Geopolitics · 6 min
When Washington controls access to models: export controls and cognitive sovereignty
Washington has applied export controls to the models themselves. A precedent that shifts the question from data to the capacity to think.
Mature open models, accessible GPUs, growing regulatory and geopolitical constraints: local inference is no longer an activist choice. It is a rational architecture option — provided you know its real terms.
As recently as eighteen months ago, hosting your own models was either an ideological choice or an extreme constraint. The situation has changed along the three axes that matter. The models, first: the current generation of open weights — Qwen, GLM, Mistral, DeepSeek and their peers — covers the majority of enterprise use cases at a largely sufficient level: summarisation, extraction, document RAG, code assistance, triage. The hardware, next: a pair of recent professional GPUs is enough to serve a quantised 30-to-70-billion-parameter-class model for a mid-sized organisation. The context, finally: between GDPR requirements on sensitive data, trade secrets, and the precedent of American export controls applied in June to frontier models, exclusive dependence on non-European APIs has become a risk that is identified, documented, and actionable in a risk committee.
Let us be precise: running a Chinese or American open-weights model on your own servers does not make the model European. But it changes what matters: the data never leaves the perimeter, no telemetry flows back, no terms of use can evolve unilaterally, and no foreign decision can withdraw weights already downloaded. Sovereignty of use — control of processing, of context, of the log — is secured. Sovereignty of design — who trained the model, on what, with which biases — remains a European work-in-progress, that of the continent's model makers, which deserves to be supported by demand.
A revealing debate followed the publication of the post-mortem of the July intrusion. Practitioners publicly asked the awkward question: is AI-assisted incident response — the very thing that made it possible to detect the attack and decrypt its payloads — reserved for GPU-rich organisations? The field reports are more encouraging than expected: mid-sized open models, served on modest configurations, are enough for useful forensic analysis; some report complete analyses carried out on a few compact machines. Sovereign response capability does not require a supercomputer; it requires having built and tested the pipeline before the incident.
Local inference has costs that must be named: hardware investment and operational skills (serving, quantisation, updates, continuous evaluation), a real performance gap with frontier models on the most demanding reasoning tasks, responsibility for security — a poorly isolated inference server is an attack surface, and the summer was a reminder: the connectors and tools you plug into models must be authenticated and bounded like any privileged access. The reasonable architecture is therefore hybrid and tiered: local by default for everything touching sensitive data, frontier APIs — ideally European, or contractually framed — for cutting-edge tasks on non-sensitive data, and a tested switchover between the two.
Local inference will not replace frontier models. It restores something more important: a negotiating position. You do not talk the same way to a provider you are able to leave.
Geopolitics · 6 min
Washington has applied export controls to the models themselves. A precedent that shifts the question from data to the capacity to think.
Strategy · 6 min
No European player will leave AWS, Azure or Google Cloud in the short term. Five moves to turn an endured dependence into a managed one.
Open source · 6 min
The July incident argued for openness as much as against it. Open source is not sovereign by nature — it can be made sovereign.
Book a meeting: we gladly turn an article into an answer to your specific case, with your hosting and compliance constraints.