When access becomes the moat

The market already voted — before the lockup

This isn't a forecast; the migration was underway before June. By early 2026, OpenRouter data had China-built models at a large and fast-growing share of total token consumption — by some widely-cited readings a majority, up from under 2% in late 2024. Treat any single percentage with care; methods and weeks vary. But a16z — no friend of Beijing — reports that around 80% of developers who use open-source AI tools reach for Chinese ones, and Alibaba's Qwen has crossed a billion downloads and now anchors a large share of new model derivatives on Hugging Face.

That's what 'default layer' means. a16z's own argument is that whoever supplies the open foundation sets the incentives and norms of the ecosystem — and right now that supplier is increasingly Chinese. Gating the US frontier doesn't reverse that drift. It removes the most attractive reason to stay.

The steelman — and it's a real one

The controls have a serious case, and fairness demands stating it plainly. Dario Amodei's foundational argument is that keeping democracies at the frontier is existential, that efficiency gains from labs like DeepSeek sit on an expected cost curve rather than a breakthrough, and that gating compute and capability is one of the few levers that actually bite. And the gap isn't folklore: a US government (NIST) evaluation found DeepSeek's latest lagging the frontier by roughly eight months on private benchmarks, and scoring better on its own self-reported evals than on the independent ones. On the hardest long-horizon and adversarial work, 'good enough' is still not good — and there are real data-governance reasons an enterprise might not pipe sensitive traffic to a Chinese-hosted API.

But notice that two of those points cut the other way. If open Chinese models are weaker, the security case for gating US models weakens too. And the data-risk worry is mooted by the very thing that makes these models spread: open weights run on your hardware, with no API home to phone.

The redistribution nobody legislated

Here's the part this site can speak to from its own data, and it sharpens the irony to a point. Fable5.app exists because a great model got gated — and by our own search data, the largest share of demand reaching it comes not from Silicon Valley but from price-sensitive developers in India, Pakistan, and other developing markets who could never access or afford the gated frontier in the first place. That's exactly the population open weights were built to serve: free to download, runnable anywhere, no Washington approval required.

Our read: export controls were designed to deny capability to a rival state. Their most reliable real-world effect may be to deny the gated frontier to a developer in Lahore — who shrugs, pulls GLM-5.2 off Hugging Face, and ships. The policy aimed at a government; the burden lands on individuals, and it pushes them straight into the open ecosystem the policy was meant to contain.

You can win the benchmark and lose the default.

The verdict (our opinion)

When capability gaps shrink to a quarter and access becomes the binding constraint, the moat moves. The durable advantage stops being owning the best model and starts being being the model everyone can actually run. Locking the frontier doesn't freeze the race — it concedes the substrate.

Where we could be wrong: four months is an average, and on the genuinely hard, long-horizon, adversarial work the frontier's lead is wider and stickier than a coding benchmark suggests — and that's the work that mints enterprise revenue. The controls could also still bite at the compute layer, where they were always aimed, slowing the next open generation before it ships. But none of that changes the direction of travel visible today: the West is optimizing for a lead it can measure, while the rest of the world optimizes for a tool it can keep.

The takeaway