We Spent 10 Days Touring Chinese AI Labs. Here’s What We Saw.
Sleeping Cots, Robot Pharmacies, and the Race for AGI
Lily Ottinger is an editor and researcher for ChinaTalk, and Kai Williams writes for Understanding AI.
From late April to early May, SAIL writers spent 10 days touring Chinese AI and robotics labs. We got off the plane at the Beijing airport, to be welcomed on the jet bridge by 30-foot signs that said, “Alibaba Cloud robustly supports Chinese companies going overseas.” AGI was in the air, and we saw plenty of AI product advertisements across the four Chinese cities we visited (Beijing, Hangzhou, Shanghai, Shenzhen). I, Lily Ottinger, was one of the few fluent Mandarin speakers in the group — I’ve been to China many times before and lived in Taiwan for almost four years prior to this trip.
Despite the fact that the SAIL writers were all in the same conference rooms, we all came away from this trip with different perspectives, in part because of our language ability and previous interactions with Chinese culture. Today, Kai Williams and I would like to give you a taste of our experiences and some specific color on the companies that so generously hosted us (including Unitree, MiniMax, Moonshot, Z.ai, and more).
Disclaimer: the offices we visited are not public, and getting all these AI tigers to agree to show us around took 6+ months of planning by the fabulous Caithrin Rintoul, who leveraged his personal connections and charisma to make this possible.
Compute constraints
Basically every AI researcher we talked to in China brought up the same complaint: their companies lack sufficient access to the advanced compute resources necessary to train and run AI models.
This compute constraint is one of the key principles of the Chinese AI ecosystem. Like in the US, Chinese companies have access to top AI talent and copious amounts of training data. Without enough compute, however, it is difficult or impossible for Chinese companies to make cutting-edge models.
And the compute gap between the US and China is only going to grow. US AI companies are only just starting to integrate Nvidia’s Blackwell chips, which are significantly stronger than any chips Chinese companies can legally obtain. While Nvidia CEO Jensen Huang has claimed that Chinese companies already have sufficient compute, none of the Chinese researchers we talked with seriously believed this talking point.
Many of the Chinese companies we talked with — especially AI-focused labs like Moonshot, MiniMax, and Z.ai — still believe in a vision of AGI. But what does it mean to chase AGI in a compute-constrained environment, especially when many Chinese researchers believe that the gap with American AI labs is growing instead of shrinking?
One approach to this conundrum is to work very hard. I (Lily) got the sense that AGI is like a religion for these researchers, where — as is also the case in the US — devotees of the Machine God pray at the altar by working overtime even when the marginal hour of labor is less impactful than the marginal GPU would be.
This led to some truly brutal revelations by the researchers we spoke with — we toured labs during the week-long Labor Day holiday and on weekends, yet around a quarter of the researchers were still grinding away at these offices, at standing and sitting desks alike, with caffeinated and sugared beverages peppered across every horizontal surface like solo cups at a frat party.
One researcher we spoke to had a persistent rash on his chest, and said he was trying for a baby with his wife until DeepSeek R1 came out last January and forced him into constant overtime. During a bathroom break at one lab, I (Lily) rounded the wrong corner and found a bunch of sleeping cots that were clearly getting a lot of use. The lobby of that same lab housed a small mountain of personal packages that employees had gotten delivered to their office rather than their home. Chinese companies do have a lunchtime nap culture, but most of the researchers we talked to were very forthright about working a lot of overtime.
But hard work isn’t the only way that Chinese companies respond to compute and resource constraints. Chinese foundation model companies have also put effort into productization earlier than their Western counterparts.
Many Chinese companies offer coding plans, often with generous token limits to attract customers away from Claude Code, but companies have also pursued more niche product directions. MiniMax builds highly lucrative AI companion products, while Z.ai has sought out B2B partnerships with social media companies and Comac, China’s commercial aircraft manufacturer, as well as contracts with government entities. Ant Group, an Alibaba affiliate, has carved out a niche in AI-based healthcare delivery, which would be a liability and patient privacy nightmare in the US.
That being said, the labs we talked to were understandably more excited to talk about their research and coding agent products. MiniMax’s presentation to us only briefly mentioned their companionship business (which is their largest source of revenue).
