This post was originally posted my Substack. I can be reached on LinkedIn and X.

Two weeks ago, Anthropic released Fable 5, the public-facing version of Mythos 5. For context, Mythos was hailed as the most capable model the company had shipped to date, with a step-change in its cyber capabilities. Days later, access was restricted after Amazon researchers showed that Fable could be jailbroken into producing information useful for cyberattacks. This ban barred all foreign nationals, including Anthropic's own non-US employees, from using Fable/Mythos. Anthropic, rather than separating US from non-US users, disabled both models for everyone as of writing.

There’s already plenty of public discourse about this ban. For investors, the more important point is that hosted frontier intelligence is now subject to state action, not just vendor policy. In this piece, I argue that just as compute and energy have defined the current AI trade, “intelligence sovereignty” (control of one’s own AI stack) will be one of the next investable themes. I’ll go through some of ramifications and investment opportunities below.

Model access becomes an intelligence sovereignty issue

Until recently, customers could treat American AI the way they treated AWS or Windows: stable infrastructure that they could build their businesses upon. That trust rested partly on the perceived stability of US policy. Fable proved that assumption wrong. Access to intelligence can be pulled with little notice. Even if Fable is re-released, customers will conclude that they can no longer solely depend on one country’s AI offerings.

Strategic customers outside the US, including some US allies, will likely seek to reduce their dependence on American intelligence. There is precedent for this. China, for example, has embarked on a multi-year initiative to migrate off US technology. This is a headwind for US frontier labs. While American labs may still win high-end workloads, the assumption that they will be the default intelligence layer for the rest of the world may no longer hold. Customers will increasingly invest in local intelligence “redundancy.”

The beneficiaries are likely to be homegrown champions at every layer of the 5-layer AI cake (applications, models, infrastructure, chips, and energy). As an example, the European champion at the model layer might be Mistral. Even though it trails frontier US models (and even Chinese models), strategic industries / government agencies might use its models because control over model weights confers strategic value.

We might be at “peak” open-source

The obvious alternative is open models. Here, the common assumption is that open-source models will keep pace with the closed frontier, giving everyone a credible fallback. The online discussions around the recent launch of GLM 5.2 might even imply that open-source has reached the frontier.

But GLM’s launch is merely a continuation of the trend that open models are several months behind the frontier. Here, open models can keep improving in absolute terms while falling behind in relative terms if frontier training costs rise, distillation targets become harder to access, and labs become less willing to release their best models. If that happens, we might soon approach “peak” open-source in terms of how close open models are vs. closed frontier models.

Yes, open-source will keep pushing the open frontier forward. But, as AI labs like Zhipu, Minimax, and Moonshot come under increasing pressure to grow revenues, they will likely start closing off their best models. This is already starting. Alibaba recently started to close off its high-end models (there’s some internal politics involved here). Furthermore, if the US restricts access to its best models to only a handful of vetted customers, then theoretically it’s harder for competitors to distill them. It is also unclear whether Chinese labs, or American open-source labs like Reflection and Arcee, can come up with algorithmic innovations to leapfrog the closed-source frontier.

The net result of this is that closed models and open models will both improve, but at different slopes. One question here is what happens to these open-source labs. Inference providers and hyperscalers might be able to optimize their stack (e.g., with custom hardware) better than the labs can, making it hard for capital-constrained open labs to compete on cost per token. Therefore, open models will end up being a marketing tool. Real revenue will likely come from the customization, packaging, and deployment of open models for enterprise and sovereign clients. It's also worth watching whether a company like Nvidia keeps putting out models that act as a capability floor, so that customers at least have a good enough open model to fall back on.

Even American companies will look to self-host

Beyond “intelligence sovereignty” for the rest of the world, some American companies will also build the capability to host their own models. Companies working in sensitive domains, such as cybersecurity and biology, are structurally disadvantaged vs. competitors who have access to Mythos / Fable-class models if they are excluded from trusted-access programs or forced to use nerfed APIs. In essence, model access becomes a competitive moat and turns frontier labs (and sometimes the US government) into kingmakers. One recourse is to use self-hosted / post-trained open models.

Given the technical difficulty of standing up high-performance inference infrastructure, I expect many companies to lean on compute providers with inference abstraction layers. This is a boon for inference players (e.g. Together AI, Fireworks, Baseten), neoclouds, and hyperscalers that have strong developer experience. This also benefits OEMs / ODMs like Dell / HPE, which serve larger organizations that can afford to build private clouds. A second-order effect beyond model access is that as an increasing fraction of white-collar work can be accomplished by open models, revenue at the low end will be siphoned from the frontier labs to compute providers. Frontier labs will need to keep widening the capability gap to sustain premium pricing. This is why both Anthropic and OpenAI have pushed to own the application layer (essentially charging for software vs. purely APIs).

Intelligence can be 'brute-forced' with energy and people

A more speculative question is how different tiers of intelligence change the shape of white-collar productivity. When a new model pushes the frontier forward, it gives the humans wielding it more leverage. The reverse is also true. Nations (and companies) without access to the best models need to make up for it with energy / compute and more humans.

Here, a weaker model might be able to reach a stronger one's capabilities by spending more compute at inference, at least for some workloads. This includes running more attempts, exploring more paths, checking its own work, and putting more human operators around weaker models to better steer them. In an intelligence-constrained regime, additional humans and compute are required to make up for the intelligence shortfall vs. a stronger base model.

This dynamic sets up an interesting comparison between the US and China. The US has stronger frontier models but has been slower to ramp up energy production. It also has fewer white-collar workers. China, by contrast, currently has strong but not frontier models. China will therefore depend more on energy, inference compute, and engineering labor to match American intelligence.

The next bottleneck is intelligence sovereignty

The fact that Fable got yanked is in some ways a “crossing-the-Rubicon moment.” So, if the AI trade so far is about owning the bottlenecks, the next trade is sovereignty. This favors regional players across the 5-layer cake, and it chips away at the revenue base of US frontier labs like Anthropic and OpenAI. This isn’t confined to software. As AI moves into robotics and physical systems, control over models, hardware, and supply chains becomes part of the sovereignty trade.