Last updated on July 1st, 2026 at 05:28 am

An update to “The Great AI Divergence,” June 2026

Executive Summary:

In March, the AI divergence between the United States and China read as a trade story. Washington was fighting with tariffs and chip export controls, treating technology as an asset to be restricted, while Beijing published a coordinated industrial plan to capture the global AI substrate, the open-weight models and infrastructure the rest of the world would build on. The diagnosis was that the US was waging a 20th-century skirmish over a 21st-century structure, and losing the part that mattered. Three months later the lever has moved. The pressure shaping the divergence is no longer mainly trade policy but AI policy, and Washington is now applying it to itself. Pulling frontier Mythos-class access and routing it through export-control machinery told every enterprise and government buyer that the American frontier they were building on could be withdrawn by policy, on a timeline they did not control. Buyers responded the way they respond to any supply risk: they went looking for a fallback that could not be export-controlled, which in practice means open-weight models they can self-host.

What was waiting in that tier was Chinese. GLM-5.2, DeepSeek V4, and the Qwen family are the strongest, cheapest, most permissively licensed open models available, so a buyer eliminating export-control exposure is now routed toward Beijing by process of elimination. The trend began in the startup tier, where inference costs bite hardest, and is surfacing at Shopify and Airbnb scale, directionally real even where the headline claims overstate the magnitude. The pattern behind it is the one to watch: every American move framed as a show of dominance, from chip controls that bred DeepSeek to model restrictions that are now pushing US firms toward the Chinese supply chain, produces the opposite of its intent. The clearest anomaly is Meta, the obvious Western open-weight candidate, which runs no first-party enterprise API and is barely in the conversation, a gap that deserves its own examination. The opening is being contested by Cohere and Mistral, whose sovereign open-weight offerings turn the same self-hosting mechanism to a friendlier flag. Whether they scale fast enough to catch the demand American policy keeps generating, before it sets as Chinese dependency, is the open question of the coming year.

————

UPDATED July 1: Commerce Secretary Howard Lutnick announced today that the two week restrictions on Mythos-grade model(s) have been lifted and presumably on GPT 5.6 and any other American model products as well. The likelihood is that this was done in order to address massive migration to Chinese models by both startups and established companies

June 30: It’s fascinating to be situated in Cambodia, while trade and technology wars erupt that have massive implications for this part of the world. You can see it here in both increments and increasingly large ways. The US dollar is accepted here as readily, and possibly more readily, than Cambodian riel, but there is not a large American presence here. The majority of tourist and foreigners are from Australia and Europe, but even that traffic has been impacted by the conflict in Iran. It’s starting to settle down, but the impact on fuel prices here, the impact on flight continuity and the very direct impact on the hotel where I live has been visible. At the same time watching the very subtle growth of Chinese AI adoption has been fascinating as is the way that China’s leveraging AI in significant efforts to better society. This operates in a near antithesis of how US companies are both managing and building their capabilities and the distinction becomes more clear and more stark by the day. The dispute has been both direct and subtly indirect.

The original argument was about trade. Washington was levying duties, restricting 2nm chips, and treating technology as one more asset in a tariff fight, while Beijing published a coordinated industrial plan to capture global AI foundational infrastructure. The diagnosis then was that the United States was fighting a 20th-century skirmish over a 21st-century structure. The impact must’ve been significant because Trump made a journey to Beijing to meet with Chinese leadership at a tenor that seemed much more cooperative than competitive. The tenor around AI has not been nearly as equitable.

Three months on, the lever has moved. The pressure shaping the divergence is no longer mainly trade policy. It is AI policy, and it is being applied by Washington to Washington. The decision to pull frontier Mythos-class access and route it through export-control machinery did something tariffs never managed: it reached inside the American stack and shook loose the demand it was meant to protect.

What the Suspension Actually Did

Export controls on chips are a bottleneck other countries engineer around. Restricting access to your own most capable models is different in kind. It tells every enterprise architect, procurement lead, and government buyer that the frontier they were building on can be withdrawn by policy, on a timeline they do not control, for reasons that have nothing to do with the product.

