The Moment Everyone Started Paying Attention
In early 2025, a relatively unknown Chinese AI lab called DeepSeek released a model that performed roughly on par with GPT-4 on several benchmarks — and claimed it cost under $6 million to train. For context, the leading Western models at the time had training costs estimated in the hundreds of millions.
The reaction in Silicon Valley was predictable: skepticism, then alarm, then a grudging acknowledgment that the numbers seemed to check out. DeepSeek had used a mixture-of-experts architecture and some clever training optimizations to achieve dramatically better efficiency. They didn't have more compute than OpenAI or Google. They just used it better.
I'm not going to pretend I've run DeepSeek's benchmarks myself. What I can tell you is that I've used DeepSeek-V3 and its successor models for coding and writing tasks over the past few months, and the gap between it and the Western frontier models is smaller than most people in the US assume. On some tasks — particularly math and code generation — it's genuinely competitive.
What Makes DeepSeek Different
A few things set DeepSeek apart from the Western AI labs:
First, the cost efficiency story is real. DeepSeek's team published detailed technical papers explaining how they achieved their training efficiency, using a combination of mixture-of-experts routing, multi-token prediction, and aggressive memory optimization. The papers are worth reading even if you're not an ML engineer — they demonstrate an approach of working within constraints that's fundamentally different from the "throw more GPUs at it" philosophy that dominates in the US.
Second, DeepSeek's models are mostly open-weight. You can download them, run them locally, fine-tune them on your own data. This matters enormously for the developer ecosystem. While OpenAI and Anthropic keep their models behind APIs, DeepSeek is betting that making models freely available will create an ecosystem that's harder to compete with than any single proprietary model.
Third, the rate of improvement has been startling. The jump from DeepSeek-V2 to V3 was significant. The jump from V3 to their latest models even more so. They're iterating faster than anyone expected.
Then There's Doubao
If you haven't heard of Doubao, that's partly the point. Developed by ByteDance (the company behind TikTok), Doubao has quietly become the most widely used AI assistant in China, surpassing Baidu's Ernie Bot in monthly active users in mid-2025.
Doubao's strategy is fundamentally different from DeepSeek's. DeepSeek is a research lab that happens to offer consumer products. Doubao is a consumer product built by the world's most sophisticated attention-capturing company. ByteDance understands distribution, user engagement, and product-market fit in ways that most AI companies don't.
The app itself is less technically ambitious than DeepSeek's models — it's not pushing the frontier of benchmark performance. But it's fast, it's free, it integrates well with ByteDance's ecosystem of apps, and it's designed for how normal people actually use AI: asking questions, generating social media content, helping with homework, planning trips. It does these things well enough, and for hundreds of millions of users, that's what matters.
The Geopolitics Nobody Wants to Talk About
I can't write about Chinese AI without mentioning the elephant in the room: export controls.
The US has restricted exports of advanced AI chips to China since 2022, with the controls tightening several times since then. The explicit goal is to slow China's AI progress. The implicit bet is that by denying access to the most advanced hardware, the US can maintain a lead in AI capabilities.
How's that working out? Mixed, honestly.
On one hand, the export controls have definitely made things harder for Chinese AI companies. They can't buy H100s or B200s in volume. They're working with older hardware, or with domestic alternatives like Huawei's Ascend chips that don't match NVIDIA's performance. This is a real constraint.
On the other hand, constraints drive innovation. DeepSeek's efficiency breakthroughs didn't happen despite the chip restrictions — they happened because of them. When you can't solve a problem by throwing more hardware at it, you have to think differently.
The longer-term question is whether export controls slow China down or just push them toward different approaches that eventually prove more robust. I don't think anyone knows the answer yet.
What This Means If You're Not In the AI Industry
You're going to have more choices. A year ago, if you wanted a state-of-the-art AI assistant, your options were basically ChatGPT, Claude, or Gemini. Now DeepSeek's models are freely available. Doubao is dominant in the world's largest market and will eventually expand. More competition means lower prices, better products, and less dependence on any single company.
It also means the AI story isn't just a Silicon Valley story anymore. The most interesting developments in the next few years might not come from San Francisco or London. They might come from Hangzhou or Beijing.
Using DeepSeek From Outside China
If you're in the US or Europe, you can access DeepSeek's models through their website (chat.deepseek.com), their API, or by downloading the open-weight models and running them yourself. The web interface is free and functional, though noticeably less polished than ChatGPT or Claude — it feels more like a research tool than a consumer product.
The API pricing is aggressive. DeepSeek charges a fraction of what OpenAI charges for comparable capability. For developers building on top of LLMs, this is genuinely compelling. A project that costs $500/month on GPT-4's API might cost $80/month on DeepSeek's. The quality difference is small enough that for many use cases, the cost savings are worth it.
There are caveats. The content moderation is different from what Western users expect — DeepSeek's models refuse to answer questions about politically sensitive topics in China, and the boundaries around what's "sensitive" are broader than you might assume. The English output is excellent but occasionally has small awkwardness that a native speaker would catch. Documentation is mostly in Chinese, though English versions are improving.
Doubao is harder to access from outside China — it's primarily designed for the domestic market and the app is in Chinese. If you read Chinese or are willing to use translation tools, it's worth trying just to understand what hundreds of millions of users are experiencing. If you don't, it's more of a phenomenon to be aware of than a tool you'll use directly.
The Bottom Line
I'm not making predictions about who "wins." The idea that AI development is a zero-sum race with a single winner is wrong. Different models will excel at different things. Different companies will dominate in different markets. The more interesting question isn't "who's ahead" — it's what everyone builds with the technology that's now widely available.
The AI story of 2026 isn't "China vs. America." It's "intelligence is getting cheap, fast, and everywhere." DeepSeek and Doubao are accelerating that trend from a different direction, with different constraints and different strengths. The result is more options, lower prices, and faster progress for everyone — regardless of which country you live in.
