same launch cycle. The pattern with Anthropic feels different, and gives a bit more goosebumps (not the good kind).

Dax, the founder of OpenCode, endorsed Thompson’s analysis on X this week. These aren’t people hostile to AI. They’re people embedded in the ecosystem, paying close attention.

None of this makes Anthropic’s models less impressive, or the government’s directive less blunt. But it adds a layer to the week’s events worth sitting with. The question of who (and how) controls frontier AI isn’t only about governments and export controls. It’s also about the companies themselves.

What connects all of this

Taken together, these events point to the same underlying question: where does the real value in AI actually sit, and who gets to decide?

Anthropic’s Fable 5 episode shows frontier models matter enormously, and that access to them can vanish overnight. Nadella argues the learning loop around those models matters even more. Fusion suggests much of the future value may sit in the layer that coordinates models rather than in any individual model itself. And Thompson’s piece reminds us that the companies building those models have their own views about who should be in control (views they are increasingly willing to act on).

The restrictions on Fable 5 and Mythos 5 have reignited debate about open-source models, sovereign infrastructure and locally controlled AI systems, not as ideological positions, but as practical hedges. An escape hatch only matters when something goes wrong. This week, something went wrong.

The more important race may not be about building the smartest model. It may be about control: of the models, the workflows, the data, the orchestration layer, and ultimately the ability to switch when circumstances change.

Last week made that race considerably more visible.