Washington Pulled Its Best Model on June 12. China Open-Sourced the New World #1 on June 13. Export Controls Are Building the Stack They Were Meant to Stop.
The day after the US forced Anthropic to pull Fable 5 worldwide, Beijing's Z.ai released GLM-5.2 under an MIT license — independently ranked the #1 open model on Earth, #4 overall, at roughly a sixth of GPT-5.5's price. That isn't a coincidence; it's the shape of the whole strategy. The open frontier is now Chinese, it's de-risking off NVIDIA onto Huawei silicon backed by a $295B state demand guarantee, and 'open' has quietly become a geopolitical instrument — state voting rights, enterprise-only model sizes, closed-weight pivots. Export controls didn't contain this. They commissioned it.
On June 12, the US government forced Anthropic to pull Fable 5 off the market for every customer worldwide. On June 13, a Beijing lab called Z.ai released GLM-5.2 under an MIT license, and according to the one benchmark house nobody seriously accuses of picking sides, it is now the best openly available AI model on the planet — Artificial Analysis has it at 51 on its intelligence index, #1 among open weights and #4 overall, behind only two Claudes and one GPT. One day apart. Washington fenced off its frontier model the day before China gave away the new leader, at roughly a sixth of GPT-5.5's price, servable on a dozen clouds. If you want one image for how the export-control strategy is actually going, that's the one to keep.
The consensus in Washington is that the chip controls are working. They are — at precisely one thing, and it isn't the thing anyone intended. The controls have not preserved an American lead at the open frontier, because that frontier is now Chinese, independently verified and cheap. What the controls did was hand Beijing a reason to build a parallel stack — its own silicon, its own software, its own state-guaranteed demand — and a deadline to finish it. Containment became a construction grant. I'll caveat my own cynicism in a minute, because the bullish-China story is also oversold. But the policy premise is the bigger error, and it's worth seeing why.
Start with the leaderboard, because the handoff is real and it keeps landing in the same country. The open title passed from NVIDIA's Nemotron to DeepSeek V4-Pro to GLM-5.2 inside a few weeks; DeepSeek V4-Pro hit 80.6% on SWE-bench Verified in April and held the crown for about seven weeks before Z.ai took it. Baidu's ERNIE 5.1 landed fourth globally on Arena's search leaderboard; Kimi and Qwen fill in behind. The best American open model, Nemotron, sits below GLM on the same index — and it's "open" with an asterisk, shipped in an NVIDIA-native checkpoint that wants NVIDIA hardware to run well. The durable top of the open tier is Chinese. Meta's Llama, the model that was supposed to keep the West's open ecosystem in front, isn't in this conversation anymore.
Now the part both sides get wrong: the chips. The triumphant version says DeepSeek trained V4 "without NVIDIA." That's false, and it matters that it's false. CSIS's independent read is that Huawei's Ascend 910C runs at roughly 60% of an H100 for inference but is, in DeepSeek's own assessment, unattractive for training — what actually happened on Ascend was full-parameter post-training and serving, not frontier pre-training. So the sovereignty is partial. What's real is the trajectory: Huawei's CANN software stack is now production-grade for transformer workloads (still a fraction of CUDA's ecosystem, but no longer a toy), and the Ascend 950, due in the second half of 2026 and reportedly approaching H200 territory, is the chip to watch — with the caveat that the eye-catching 17x training-gain figures are Huawei's own and independently unverified. Underwriting all of it is a $295 billion state grid build with an 80%-local-supplier mandate, sovereign-debt financed, after Beijing blocked H200 imports specifically to force domestic demand. That's not trade policy. It's a guaranteed customer for a chip that isn't fully competitive yet, which is exactly how you fund the ramp until it is.
The third thread is the one almost nobody names: "open" here is a strategy, not generosity. DeepSeek raised $7.4 billion, and the catch is the governance — China's National AI Industry Investment Fund holds the sole voting rights, while Tencent, CATL, and the commercial investors took economics only and a five-year lock-up. Open weights, state-controlled vendor. Alibaba closed Qwen 3.7 Max's weights after earlier Qwen versions had seeded Western developer stacks — and it built the thing as a Claude drop-in, native Messages API and all, so the off-ramp from Anthropic runs straight into a model whose next version you can't self-host. Kimi K2.7 is "open" at a trillion parameters and 340 gigabytes, meaning open to anyone with a data center. ERNIE's much-quoted "6% of the cost" conveniently excludes the parent model's training bill. Openness, in other words, is being used as a distribution wedge and a geopolitical lever, and it should be read as one.
Here's the discipline the story needs, because the cynicism cuts both ways. The bullish-China hype overstates sovereignty: Ascend still can't train frontier models well, the "without NVIDIA" claim is wrong, and the 1-million-token "fully usable" context on GLM-5.2 is the same long-range-coherence promise everyone makes and few keep — treat it as unproven until someone stress-tests it. Vendor "#1" claims from Z.ai, Baidu, and Alibaba should be kept strictly separate from the independent Artificial Analysis and Arena numbers; only the latter carry weight here. The closed frontier, for now, is still American — Fable 5 sits at 60 on the same index, Opus 4.8 at 56, comfortably above GLM's 51. The lead the US still holds is real. It's just at the closed frontier, not the open one, and that's a narrower moat than the policy assumes.
Where this goes is a two-part watch. First, the Ascend 950: if it delivers credible frontier training throughput at cluster scale — not just respectable per-chip inference — the last real bottleneck in the parallel stack starts to give. Second, Z.ai's forecast of an MIT-licensed "Open Fable," an open-weight equivalent to Claude Fable 5, by the end of 2026. If a Chinese lab actually ships a Fable-class model under a permissive license, trained on sovereign silicon, the containment narrative doesn't bend — it collapses, because the whole theory was that compute access gates capability. I think the open-frontier handoff is durable and the silicon catches up slower than Beijing wants and faster than Washington expects. What would change my mind is concrete: Ascend's training gap persisting into 2028, CANN failing to close on CUDA, and these open leads turning out to be benchmark-saturated rather than useful in production. Until then, the most quotable fact in AI policy is the calendar. June 12, then June 13. The fence went up, and the next day the better model walked around it.