China’s open source wave

Chinese companies are now leaders in open weight models.

Alibaba’s Qwen family covers many sizes and use cases and is marketed as a one stop solution for developers and enterprises, with full open source style access to weights and tooling. You can see this directly in their GitHub repository, the Qwen 2.5 Max announcement, and an overview of the Qwen model evolution.

DeepSeek releases competitive models with weights that anyone can download, fine tune and run locally on Hugging Face. Recent versions like DeepSeek V3 and DeepSeek R1 aim at high level reasoning and are openly downloadable as weights, including DeepSeek V3, DeepSeek R1, and the Open R1 launch recap.

In other words, if I am a developer today and I want a strong free base model, it is very likely I will reach for a Chinese model.

Why I think China is doing this

Lock in developers and ecosystems

In my view, the first reason is simple. If millions of developers build on your models, you quietly set the standard.

Open weights make it easy to:

• Embed Qwen or DeepSeek inside products
• Reuse their tokenizers, tools and formats
• Host them on friendly cloud and hardware stacks

Alibaba explicitly positions Qwen as an open platform for developers on top of Alibaba Cloud and documents all the variants and use cases on its official resources, including technical documentation and the Qwen 2.5 Max blog post.

So I think the game is this. Give away the model. Make money on cloud, chips and enterprise projects around it.

Turn export controls into a software play

On the hardware side, the United States is trying to slow China down by restricting the export of advanced AI chips like Nvidia’s A100 and H100, and later even the cut down A800 and H800 variants. Policy notes and reporting describe how Nvidia had to design the H20, L20 and L2 chips specifically for the Chinese market under these constraints, including a Congressional Research Service brief and analysis from ECIPE.

In that context, open models are a kind of judo move. Even if China cannot always buy the very best chips, it can:

• Train strong models on what it has
• Release the weights to the world
• Let everyone run them on any hardware they want

In my opinion, this turns some of the hardware pressure into a software and standards game, where China has more room to maneuver.

What this means for developers and poor countries

From a developer point of view, I think the impact is huge.

If you live in a poor country and you cannot afford to pay for expensive US APIs, Chinese open models are very attractive. You can:

• Download them once
• Run them on your own GPUs or on a cheaper local cloud
• Fine tune them on your language and use case

Recent coverage and community projects around DeepSeek and Qwen show how they are becoming viable drop in replacements for many closed models, especially when cost matters more than squeezing out the last few benchmark points. Highlights include DeepSeek V3, the Open R1 announcement, and Qwen’s open resources.

In my opinion, this is soft power. If most universities, startups and governments in the Global South standardize on Chinese open models, China gets long term influence over tools, formats and even research directions.

Taiwan, the chip supply chain, and the West

Hardware reality

AI still runs on hardware. Most advanced GPUs are fabbed by TSMC, which builds Nvidia’s latest parts; some capacity is expanding to the US, but Taiwan remains the hub.

The hardware constraint remains: training and large-scale deployment depend on a few fabs, ship routes, and export rules.

Implications for the West

• Export controls: harder to enforce once capable open models spread.
• Competition: pressure to match “good‑enough” open models on quality, cost and openness.
• Response: compete in the open—invest in local models, safety, trusted governance—and keep building onshore capacity.

Europe and especially France

Europe leans on regulation and open source more than on chips. The EU AI Act sets strict rules, and many argue open source is central to AI sovereignty.

France is the concrete example: Mistral AI positions high‑performance open and commercial models as a European foundation. Europe’s play is to use open ecosystems to stay in the game, even if the cutting‑edge chips and largest models still come mainly from the US and Asia.

In my opinion, China’s open-model push is a strategic move in a larger game that mixes developers, poorer countries, export controls, and the chip supply chain.

If countries care about sovereignty and long term leverage, they cannot ignore open source. They either build their own open ecosystems, or they will end up living inside someone else’s.