Buyers respond to that the way they respond to any supply risk. They look for a fallback that cannot be export-controlled, and in practice that means open-weight models running on infrastructure they own. The market had already started to read the signal; US startups were already beginning to choose Chinese models to host their weights. And then the Mythos rug pull occurred. One European outlet described the Anthropic shutdown as handing Mistral its sovereignty argument on a plate, and Mistral’s CEO sat down with the heads of Anthropic and OpenAI at the G7 days later from a position of unusual strength. The pitch that had sounded like nationalist marketing became the most concrete enterprise argument in AI almost overnight.

But what was waiting in the open-weight tier when buyers went looking? The strongest, cheapest, most accessible open models are Chinese. GLM-5.2 ships under an MIT license at 744-billion parameters. DeepSeek V4 and the Qwen family are frontier-adjacent, permissively licensed, and already the default for cost-conscious startups and much of the Global South. A buyer who wants a self-hostable model with no export-control exposure and frontier-class performance, today, is being routed toward Beijing by process of elimination.

The Trend

The pattern started in the startup tier, where the math is unforgiving and loyalty is cheap. Chamath Palihapitiya said he moved his company’s workflows off Amazon’s Bedrock to Moonshot’s Kimi K2 because it was more performant, and cost-conscious founders who could not justify frontier prices began defaulting to DeepSeek and Qwen well before any enterprise did. What looked like a fringe efficiency play among small, price-sensitive teams in early 2026 is the same behavior now surfacing at Shopify and Airbnb scale.
Then, a list circulated this week claiming a wave of Western companies had moved AI workloads to Chinese models. Lindy to DeepSeek, Cursor to Kimi, Shopify and Airbnb to Qwen, Siemens to both, Microsoft testing DeepSeek. It is directionally real and methodically loose, and it deserves to be read as a signal hardening toward evidence rather than a finished migration.

The verifiable entries are scoped deployments, not wholesale switches. Cursor’s in-house Composer models are built on Moonshot’s open-weight Kimi K2.5, confirmed after a developer found the model identifier in API responses. Shopify replaced one GPT-5 pipeline for merchant-data extraction with a self-hosted Qwen3 system and reported a 75-fold per-unit cost reduction, with most of the gain coming from system redesign as much as the model swap. Airbnb runs 13 models and leans on Qwen for its customer-service agent, which now resolves queries in seconds and has drawn a US House committee probe. The arrow notation flattens all of this into defection. The accurate version is narrower and still points one direction: when a cost-sensitive or sovereignty-sensitive workload comes up for a decision in mid-2026, a Chinese open-weight model is increasingly the answer.

That is the infrastructure thesis confirming itself. The mechanism is the one the March plan described, dependency built quietly at the infrastructure layer. What has changed is the accelerant. Beijing’s pull is real, but American policy is now doing the pushing.

The Pattern Behind the Policy

The United States keeps reaching for isolation as a demonstration of dominance. Chip controls, model-access restrictions, tariff walls, each is framed as leverage, a way to deny rivals and assert command of the stack. And each produces a version of the same result. The controls on advanced chips pushed Chinese labs into the efficiency breakthroughs that produced DeepSeek. The model-access suspension is pushing American and allied enterprises toward the Chinese open-weight ecosystem. The instrument meant to project strength keeps functioning as a referral.

A nation cannot wall off a supply chain it no longer monopolizes by restricting the parts it still controls. Restriction only teaches the market to route around the restrictor. The harder Washington squeezes access as a show of dominance, the more it manufactures demand for the alternative it was trying to marginalize.

The Meta Question

Which makes one absence genuinely strange. Llama is (aside from new Canadian/European entrant North by Cohere) the obvious Western open-weight candidate. It is permissively available, broadly capable, already embedded in enterprise pipelines, and produced by a company with effectively unlimited capital. If displaced demand were flowing to a Western open-weight option, Meta should be the first beneficiary. It is barely in the conversation.

