From Hangzhou to your GP: Asian AI and the future

When you think of AI, you probably think of American names. But some of the most exciting work is happening in Asia, and much of it is free for anyone to use. In this article you'll read who these players are, what's going on with chips and memory, and why AI can make life better for billions of people.

DeepSeek: the newcomer that surprised everyone

DeepSeek is a young AI company from Hangzhou in China. It was founded in 2023 by Liang Wenfeng and grew out of High-Flyer, an investment fund he had set up earlier that invests using AI. In early 2025 it released a model that could reason at the level of the big American models. According to DeepSeek itself, training cost only a fraction of what its competitors spend. Experts are still debating exactly how that adds up, but the shock was huge. On a single day, 27 January 2025, chipmaker Nvidia's market value fell by almost 600 billion dollars.

What makes DeepSeek special isn't just the quality, but above all the efficiency. The company showed that clever techniques can get you a long way without the most expensive computers.

Alibaba and Qwen: a whole family of models

Many people know Alibaba from the online shop AliExpress. The company also makes a series of AI models called Qwen. They come in all sizes: from small enough for a laptop to big enough for heavy tasks. Developers around the world build their own applications with them. According to Hugging Face, the largest website for downloading open AI models, Qwen is in 2026 even the most downloaded family of open models there.

What does open source mean here?

In software, open source means the source code is free to view and use. With AI it's a bit different. Usually companies release the "weights". Those are the billions of numbers in which the model has stored what it has learned. That's why experts often prefer to speak of "open weights". You can download the model and use it yourself, but you don't always know exactly how and with what it was trained.

The upside: anyone can use it. A business can run such a model on its own computer, so customer data never leaves the building. Researchers can build on it, and small businesses don't have to depend on one big supplier.

The downside: you don't always know what data the model was trained on. Some Chinese models dodge politically sensitive topics. And once something is freely available online, anyone can use it, including people with less good intentions.

The exciting part: the best open models are lagging less and less behind the best closed models. Smart AI is becoming something the whole world can use, not just a handful of companies.

AI on your phone: future talk or already real?

A bit of both. Small AI models already run on phones. Apple and Google put one on their newer devices themselves. They help you summarise a message, improve a text or search your photos, for example, without anything going to the internet. Small open models can also be put on a good phone or laptop these days.

But the honest story is that the very smartest models are too big for a phone. They need far more memory and computing power than fits in your pocket. What's happening now is that small models are improving at a rapid pace. What can only be done in a data centre today might fit in your phone in a few years. Exactly how fast that will go, nobody knows.

Memory chips: are we stuck?

AI doesn't just run on processing chips. It also needs a great deal of fast memory, to hold the model and all the calculations while it works. The best AI chips use a special, stacked type of memory. Almost all of it comes from three companies: SK Hynix and Samsung in South Korea and Micron in the United States.

Because everyone is building AI data centres at the same time, demand for that memory is enormous. Factories are switching to AI memory, and as a result ordinary computer memory has also become scarcer and more expensive. Since October 2025 prices have been rising sharply: ordinary PC memory became about three times as expensive within a few months. And it isn't over yet. Market researcher TrendForce expects memory to become another 10 to 20 percent more expensive by the end of 2026, although the increase is slowing compared with earlier this year. The shortage of AI memory is expected to last until 2028. Laptop and phone makers are feeling it too. It isn't a real blockage, but it does slow things down and it costs more.

Why does Nvidia build almost all the best AI chips?

Nvidia originally made chips for games. They happened to be very good at exactly the kind of computing AI needs: lots of small sums at the same time. On top of that, back in 2006, twenty years ago, Nvidia built software that lets programmers use those chips easily. As a result, almost all AI software works with Nvidia by default. Switching to another brand is a lot of work, and that is Nvidia's biggest advantage.

There's a Dutch part to this story too. Nvidia has its chips made by TSMC in Taiwan. TSMC can only make those chips with the advanced machines of ASML from Veldhoven in the Netherlands. No Dutch technology, no AI chips.

How far along is China?

The United States bans the sale of the newest AI chips and the most modern chipmaking machines to China. Since early 2026, an older Nvidia model, the H200, may be sold again, but only with a licence and under strict conditions. China is also building its own chips, mainly through Huawei, increasingly in its own factories. They are usable and getting better. In the first half of 2026, Huawei already sold almost as many AI chips in China as Nvidia. But in terms of power they still lag behind: according to researchers, the best Chinese chips resemble what Nvidia made about five years ago, and Chinese chip factories are three to five years behind TSMC. The biggest obstacle is that China has not been allowed to buy ASML's most advanced machines since 2019. China is working on its own version, but it is far from ready for use in factories. China partly makes up for this with clever software, as DeepSeek showed.

The best part: what AI can mean for billions of people

When it comes to AI, many people think of chatbots and clever texts. But the biggest gain may lie in health.

Finding new medicines faster. Developing medicines often takes more than ten years. With AI, researchers can predict much faster which substances are promising. A well-known example is AlphaFold, which can predict the shape of proteins. It earned two of its creators, Demis Hassabis and John Jumper, the 2024 Nobel Prize in Chemistry.

A first look at a spot. Say you notice an odd spot on your skin. You take a photo, and the AI says: "I'm not a doctor, but this is something to show your GP." Or instead: "This looks harmless, but keep an eye on it." AI doesn't replace the doctor, but it helps people see a doctor in time.

Care where there is no doctor. In the Netherlands, you can see a GP easily. There is roughly one doctor for every 260 inhabitants. In Uganda it's one doctor for more than 5,000 people, and in Niger even one for about 26,000. In rural areas it's often much less still. With a simple phone or a terminal in the village and an AI assistant, people can ask basic questions. What should I do about a fever? Is this wound infected? When should I go to a clinic? A nurse with AI at her side can help far more people than without it.

This is exactly where open models make the difference. If good AI can be downloaded for free, countries and hospitals without big budgets can get started with it too, in their own language and on their own devices.

Finally

The future of AI is being made not only in Silicon Valley, but also in Hangzhou, South Korea, Taiwan and Veldhoven. There are real questions: about chips, about fairness, about who uses which models. But the direction is hopeful. Smart help is coming within reach of more and more people, and that can make everyone better off.

As of October 2026.