Inside China's AI Labs with Interconnects & SAIL Media
Firsthand reflections from a visit to China’s most prominent AI labs, exploring the human side of the Chinese AI ecosystem, the technical constraints they face from chip supply, and how their research culture compares to the Bay Area.
A recent trip to China’s top AI labs reveals an ecosystem that is surprisingly similar to the Bay Area, yet shaped by unique local dynamics and technical constraints. The most significant takeaway was a “humanizing factor”—the researchers are relatable, humble, and deeply focused on the technical work, a stark contrast to the often politicized discourse in the West.
The Research Culture: A Familiar Scene
The AI research hub in Beijing feels remarkably like the Bay Area. Labs are often within walking distance of each other, fostering a fluid talent ecosystem where researchers might interview at every major company before choosing one based on “the vibes.” Tsinghua University serves a role analogous to Stanford or Berkeley, anchoring the community.
Similar to the US, different labs adopt varying research philosophies:
- Bottom-up: Researchers are encouraged to constantly explore new ideas, which are then escalated up the chain.
- Top-down: Leadership sets a clear research direction that the teams then execute.
Collaboration is common, with researchers across different labs maintaining friendships and sharing ideas, fostering a research landscape that has largely shifted from academia to well-resourced corporate labs.
Competitive Landscape and Market Dynamics
The domestic market is vibrant and insular. The entire industry is “obsessed with Doubao,” a chatbot app from ByteDance, which is viewed as a “fearsome economic competitor.” In contrast, DeepSeek holds a place of high esteem for its scientific contributions and cultural role.
There is significant domestic demand for AI services. Zhipu AI, for example, showcases a live dashboard with business metrics, at one point displaying that they serve 5.5 trillion tokens daily, with the vast majority of users being in China. This strong demand has led them to raise API prices multiple times. Interestingly, smaller startups like Zhipu AI seem almost proud of being placed on the US entity list, viewing it as a mark of global recognition that boosts their reputation.
Companies like Meituan (a delivery giant) and Ant Group are building their own solid, if not frontier, models. Their motivation is primarily to avoid spending millions on tokens from third-party APIs and to leverage their in-house talent. This mirrors the “monetize your compliments” strategy seen with companies like Meta, where releasing models openly is a natural starting point when direct API sales aren’t the primary business model.
Technical Realities: Chip Constraints and Preferred Tools
The impact of US export controls is a top-of-mind issue for researchers on the ground. They are “very chip constrained,” which clearly affects how they design and iterate on models, leading to longer pre-training runs. The universal preference is for NVIDIA chips, and the lack thereof is a significant bottleneck.
Despite nominal bans, the software toolchain is very similar to the West’s.
- Claude is heavily favored, with many researchers calling “Claude code” the best command-line interface (CLI) that exists.
- Hugging Face is the default platform for releasing and accessing models. Its network effects are so strong that its domestic clone, Model Scope, is only used when Hugging Face is inaccessible.
The Data Industry and Government’s Role
Contrary to some narratives, the data industry in China does not appear to be as central as it is in the US. There was little indication that labs were spending heavily on data from third-party vendors. Instead, many perform data creation and annotation in-house to ensure quality control, suggesting the “data industry” might be less of a moat than is often assumed.
The government’s role is also more nuanced than one of top-down control. It’s highly decentralized, with different municipalities within cities like Beijing competing to attract top AI companies. They might offer funding or make it easier to open offices. For instance, the Beijing Institute for Artificial Intelligence (BAAI) is directly funded by the local city government. This support acts as a signal of importance, but it does not extend to dictating research agendas or how models should behave.
A Purely Technical Focus
Perhaps the biggest cultural difference is the researchers’ focus. They are driven by engineering challenges and the desire to “build good models.” Questions about the broader economic impact, the labor market, or other political implications—common topics in the Bay Area—were consistently met with a pivot back to the technical work of scaling and improving their models.