In mid-June 2026, the US government blocked international access to Fable 5 and Mythos 5, Anthropic’s most advanced security models. Three weeks later, the restriction was lifted. What happened in between shows, more clearly than any analysis could, why waiting in AI is not risk management, it’s the biggest risk.
Anthropic had no choice but to shut the models down entirely. The conditions imposed, such as access limited to US citizens only, were unworkable for a company with an international workforce, both internally and externally. In those same three weeks, Chinese AI provider Z.ai released GLM 5.2: a model with comparable capabilities for detecting security vulnerabilities, launched as an open-weight model, freely available worldwide. The result was straightforward and damaging. Attackers had a powerful tool. Defenders were legally barred from using the superior model to prepare against it.
On July 3, 2026, the export restriction was lifted. Three weeks of controls did not achieve their purpose.
What three weeks of export restrictions actually show
This is not an anomaly. A bill that passed a US House committee in spring 2026 would require exported AI chips to report their location and allow for remote shutdown . It is not yet law, but the direction it signals is clear. Politically driven technology volatility through export bans, model shutdowns, regulatory reversals is a permanent feature of the AI market. It is not something that can be planned around, and it will not stabilize.
The instinctive response is caution: wait until conditions clarify. But that response misidentifies the actual risk. Organizations that wait for stability before building capability are waiting for a moment that will not arrive. In the meantime, competition continues and the gap compounds.
Capability accumulates — so does standstill
Stefan Fritz describes this mechanism directly. AI competence is not learned through reading, strategy papers, or waiting for a mature European alternative. It is learned through practice: applied to real cases, pushed to the edges of what current tools can do.
That has a compounding consequence. A company that trains two hundred employees on real agent workflows today builds a lead that a competitor cannot simply acquire, because the knowledge is distributed, practically learned, and embedded in people. A company that waits two years does not start the learning curve two years later. It starts against someone who has been climbing for two years. The gap grows every month, because the tools improve and the learning curves accumulate.
There is a second effect: talent follows the best tools. Organizations that restrict access to current AI risk losing exactly the people they will need tomorrow — and often do not notice until those people have moved on.
Architecture decides, not the model
The right response to technology volatility is not less AI, but better architecture. Sovereignty does not emerge at the model layer. It emerges at the layer below : in orchestration, in knowledge management, in the workflows that an organization controls itself.
Building that layer creates optionality. When an export ban takes effect, when a superior model emerges, or when a European alternative becomes ready, the model at the top can be swapped. It becomes an interchangeable component of a system the organization owns.
We encourage our portfolio companies to build expertise today rather than wait for a more favorable entry point, which, structurally speaking, does not exist. The AI market is inherently volatile. Our “buy-and-build” approach is a direct response to this, because expertise cannot be built organically alone when the market is growing faster than any organic growth rate.
Companies that push today’s best AI to its limits while building the architecture for interchangeability win on two dimensions: a competence lead today, and freedom of choice tomorrow. Both require that the decision is made now.