Part of the answer is structural. Meta operates no first-party enterprise API by design. It distributes weights and lets the ecosystem host them, which means there is no Meta contract to migrate to, no commercial relationship to deepen, no account team capturing the shift. Llama adoption shows up as inference revenue on AWS and Azure, not as a Meta procurement win, so the company is structurally invisible in exactly the migration it is positioned to lead.

But structure does not fully explain it, and the gap is the interesting part. Why a firm with this much money, this much distribution, and a genuine open-weight head start is not the default sovereign fallback for Western enterprise is a question worth its own piece. The answer may say more about Meta’s strategic posture, its licensing terms, or its enterprise credibility than anything about the models themselves.

Where the Opening Sits

The vacancy is being contested by two vendors built for it. Cohere has shipped North Mini Code under Apache 2.0, runnable on a single H100, and Command A+, an open-weight model it frames explicitly against open-source development concentrating in, as it puts it, a small number of jurisdictions, namely China. Mistral has the Devstral and Small lines under Apache 2.0, EU-hosted, self-hostable, with HSBC and BNP Paribas already deployed and a sovereignty pitch that the Mythos suspension turned from slogan into sales engine.

There is a tension here the argument has to own. Sovereignty is doing two opposite jobs depending on whose weights are involved. Chinese open weights are the dependency to fear; Canadian and French open weights are the escape from it. The mechanism is identical, self-hosted models you control. What differs is the flag on the origin and the jurisdiction that can compel the maker. That is a coherent (as always, no pun intended) distinction for a Canadian or European buyer, and it is the entire commercial case for Cohere and Mistral. It is also the precise reason the substrate question is not settled. The West has credible sovereign open-weight options now. Whether they scale fast enough to catch the demand American policy keeps generating, before that demand sets as Chinese dependency, is the open question of the next year.

The divergence did not close. The cause moved, from a war Washington is fighting abroad to one it is waging on itself.

Original Post: The Great AI Divergence: Beijing is Building the Foundation While Washington Fights (Tariff) Wars

It’s one kind of uncertainty to look around and get the sense that the world is moving so quickly that it’s almost impossible to keep up with daily developments.

It’s completely another kind of uncertainty to realize that that disorientating environment is itself significantly behind the most advanced society in the world.

And yet, this is where we in the West find ourselves; divided, behind, and technologically and demographically disadvantaged. While American headlines are consumed by the latest circular financing round between a foundation model and a chip giant, or political disputes over Pentagon contracts, or violent attempts at imperialism, Beijing has published a definitive blueprint for the next decade of geopolitical and economic dominance.

This month’s release of China’s 15th Five-Year Plan (2026–2030) is not an incremental update. It is a structural declaration of intent that reframes the entire tech competition. It reveals that China has recognized the true nature of the “war” for technological sovereignty, while the United States is bogged down fighting a 20th-century tariff skirmish. For senior executives, the implications for supply chains, standard-setting, and future market access are profound.

The Signal vs. The Noise: A Study in “Nudgment”

The Five-Year Plan is a governing document through which Beijing aligns ministries, provinces, industrial capital, energy systems, manufacturing policy, and domestic technology champions around a shared economic objective. Reuters notes that the new plan sharply elevates AI, quantum technology, robotics, biomedicine, 6G, and related sectors as part of a broader effort to create “new quality productive forces.”  
 

To understand where the global economy is heading, executives must distinguish between noise and signal. Nudgment is the ability to understand the significance of emerging data signals (nudges) before they harden into evidence (judgment). AI is paradoxically the reason we need nudgment (proliferating data signals) and its strategic solution (advanced pattern recognition capabilities). In an increasingly probabilistic world, nudgment theory says we must become experts at data discernment and understanding which signals matter.

 * The US Noise: Currently, US (and Western) AI policy is reactive and decentralized. It is composed of reactive chip export controls, circular financing deals to prop up valuations, executive orders that are difficult to enforce, and a strategy often communicated through the vibe-based X posts of a few Silicon Valley founders.

 * The China Signal: Conversely, the 15th Five-Year Plan is a single, 141-page mandate. It is the definitive signal to every provincial governor and industrial conglomerate.

