Made in China
Introduction
Day by day, it is becoming abundantly clear that the AI models powering Western companies are no longer the ones made in the West. That reality has slipped in almost unnoticed. While the public conversation circles around frontier labs and safety debates, engineers inside North American companies are quietly building on Chinese open-source systems like Qwen, DeepSeek, Yi, and Kimi. They are doing it for one simple reason: these models deliver what they need at a price they can afford.
This shift is not driven by politics or ideology. It is driven by economics, performance, and the freedom that comes with open weights. And it raises a question that almost no one in the industry is willing to ask out loud: what happens when the foundation of the Western AI ecosystem is built on models the West neither owns nor controls?
Some Evidence
The strongest proof of this shift comes from a joint study by MIT and Hugging Face, which found that Chinese-developed open models accounted for 17.1% of global open-model downloads in the 2024-2025 period, surpassing the Western share of 15.8% for the first time ever. Most of this activity is driven by two families: Alibaba’s Qwen models and DeepSeek’s reasoning models, which together represent the bulk of China’s open-weight usage.
Two of the most widely used AI coding tools in the West, Cursor and Windsurf, both with millions of users, rely heavily on Chinese open-weight models under the hood. Cursor’s new “Composer” model, which the company describes only as “our first coding model,” has been identified by developers as a fine-tuned version of Zhipu’s GLM, likely GLM-4.6, after users noticed the model generating Chinese-language reasoning traces. The connection is further reinforced by Cursor’s built-in ability to run GLM models directly through a Zhipu AI (Z.ai) API key. In short, two of the most popular coding assistants in the West, both founded and based in North America and used daily by millions of developers, are powered by Chinese base models at their core. While I do find this slightly amusing, I also find a peculiar tension in relying so deeply on what cannot be remade at home. Stranger still to call it “ours.”
Despite this growing reliance, almost none of these systems can be recreated domestically. The West still leads at the frontier with closed models like GPT-5 and Claude, but it lacks open-weight pipelines that match the scale or accessibility of China’s releases. The data, training mixtures, and optimization methods behind Qwen, DeepSeek, Yi, and GLM are not reproducible, and few Western equivalents exist in open form. In practice, Western companies are building on foundations they could not rebuild if access vanished tomorrow. What looks like openness is, in reality, a dependency on models whose underlying parameters are controlled abroad.
Amazon Bedrock now offers Qwen3 and DeepSeek models as standard managed endpoints. Nvidia NIM distributes DeepSeek-R1 as an optimized reasoning container. Databricks provides first-class support for serving Qwen models within DBRX workflows.
Performance relative to cost is a major driver of this adoption. DeepSeek-R1 delivers reasoning results competitive with frontier Western models while requiring dramatically less training compute and offering up to 27x lower inference costs. Analysts across outlets such as the Financial Times and The New Yorker described R1’s release as having “shocked Silicon Valley” because it demonstrated that top-tier reasoning does not require frontier-scale compute or closed architectures.
Licensing further accelerates adoption. Most leading Chinese open-weight models, including Qwen, DeepSeek, Yi, GLM, and InternLM, use Apache-2.0 or MIT licenses, which allow unrestricted commercial use, modification, embedding into closed products, and redistribution without obligations. This stands in contrast to Western models, which more often rely on restrictive licenses or API-first access, adding friction for companies that want to self-host or customize.
China also leads in model production. As of mid-2025, Chinese organizations have released 1,509 of the 3,755 large language models published worldwide, the largest national share by a wide margin. As mentioned before, MIT and Hugging Face also report that China’s contribution to new open-weight uploads is rising in parallel with its download share, showing both its demand and supply leadership.
But how did this shift come about? Well, what is certain is that it was no accident. The real cause lies in what has been noticed and ignored, because the modern operator tends to asks nothing of the machine’s origin and everything of its output. And haven’t we grown used to taking performance at face value anyway?