Overall, the competition is intense. These labs will have to weather successive consolidation crunches (as has been seen in other Chinese industries, like EVs) to survive long enough to build AGI. I came away with a different feeling than Nathan Lambert did — there’s definitely a sense of rivalry between Chinese AI labs. When we talked with MiniMax, one employee commented that compared to Z.ai, MiniMax has “half the headcount, 60% the cash burn, and two more modalities.”
The core competency is making money
Like the US, China has plenty of companies which focus on AI model development specifically, like Z.ai, Moonshot, and MiniMax. But China also has a robust ecosystem of tech giants who have decided to build their own foundation models, as Meta and Google do.
For instance, the food-delivery app Meituan is building an LLM to use on their own platform because they have cash to burn and would prefer not to use a model they don’t own. And the phonemaker/appliancemaker/carmaker Xiaomi has an LLM team to build models for integration into their phones/appliances/cars.
This reflects a broader pattern as well. Flush with state capital and engineering talent, China’s tech giants are happy to fall in line and pursue new product areas when Beijing declares them to be an industrial policy priority. (This also explains why China’s most famous liquor brand has dipped its toes into the semiconductor business.)
In the US, the specialized model makers seem like they are ahead. Anthropic and OpenAI make the most revenue, and they tend to focus on their “core competency” of foundational model innovation as opposed to making wrappers. But the philosophy of Chinese companies, as noted by Dan Wang, is that making money is their core competency.
In China, it is less clear whether specialized model-makers can survive — hence the incentive to serve up products in addition to foundational models. While some of the general tech companies making LLMs don’t seem like they have a particularly strong vision for the trajectory of the technology, the large tech companies have very important advantages over less-established companies.
One key reason is distribution. Compared to the US, the Chinese tech ecosystem routes disproportionately through large consumer apps like WeChat and Douyin (the domestic version of TikTok). This makes it difficult for new companies to gain a large market share as the large companies can restrict advertising and block upstarts.
For instance, Z.ai made a model which can control a smartphone to do tasks on behalf of the user. However, several of the major Chinese apps like WeChat blocked Z.ai’s model from taking screenshots in their app, basically crippling its ability to be useful.
The shadow of ByteDance, TikTok’s parent company known for their “Wolf-like” corporate culture, loomed over our visits. ByteDance’s LLM is called Doubao, and it’s the most popular chatbot in China by a wide margin with almost 350 million monthly active users. It’s trained to speak in different Chinese dialects, which requires a strong voice model because Chinese dialects other than Mandarin and Cantonese are largely not written (by comparison, GPT’s voice mode is still powered by GPT-4o). ByteDance’s strategy has been to optimize for engagement over raw intelligence. Many Doubao users are older people using it as a search engine in the wake of the persistent enshittification of Baidu, China’s biggest search engine. Several researchers we spoke with at other labs lamented being asked by older relatives how their work was any different from Doubao. Doubao recently introduced paid subscription tiers, which has brought unusual backlash from users who allege that Doubao is “too dumb to pay for.”
ByteDance is confident that they’ll be the champions of the war between China’s dozens of foundation model companies. Caithrin relayed an anecdote from dinner he went to. The Kimi researchers upon arrival: humble. The DeepSeek researchers upon arrival: humble. The ByteDance researchers, conversely, arrived with their chests puffed out and cigs hanging from their lips.
As another model maker put it: “We are all midsized compared to ByteDance.”
Unitree and Galbot: two sides of the robot stack
Robotics is where China’s AI ambitions meet physical reality. Before China’s AI+ plan, Xi Jinping seemed to believe that only hardware counted as real technology. That has led many foundation model makers (like Z.ai and Ant) to have perfunctory robotics wings. A huge mass of undifferentiated robotics companies have raised hundreds of millions of dollars purely because their founders had degrees from Tsinghua University.
An exception to this trend is Unitree — the national frontrunner in robotics hardware that has had their bots featured in China’s Spring Festival Gala (the most-viewed television program on Earth) for the last two years. Unitree is based in Hangzhou (the same city that birthed Alibaba and DeepSeek), and its founder was not educated overseas or in one of China’s top universities. Their rise was predicated almost entirely on hardware innovation — which lets them build useful hardware very cheaply and at a profit.
