The Great Decoupling: How U.S. Export Policy Is Building the Chinese AI Stack It Meant to Stop

A food-delivery company trained an AI model that beats GPT-5.5 at writing code, and it did it without a single Nvidia chip.

That sentence should stop you, and not because Meituan — yes, the app that brings you dinner — decided to open-source a 1.6-trillion-parameter model called LongCat-2.0. Companies ship models every week now; the leaderboards are a landfill. It should stop you because of how they did it: pre-trained end to end on roughly 50,000 domestic Huawei Atlas-950 accelerators, stitched together with HCCL, China's homegrown answer to Nvidia's networking layer. It leads OpenRouter on real usage. It scores 59.5 on SWE-bench Pro against GPT-5.5's 58.6. And the company that built it does not, in any meaningful sense, exist to build AI.

Here is what I think that means, and I'll defend it: the AI world is not drifting toward a U.S.-China rivalry. It has already split into two stacks, and it is splitting the same way at every single layer — models, silicon, fabs, media. The isolated business stories everyone is filing — a Chinese model tops a leaderboard, Korea announces a fab plan, OpenAI kills Sora, Kuaishou spins off its video unit at $18 billion — are not separate. They are one story. 2026 is the year the decoupling stopped being a talking point and became structural. And the U.S. export regime, the entire apparatus built to preserve an American lead, is the thing accelerating the fork.

Let me walk down the stack, because the pattern only becomes undeniable when you see it repeat.

Layer one: the models already decoupled

Start with the part that's already over. The number-one open-weight model in the world is Chinese, and it has been for long enough that the qualifier is getting stale. GLM-5.2 tops the Artificial Analysis composite at 51 to DeepSeek V4-Pro's 44 — that's a Chinese model beating another Chinese model, with the best American open model further down the list. This is not a benchmark China is chasing. It's one they own.

LongCat-2.0 is the escalation. Until this week, the honest caveat on Chinese AI was that "trained without Nvidia" meant post-training — the fine-tuning polish at the end, not the brutally hard pre-training run where you actually build the thing. That caveat is now dead. LongCat is the first confirmed Chinese frontier-grade model pre-trained end to end on domestic silicon. A non-lab did it. On chips Washington's rules were designed to deny.

And here is the part the West keeps misreading: none of this "open" is generosity. When DeepSeek raised its latest round, the structure handed China's state fund sole voting rights. Enterprise-grade model sizes are gated. "Open weights" is not a gift to the global commons — it's a distribution channel with a flag on it. More on that later, because it's the whole game.

Layer two: the silicon is forking in real time

The H200 story is the tell. Beijing didn't just tolerate Washington's export restrictions on advanced chips — at a certain point it blocked H200 imports itself, to force domestic buyers off the American supply and onto Huawei's Ascend line. Read that twice. The export war reached the point where China started doing the export controllers' job for them, because a captive domestic market is exactly what you need to fund a homegrown chip ramp. LongCat's 50,000 Atlas-950s are the payoff on that bet.

On the other side of the wall, the U.S. is building its own inference silicon — Etched's Sohu, the lab-designed accelerators — for a market that increasingly won't touch Chinese chips and can't sell into China anyway. Two silicon ecosystems, each optimizing for a customer base the other can't reach. That's not competition. That's a partition.

Layer three: compute became industrial policy

Money reveals intent, and the money here is national. China is standing up a roughly $295 billion compute grid, financed with sovereign debt — the state balance sheet, not venture capital. Korea has trumpeted a $576 billion drive. Both numbers deserve a hard look, and I'll give Korea's the scrutiny it needs in a moment, because a chunk of that is re-badged private Samsung and SK Hynix capex wearing a national-strategy costume. But the direction is unambiguous: compute has been reclassified from a business input to a strategic resource, the kind governments build grids and reserves around.

Which brings us to the crack in my own thesis — the thing that keeps this piece honest.

Where the split stops: Korea owns the floor

If the decoupling were complete, I'd be writing hype. It isn't, and the reason is memory.

High-bandwidth memory — HBM, the fast RAM that sits next to every AI accelerator — has not split. Samsung and SK Hynix supply something like two-thirds of the world's frontier memory, and they supply both stacks. China's Atlas-950s and Ascend parts still need HBM where Korea dominates and China's domestic answer, CXMT, is nascent. Western frontier GPUs — Nvidia's Rubin line, HBM4 — depend on the same Korean suppliers. The second shared chokepoint is TSMC's 4nm node, which even the plucky American upstarts like Etched can't route around.