Western Apache-Licensed Models
The West has produced strong Apache-licensed models as well. Meta’s Llama 3.1 family includes several fully open-weight variants under Apache-2.0. In vision and 3D modeling, Meta’s SAM 3D is Apache-2.0 and one of the most permissively licensed spatial-reasoning systems available. Smaller labs and independent researchers in the West continue to release Apache-licensed checkpoints in coding, embeddings, and multimodal tasks.
But while the West has produced a handful of high-quality Apache-licensed models, they do not match the frequency, breadth, or volume of releases coming from China. Western frontier work is still dominated by closed, API-restricted systems, whereas Chinese labs continue to publish open-weight models across every major category, including reasoning, coding, multimodal, embeddings, 3D vision, and diffusion. The release cadence also differs dramatically. Chinese groups such as Alibaba, DeepSeek, 01.AI, Zhipu, and Shanghai AI Lab push major updates every few weeks, while Western open-weight releases arrive far less often. The West has open models, but China has built a full open-source ecosystem.
However, the permissiveness of today’s licenses should not be mistaken for permanence. Apache-2.0 access is a choice, not a guarantee, and nothing obligates future generations of these models to remain open. If the next iterations adopt restrictive terms or move entirely behind domestic platforms, Western companies will have no leverage to secure continued access. The ecosystem is building on licenses that can change far faster than the infrastructure depending on them.
Western open-model work exists, yes, but it is arriving like a stream against a river in full flood. And a stream, no matter how clear its water, cannot redirect the force of a world moving in another direction. Once the river chooses its way, whatever it may be, the notion of steering will quickly dissolve and drift with it.
Our Perspective
Over the past month at Ventures Edge, we’ve experimented heavily with Qwen, Hunyuan, ByteDance’s open models, and Meta’s new SAM 3D for work involving Gaussian splatting, diffusion-based 3D modeling, diffusion workflows, and text-generation tasks. These models give us open access to the weights, strong performance, and rapid iteration cycles, which makes them ideal for experimental research. They were simply the easiest models to work with in practice, supported by far more community resources, far more shared experiments, and far more fine-tuned variants than anything available elsewhere. In many cases, choosing them felt less like a decision and more like the path the ecosystem had already cleared.
The rapid spread of open-weight models creates undeniable benefits, but it also introduces clear risks that policymakers and researchers have begun to highlight. Open-source models can be copied, modified, and deployed by anyone, which lowers the barrier for misuse in areas like fraud, disinformation, automated harassment, or unsafe content generation. Unlike API-based systems, self-hosted models operate with no logging, rate limits, oversight, or enforcement mechanisms, creating a regulatory blind spot. If a model is fine-tuned or repurposed for harm, there is often no way to detect it.
The safety gap is another concern. Open-weight models are frequently deployed without the moderation layers or guardrails that proprietary providers in the West are forced to build, which increases the likelihood of biased, toxic, or hallucinated outputs reaching end users. The ecosystem expands faster than it can be monitored.
There is also a supply-chain dimension. As more Western companies build on a small cluster of widely used open-weight models, a bug, vulnerability, or silent model change could cascade across many downstream systems. And because these models fall outside Western regulatory reach, policymakers have limited visibility into how they evolve or where their derivatives end up. These risks do not negate the value of open-source AI, but they underline an uncomfortable reality: unrestricted models accelerate innovation, and they also accelerate the potential for misuse, accountability gaps, and systemic dependencies that no single government currently has the tools to fully manage.
Our Question to the Reader
As you may have already read, our recent article on the AWS outage showed how the legs of the internet shake when its crutches snap. As AI settles into the foundation beneath those legs, the question grows sharper when the weights are built on a side of the world that is not our own. What happens when we dig into the ground beneath our software and find nothing familiar, nothing we can claim to understand, only the same small, ironic, and all-too-familiar inscription repeating in its quiet certainty: Made in China?