The common belief: open-source AI routes around government control. You can’t un-download a file. The semiconductor analogy says otherwise — and June 12 proves it.
The 45nm Chip Analogy
In semiconductors, the equivalent of “open” was always the commodity tier — older process nodes, widely diffused designs, academic papers anyone could read. The state never tried to fence that tier. Why? It wasn’t considered strategic. 45nm chips ship everywhere, even today, because nobody important cares to stop them.
Open-weights AI is in the same position. Today’s best open models — Llama, Mistral, Qwen, DeepSeek — are roughly comparable to last year’s closed frontier. They flow freely because they’re one step behind the line the state has decided to fence, not because they’re immune to fencing.
The correct read: Open-source is currently below the fence because the frontier is above it. As open models close the gap, the fence either descends to capture them — or the industry splits in two.
Three Fencing Mechanisms Already Visible
Capability threshold
At the frontier, openness isn’t the variable that matters — capability vs. risk is. Below the threshold, open-weights flows. Above the threshold, the fence applies regardless of the license on the file.
Why DeepSeek Terrified Washington
This is why DeepSeek’s open releases were read in Washington as a national-security event, not a research event. A non-US lab releasing a frontier-class open model is a direct strategic act — the cleanest available move for a state outside the US hub to deny the hub’s fencing power.
The chip-war pattern predicts this precisely: when the frontier escapes the fence, the fence either expands to grab it back, or the world accepts that the frontier has split into two competing systems. Both processes are already in motion.
The Bottom Line
Open-weights is a real release valve and a genuine source of diffusion. It is not a structural escape from this architecture. It is the part of the stack the fence has not yet reached — for reasons that are economic and tactical, not principled. That distinction matters more every quarter.
Source: Business Engineer — The Geopolitical Fencing of Frontier AI