So the honest map looks like this: the split is real from the models all the way down through the logic silicon — and then it hits the bottom of the stack and stops. Both blocs still meet at Korea. A genuine, clean decoupling requires severing HBM and TSMC, and neither side is remotely close. Anyone selling you a finished two-worlds story is skipping the floor both worlds still stand on.

Layer four: the media bet, funded on one side and buried on the other

Now the layer that prompted this piece, and it rhymes with everything above so precisely it's almost a diagram.

On July 2, Kuaishou spun its Kling video unit into a standalone company and raised somewhere between $2 and $3 billion at an ~$18–20 billion valuation — the largest dedicated capital event AI video has ever seen. The investor list is the actual story: Alibaba, Tencent, and Baidu — Kuaishou's direct rivals — all put money in, alongside Abu Dhabi's BlueFive Capital. Chinese Big Tech collectively bankrolling a competitor's video champion, with Gulf money riding along. And it's commercially real, not a demo reel: Kling's annual recurring revenue hit roughly $500 million in March, up from $300 million in January.

Run the split against the same calendar. As China capitalizes a video champion, the American incumbent exits: OpenAI shut down Sora over unit economics — $1 million a day to run, $2.1 million in lifetime revenue. That is one of the worse ratios I have seen attached to a flagship product, and OpenAI, to its credit, read the spreadsheet and pulled the plug. Western AI video is now a price war among Veo, Grok Imagine, and Runway. Eastern AI video is a funded national bet. Same technology, opposite strategy, and the strategies map cleanly onto the two stacks.

The thing almost everyone will miss

Here's the reframe I'd stake the piece on: "open" and "sovereign" are the same strategy wearing two faces.

The West reads "open weights" as generosity and reads "export control" as strength. The corpus says both readings are backwards. Open weights — GLM, DeepSeek, LongCat, Kling — are a geopolitical distribution channel: seed the world's developers and creators with your model, capture the ecosystem, and keep the control levers (state voting rights, gated sizes, the option to close the weights later). Sovereign compute — Ascend, the Atlas-950, the $295 billion grid — is that channel's supply chain. Open is the Trojan horse; sovereign is the fortress it rolls back into. They are one policy.

And export control? On the evidence, it is an accelerant, not a creator. I want to be exact about how strongly to say this, because the overfitted version — "export controls conjured the Chinese AI stack from nothing" — is wrong, and the underfitted version — "the controls are working" — is wishful. The truth sits in the timelines. The H200 block preceded the Ascend ramp that LongCat rode. The Fable 5 restriction landed on June 12; GLM-5.2 shipped on June 13, one day later. Those coincidences are real and documentable. But China's sovereign push — DeepSeek, Huawei's Ascend roadmap — predates 2026 by years. So the defensible claim is this: each denial accelerated a substitute that was already in motion. Export policy didn't build the wall. It poured the concrete faster.

What to watch

One question decides whether the two stacks become two worlds: memory. Watch CXMT. The day China's domestic HBM stops being nascent and starts being credible is the day the floor cracks and the decoupling runs all the way down. Until then, both blocs are standing on a Korean foundation, and every triumphalist chart on either side has an asterisk pointing at Samsung and SK Hynix.

I think the fork keeps widening through layers one, two, and four, and stalls at layer three's memory floor for longer than either capital feels comfortable admitting. I've been wrong about the pace of Chinese hardware before — I did not expect a food-delivery company to be the one to kill the "post-training only" caveat this year. Which is rather the point. The substitutes are no longer arriving on schedule. They're arriving early.

A note on what's proven versus what's asserted, because a story about a decoupling shouldn't launder claims across the wall it describes. Independently verified: GLM-5.2's leaderboard position, LongCat leading OpenRouter, Kling's ~$18 billion raise, Sora's shutdown economics. Vendor- or state-sourced and not yet independently reproduced: LongCat's all-domestic pre-training and 59.5 SWE-bench Pro score (Meituan's own technical report), DeepSeek's sole state voting rights (effectively single-source), and a good chunk of Korea's $576 billion (re-badged private capex, not new state money). Treat those as claims, not facts. The thesis survives even if you discount them — because the one number that anchors the whole piece, the Korean HBM chokepoint that binds both stacks, is the one nobody is disputing.