Like the AI model companies, Unitree researchers are working overtime. But we didn’t detect the same AGI-fueled fire in their bellies. And that’s fine — Unitree needs a definition of AGI like a fish needs a bicycle. Another robotics company, Galbot, took us to a warehouse where a robot autonomously picked cold medicine and contact lenses from the shelves and packaged them together according to customers’ orders submitted via app. As we walked around the facility, delivery drivers were coming and going to pick up orders packed by the robot. Galbot told us they delivered over one million orders in 2025, with 20% of those orders coming in overnight when traditional pharmacies are closed. None of those orders required frontier models or superhuman intelligence — China’s robotics ecosystem is taking advantage of AI right now and making money despite compute constraints. As Afra Wang noted, there are tons of industrial applications for AI+robotics that will be invisible to US companies because the US has lost its industrial base. Those innovations, it seems, are destined to come from China instead.
Unitree’s revenue used to be dominated by sales to academic research labs, but as their IPO prospectus revealed, that is becoming less and less true as commercial sales have ballooned. Some of these customers are industrial, but as Unitree and Galbot pointed out to us, the plurality of commercial buyers are purchasing robots for entertainment applications. It’s surprisingly satisfying to watch a robot dance — or struggle to make coffee. Galbot told us that it had opened over 100 stores last year where robots make beverages. Apparently, the novelty sells.
We began our Unitree visit by rolling up to the wrong place. This old office was gritty and unassuming, and served as Unitree’s primary facility up until about a year ago. Today, it houses R&D and order fulfillment operations.
Directions corrected, we showed up at the new office where all the shiny stuff happens. We were welcomed at the door by a humanoid greeter.
Unitree walked us through several demos of humanoids and quadrupeds. All of the text descriptions were in Chinese and google-translate-level German because of a recent visit by a German delegation, which also led Unitree to display several Siemens appliances in their kitchen demonstration environment. We attempted to break through the corporate SOP by asking if we could ride on the quadrupeds.
ChinaTalk’s Lily Ottinger takes a ride on a robot dog. Presumably, the German delegation did not also ask for rides. (Filmed by Jasmine Sun)
The most exciting Unitree demo was the robot boxing match. When one of the bots got its leg stuck in the ropes, human staff intervened to get it untangled and then laid it flat on the floor so we could see it stand up by itself.
Labubu is AGI-pilled?
In my mind, the best piece of merch we received from any lab was the Kimi Labubu, but in fact, these would not be the only Labubus bestowed upon us during the trip. MiniMax also gifted us a box of Labubus (though without branded outfits), noting that the founder of Popmart (the company behind Labubu) and MiniMax co-founder Yan Junjie both hail from Henan province. MiniMax also told us that Popmart was using their models to make animated Labubu advertisements. I can’t imagine that it would go down well in the US if a toymaker were using AI-generated video instead of human animators.
Similarly, at the Huaqiangbei electronics market in Shenzhen, many sellers were using promotional images that were clearly AI-generated (by models at least a couple of generations old). What this communicated to me was that these sellers were early adopters of AI products.
Vibe dump
Some employees respond to the intense overtime requirements by buying nice things — at one dinner, I sat between a DeepSeek researcher and a Z.ai researcher, who each had a brand new Huawei folding phone that retailed for 10,000 RMB (~US$1,470). One of our hosts came to a casual lunch wearing a Van Cleef & Arpels bracelet. Other researchers respond with camaraderie: “AGI is the friends we make along the way,” as one Moonshot researcher put it.
Huaqiangbei
The Shenzhen electronics market is pretty deeply integrated into the supply chains of the war in Ukraine. Sellers posted advertisements in Russian boasting the infrared sensing capabilities of their drones, and groups of injured Slavic veterans perused these stalls with intrigue.
Xiaomi
Xiaomi was the only lab we visited that had exclusively squat toilets in the bathroom, so clearly someone at the company believes strongly in them.
All three Xiaomi researchers we spoke with were PhD students; apparently 80% of their LLM team are currently PhD students. The average age is 25. Later, a couple of PhD students told me that because Chinese PhD stipends are much lower than in the US, students have much more pressure to take internships, and are encouraged to do so by their advisors.
All of the Xiaomi researchers we spoke with were dressed in black, and we met in a slick black conference room.
Modelscope:
Open source is Chinese industrial policy, but it’s being implemented in style at ModelScope, which is going to be China’s HuggingFace equivalent. With bean bag chairs and a hip events board, it wins the award for the coolest office space we saw on this trip.
Z·Pilot
Z·Pilot is an electronics retailer based in Shenzhen. Here are some of the products we saw at their store.
 | A guest post by
|
 | A guest post by
|