The numbers tell the story. Artificial intelligence is mentioned 52 times in this plan, compared to just 11 times in the 14th plan released in 2021. This isn’t an evolution; it’s a nearly fivefold escalation in strategic priority within a single planning cycle. The central theme has shifted from general technological innovation to building a modernized industrial system where innovation is specifically integrated into scalable, high-value production capacity. Beijing is gambling its entire economic future on AI. Or, alternatively, you could say it has seen the signal of the future and it is building it.

Sovereign AI: Foundation vs. App

The plan clarifies a critical concept that many Western businesses are failing to grasp: there are two fundamentally different models of AI sovereignty emerging.

 * The American “App Layer” Model: The US is currently winning the consumer-facing, “flashy” AI race. We have the best chatbots, the most viral image generators, and highly proprietary models (OpenAI, Anthropic) designed for American enterprise subscription services.

 * The Chinese “Substrate” Model: China has recognized that whoever controls the plumbing controls the system. They are not trying to compete for the American consumer market. They are building the global substrate.

The Chinese open-source strategy, spearheaded by models like DeepSeek and Alibaba’s Qwen, is a masterstroke of standard-setting. They are building free, high-quality open-source infrastructure that is becoming the default choice for the “Global South” and cost-conscious startups globally. If your entire digital economy runs on a Qwen-derived model, optimized for Chinese standard hardware, you have effectively “seceded” from the Silicon Valley stack. Beijing is creating a technological dependency loop that tariffs cannot break.

Decoupling Productivity from Population

The plan’s most aggressive measures are practical, not theatrical. The center-of-gravity is the “AI Plus” initiative, which sets a devastatingly clear target: integrating AI into 90% of China’s economy by 2030.

Read that again. 90%. Unimaginable, even seemingly undesirable (due to documented LLM flaws shortcomings) in the West. But let’s look at The Chengfan Port Ecosystem.

In March 2026, the Meishan terminal at Ningbo-Zhoushan officially shifted its container inspection workflow from a manual, human-centric process to a fully automated AI loop.

• The Agent Layer: Instead of customs officers walking the stacks with clipboards, they now deploy quadruped robot dogs (developed by Unitree). These aren’t just remote-controlled drones; they are edge-computing nodes running the Chengfan large model, a specialized LLM trained specifically on port logistics and customs regulations.

• The Substrate Integration: The robots navigate autonomously using 5G-Advanced (5G-A) networks. They verify container numbers, seal codes, and damage profiles in real-time. This isn’t just “vision AI”; it’s integrated into the National Integrated Computing Network, meaning the data collected by a robot dog in Ningbo can immediately optimize a logistics schedule for a factory in Chongqing.

• The Efficiency Gap: A task that previously required six personnel and over an hour of physical labor is now completed by a single autonomous unit in 20 minutes, with over 99% accuracy.

Compare this to the recent warnings across Amazon of AI error blast radius and the distinctions of advancement and competence could not be more clear. Is there an example of anything even remotely as sophisticated, at this scale, with this critical a function, anywhere in the West?

Integration at this scale is designed to, in part, solve a fatal demographic handicap also affecting the West: a shrinking labor force. But while we are planning deportations and “renovation”, the 15th plan envisions deploying robots to perform manual jobs in labor-scarce sectors and creating “hyper-scale” compute clusters powered by abundant, cheap electricity (a resource the US grid is struggling to provide). By integrating AI into logistics, energy distribution, and advanced manufacturing at a 90% penetration rate, China is aiming to decouple its GDP growth from its population decline.

Beijing’s strategic coherence is also visible beyond its borders. Starting May 1, 2026, China will implement zero-tariff treatment for imports from 53 African countries with which it has diplomatic relations, a move that expands Chinese market access across nearly the entire continent. Read alongside the 15th Five-Year Plan, the message is unmistakable: while Washington escalates trade friction, Beijing is building influence through market access, long-horizon alignment, and the steady construction of economic dependency.  

