The apps you have probably never opened
There is a quiet gap between what Americans think is happening in AI and what is actually happening. In the US, the names are ChatGPT, Claude, Gemini, and Midjourney. In much of the rest of the world, a different set of apps leads. Most of them come from China.
They are enormous. They are polished. And almost nobody in a US coffee shop has heard of them. That gap is worth closing, because it tells you where the competitive line is moving.
The reason the gap exists is boring, not sinister. Language comes first. Most of these apps assume a Chinese-language user, and the English builds that exist are often afterthoughts. App store policy comes second. Neither point is a verdict on quality.
Start with the numbers, because they surprise people. ByteDance’s Doubao reported roughly 168 million monthly active users at the start of 2026. That put it second in the world behind ChatGPT. Its overseas version, Dola, passed 70 million downloads in a single quarter and briefly became the most downloaded AI app in Southeast Asia.
These are not niche tools. They are the daily default for more people than the entire US population.
That is what a market that size buys you. A product can be the default for hundreds of millions of people and still be invisible in Ohio. Distribution decides which apps an American hears about, and distribution points somewhere else.
Kimi is worth a closer look. Its pitch is simple. Feed it a book or a year of notes. It keeps all of it in view. That is a different job than chatting, and it does that job well.
DeepSeek, the lab that I wrote about as a price disruptor, sits in the global top five of web AI tools. It pulls a large share of its users from outside China, including US companies that pay for direct access. These are not hobby projects.
What each one is for
Doubao is the general assistant, the ChatGPT equivalent, tuned for everyday Chinese-speaking users. DeepSeek is the cheap, open, capable model that businesses reach for when US pricing feels like a tax. Kimi is the long-context reader, built for working through very large documents. Each one owns a lane.
Kling AI, from Kuaishou, generates video and pulled most of its revenue from outside China. Even CapCut, which many Americans use without realizing its origin, is a ByteDance product. Qwen, from Alibaba, is the open-weight model most developers outside China actually download. The list is long, and it keeps getting longer.
Kling sells to creators outside China more than it does at home. That is the tell. The product is good enough to cross a language barrier alone. CapCut already proved it.
There is a bundling effect behind some of this that Western labs cannot copy quickly. A new model in China can reach a phone maker, a cloud platform, and an app store that already share a corporate parent. In the US the same distribution deal would take a year of negotiation.
None of these is a knockoff. They are different products aimed at different habits. Several are genuinely ahead on the one dimension they care about. The mistake is assuming that an unfamiliar name means a worse product. In AI, unfamiliar usually means not marketed to you yet.
What does that mean for people who build software? Mostly this. Watch the open weights.
A model you download and run yourself is not a locked door. It is a shortcut past the pricing maze. And it costs you nothing to test one.
Why should that matter to someone who is happy with Claude? Two reasons. Competition is already pushing prices down everywhere, and the models with the most users collect the most real-world usage data. Both effects reach you even if you never open a Chinese app.
Why you still might not use them
My honest take is that the US and China are drifting into two separate AI worlds, and most users only ever see one. That is not a political statement. It is a practical one. The app you reach for depends on where you live and which app store you trust.
Availability is uneven. DeepSeek’s web app works broadly and is easy to try. CapCut is already on most phones. Others are geo-limited, or simply not tuned for an English-speaking user, because the model behind them assumes a Chinese-language context first. The door is open, and not for everyone.
The friction is worth naming, because it is smaller than people assume. Sign-up flows often expect a phone number from the right country. Support pages are in Chinese. Billing works best with a local payment method. None of that is a wall, and all of it costs you an evening.
So what would actually change the picture? Something boring. One English-first build with a clean sign-up would move a lot of users, because very little of this gap is still about capability.
There is a fair question underneath all of this too. Where does your data go, and who can read it? Every AI company answers that question badly. The ones outside your legal jurisdiction answer it worst.
So the honest conclusion is not that you should abandon your current tools. It is that the field is wider than the headlines suggest. The gap is not in quality. It is in attention.
If you want a practical test, try one. DeepSeek’s web app needs no account trickery, and it answers coding questions as well as anything built in the US. CapCut is probably already on your phone. The point is not loyalty to a flag. It is awareness.
Watch the open-weight side especially. Qwen ships weights anyone can download and run, and that is not a small thing. Meta’s Llama spent two years as the default choice for that crowd. It is not the default any more.
For a developer, that shift is concrete rather than abstract. The model you fine-tune next month may well have been trained in Hangzhou, and the licence attached to it may be more permissive than anything Meta ships.
But be careful with the assumption sitting under most of the coverage. The leading American model is not the leading world model by default. It is the model with the best English marketing, which is a different thing entirely.
A builder who only watches Silicon Valley is watching roughly half the game. If you remember one thing, remember this. The next time someone tells you the AI race is already decided, open the charts for the regions with the most users. The answer there rarely matches the one you hear at home.
The apps that win are built by companies most Americans could not name. And they are very good. The company that takes the next two years of adoption may not be the one with the best benchmark scores, but the one already sitting on the phone in your pocket.
None of this is settled. The next year will decide most of it. What will not change is where the users already are. That number is larger than any English-language launch has managed.
Start with one app. Try it for a week. Then decide what you think.
