The AI Decoupling Just Went Directional — and It's Climbing Out of Reach of the Supply Chain

A month ago I argued that AI was splitting into two stacks — a US one and a China one — the same way at every layer: models, silicon, fabs, memory. That was the geography of the thing. Same product, separate supply chains, a wall running top to bottom. I also argued the wall stopped at the bottom, because both sides still buy their high-bandwidth memory from Korea and their advanced wafers from TSMC. The split was real, and it was capped.

I need to update that, because in the last week the story changed shape. The decoupling didn't just widen. It changed direction. The two sides are no longer building the same thing in separate factories — at the top of the stack, they're building toward opposite ideas of what AI is even for. And the layers where that's happening are precisely the ones no shared supply chain can reach.

Here's the single contrast that made me rewrite the thesis. This same month, in the United States, Visa and Mastercard wired payment rails into chatbots so that AI agents can transact on your behalf, and Google replaced its search box with an agent that acts rather than lists. Also this month, four US companies stood up armies of forward-deployed engineers to embed AI inside the enterprise — Microsoft alone committed $2.5 billion and roughly 6,000 people. The West is racing, as fast as it can, to make AI autonomous, transacting, and human-served.

And in China, the government set a deadline of July 15 to switch consumer AI agents off.

The stack now has three zones, not two

Before the reframe tempts anyone into a clean East-versus-West morality tale, let me impose the discipline that keeps this honest. The divergence is not uniform. The stack now has three distinct zones, and they behave differently.

The bottom is still converged. Both blocs buy high-bandwidth memory from Samsung and SK Hynix in Korea, and both depend on TSMC's advanced nodes. The physical foundation is shared, full stop. The decoupling even leaks upward from there: China's best open model, LongCat-2.0, is freely downloadable in the West and Chinese models now carry a large and growing share of global developer traffic. Anyone selling you a hermetic two-worlds story is skipping the floor both worlds stand on.

The middle is parallel, not directional. Models and silicon are bifurcating by geography — LongCat trained on domestic Chinese chips, US frontier models on Nvidia — but both sides are racing in the same direction: build the most capable frontier model you can on whatever silicon you control. That's two teams running the same race on separate tracks. Calling it "opposite directions" would be exactly the narrative overlay I want to avoid.

The top is where it goes directional. And that's the news.

Axis one: the West productizes the agent, China switches it off

China's Interim Measures for the Administration of AI Anthropomorphic Interactive Services — a joint act of the cyberspace, planning, industry, public-security, and market regulators — take effect July 15. The consequences are concrete and brutal. Alibaba's Qwen disables humanlike and user-created agents starting July 10; Doubao, ByteDance's, turns its agents off July 15, goes read-only until October, then deletes them, redirecting users to a different, compliant app. No migration path. The agents don't get tuned. They get erased.

The mechanism is the part worth slowing down for, because it reveals intent. The rules mandate anti-addiction systems, usage notifications, and instant-exit mechanisms — and those requirements are architecturally incompatible with a persistent-memory companion agent. A companion built to hold a continuous emotional relationship cannot implement the friction the law demands; the two things cancel out. So China is not regulating the consumer agent. It is regulating the product category out of consumer reach, with an explicit driver of banning virtual-intimacy services aimed at minors. The same layer the US is wiring a credit card into, China is making illegal to build.

Axis two: the US ships engineers, China ships a download link

The second new axis is go-to-market, and it's just as opposed. The US frontier is winning the enterprise by shipping people — the forward-deployed-engineering armies above, Palantir's playbook bolted onto foundation models. Embed humans, capture the account.

China is doing the reverse. Chinese labs now serve a large slice of global developer traffic — over 45% of OpenRouter's, up from under 2% a year ago — and they do it through open weights, low prices, and marketplace distribution, not embedded staff. One Moonshot executive reportedly said the company simply won't do heavy last-mile delivery; it will win through the marketplace instead. I'll be careful there — that crisp refusal is a single exec's line, not a declared industry policy, so don't hang the axis on the quote. Hang it on the traffic. The revealed strategy is unmistakable and it generalizes: Chinese labs distribute the way you distribute when you have neither the capital to field consultant armies nor the access to place them inside the Fortune 500 — you flood the marketplace and let price and openness do the work. Two continents, one technology, opposite go-to-market philosophies.

Why the split can go directional up top when it's stuck down below

Here's the connection that ties FEATURE-001's cap to FEATURE-002's escape, and it's the thing most coverage will miss.

Last time, the decoupling ran models to silicon to memory and stopped at memory, because Korea binds both blocs and a supply chain is a physical chokepoint you can't wish away. That logic still holds — for the bottom of the stack. But the two new axes sit above the model layer, at agents and go-to-market, and those layers are governed by law and product culture, not by fabs. China can force-quit a companion agent regardless of whose memory it runs on. The US can wire Visa into ChatGPT regardless of whose fab printed the chip. The binding chokepoints cap the bottom of the decoupling. Nothing caps the top, because nothing physical is holding it together up there.

That's the whole point, and it inverts the reassuring read. The comforting version of the decoupling story was: the shared foundation limits how far apart the two stacks can drift. True — at the foundation. But the divergence is migrating into the layers the foundation doesn't touch. Both sides can keep buying the same Korean memory forever and still build fundamentally opposite products on top of it.

What I think

I think the middle of the stack stays parallel for a good while — both blocs keep racing to build frontier models on their own silicon, and that's geography, not philosophy. I think the bottom stays shared until China's domestic memory catches up, which is not soon. But the top keeps diverging, and it accelerates, because the forces driving it — regulation on one side, product-and-capital culture on the other — have nothing to do with the supply chain and everything to do with two states' incompatible answers to a single question: what is consumer AI for? The West's answer, right now, is an autonomous agent that transacts and remembers you. China's answer, as of July 15, is that a machine pretending to be your companion is a harm to be switched off.

What would falsify me: if China's anthropomorphism rules turn out to be enforced loosely and the agents quietly return, or if a US regulator decides autonomous transacting agents are a consumer-protection problem and claws that axis back toward China's posture. Either would collapse a directional axis into a parallel one. Watch the July 15 deadline actually bite, and watch whether Washington gets nervous about agents that spend money on their own. Until then, the map I'd hand you is three zones: a converged floor, a parallel middle, and a top that's pulling in opposite directions — out of reach, now, of the one thing that was supposed to hold it together.