The Strategic Divergence is Not Just Technological

Beijing’s strategy is made up of models and industrial automation, *combined* with demographics and geopolitics. It is deepening commercial ties with the world’s youngest continent at the very moment its own population is ageing. That matters because who builds the best AI systems is less important than who develops a credible response to demographic decline. They are not unrelated. Reuters has reported that China’s current policy direction combines productivity gains from AI with expanded elderly care, pension support, and development of a “silver economy” as the country prepares for more than 400 million people over 60 by 2035. 

The contrast with much of the West is increasingly stark. Across OECD countries, 14.8% of people over 65 live in relative income poverty, even before accounting for the wider pressures of housing costs, care shortages, and fiscal strain. At the same time, Washington is intensifying deportation policy and parts of Europe are hardening immigration rules, despite the fact that ageing economies need labor, taxpayers, and younger households. China, whatever its many faults, is pursuing something recognizably strategic: trying to preserve output through AI, preserve dignity through eldercare and pensions, and extend its long-term economic runway through deeper integration with younger external markets.  

Fighting the Wrong War

The core fallacy of current US policy is believing that technology is merely another trade asset. The US is fighting a trade war, levying 145% duties on EVs, imposing trade restrictions, and restricting the flow of 2nm chips.

China looked at this and came to a different conclusion: the real war is for structural, productive capacity. Tariffs are a nuisance to be work-around; chip restrictions are a bottleneck to be solved through domestic development and creative engineering. However, an economy that has achieved 90% AI integration across all industrial sectors has undergone a structural transformation.

This transformation makes the tariff question irrelevant, because it fundamentally changes the productive capacity of the entire economy. A nation with the “best” chips mostly used for generating viral videos cannot compete with a nation that has “slower” chips that are actively optimized across 90% of its productive supply chain.

For Western boardrooms, the warning signal should be flashing. China has a five-year, coordinated industrial plan to capture the global AI substrate. The West has a fragmented ecosystem obsessed with “vibe-checks” and quarterly profits. If awareness does not develop as to the true nature of this “war,” it will discover it is no longer playing the same game.

The difference between what’s manifesting in China through the pages of the latest five-year plan and what we’re seeing in places like Iran and Venezuela and Minneapolis is coherence. There’s a coherent approach in China to everything from technology to ageing to population growth. The signals are devastatingly clear. Across the West no such coherence is in place; it much more closely resembles chaos, and given the longstanding dynamics and geopolitical tension, that is something we should all be watchful for. What is to be done is another matter entirely.

UPDATE July 1

The order that drove the whole episode is gone. On June 30 the Commerce Department lifted its export controls on Fable 5 and Mythos 5, about two weeks after imposing them, with Fable returning to global users and Mythos restored to a set of approved US organizations. Part of what forced the climbdown was the cost becoming visible, with executives and investors noting openly that freezing American frontier models handed time to the Chinese open-weight developers working to close the gap. Washington restricted its own most capable models as a show of control, watched the demand begin routing toward Beijing, and reversed course inside a fortnight.

What stands out on a second read is who was never in the blast radius. The directive hit Anthropic’s top tier, and OpenAI pulled its own cybersecurity model under similar pressure. Meta was not asked to suspend anything, agreed to no restriction, and never figured in the story at all. The likeliest reason is not that Meta was spared but that it had nothing operating at the level the controls were written for. The order targeted models capable of finding security vulnerabilities at scale, the capability that raised the national-security alarm, and Llama does not play there. So Meta sits outside the restriction for the same reason it sits outside the migration it looks positioned to win. The trait that would make it the obvious safe harbor, weights already downloaded and running on hardware no administration can reach, comes attached to a model that does not compete at the tier now defining the frontier. Being too far from the frontier to be restricted is also being too far from it to be chosen.

That last line is the turn: Meta’s safety from the ban and its absence from the demand are the same fact seen twice. Want me to compress this to a single paragraph for placement, or is the two-beat structure right? And confirm I read “the previous update” correctly as the Cohere/Mistral piece, since I referenced it by name.