Artificial Intelligencer: Funding frenzy, chip wars, and robot dreams: Inside China‘s AI boom
ARTIFICIAL INTELLIGENCER-Funding frenzy, chip wars, and robot dreams: Inside China‘s AI boom By Krystal Hu
Sept 23 (Reuters) - OpenAI is continuing to attack Anthropic's home turf by trying to capture the most valuable use case in AI: coding.
Airbnb ABNB.O is expanding its engineering and product teams’ access to GPT-6 Astra, OpenAI’s flagship model, CTO Ahmad Al-Dahle told me, as early testing shows how powerful the model can be.
Al-Dahle said the company will be adding it to the other models on offer for its engineers, encouraging them to use whichever model works best for the job. As he put it: "There isn't a sort of attachment to one model versus another. There's only what works better."
Airbnb’s move is an example of how a new powerful model could help OpenAI level the playing field against Anthropic's Claude Code, which has captured much of the coding market so far.
Airbnb also uses and post-trains open-source models for specific use cases, and says they can outperform frontier models in those cases.
Many of the best open-source models, notably, come from China.
Ahead of President Trump's meeting with President Xi on Thursday, I called up the two reporters who probably know China's AI scene better than anyone — our Beijing colleagues, Eddie and Laurie. How did China get so good at open source, so fast? And do people there feel the same AI anxiety Americans do?
Scroll down for my Q&A with Eddie and Laurie — it'll get you up to speed on China's AI development in just a few minutes.
Secure your front row seat for exclusive Reuters AI conversations–sign up here to watch my colleagues’ live interview with US FTC Chairman Andrew Ferguson, and Joe Lonsdale, Founder of 8VC and Co-Founder of Palantir Technologies at Reuters Momentum AI this Friday.
OUR LATEST REPORTING IN TECH AND AI
Ten days that changed the course of AI
In Silicon Valley, AI data center boom meets local resistance
AI is changing how Americans choose snacks
Meta testing a 'human concierge' for its new personal AI agent, Muse
Ex-Google safety chief warns AI could harm children more than social media did
Three out of four Americans say AI firms not doing enough to prevent disaster
FUNDING FRENZY, CHIP WARS AND ROBOT DREAMS IN CHINA
Eduardo Baptista and Laurie Chen are Reuters correspondents based in Beijing, where they cover China's technology and political landscape up close. They both speak fluent Mandarin and have a track record of breaking scoops by tracking China's AI boom from the inside. Our conversation has been edited for length and clarity.
Q: Why is there so much AI fundraising activity in China?
Eddie: It all comes down to compute. Over the past year, Chinese labs realized the same thing US labs did: if you want to compete at the frontier, you can’t avoid spending billions on advanced chips and data centers.
Take DeepSeek. It used to pride itself on not raising external capital, relying instead on Liang Wenfeng’s quant hedge fund High-Flyer. That’s no longer realistic. The next generation of models is so compute‑hungry that even DeepSeek has gone on a fundraising push and is now preparing for a potential IPO.
You see the same pattern with MiniMax, Z.ai, Moonshot, and others looking at Hong Kong listings or secondary share sales. All of this is essentially for raising money to buy compute.
Q: How much of this is funded by Beijing versus private capital?
Laurie: It’s a genuine mix of three: private capital and VCs; big tech and large private companies, as well as state-backed and state-owned funds.
Chinese researchers know they operate with far less compute than labs like OpenAI or Anthropic. That forces them to be more disciplined: fewer wild experiments, more targeted bets.
On the state side, you have vehicles like China’s national semiconductor investment fund taking stakes in companies such as DeepSeek. The government is strategic, not omnipresent — it wants influence in key players, but a lot of the frenzy is still driven by market investors chasing the next AI champion.
Q: There is lots of debate on AI safety in the US. Is China having the same debate?
Eddie: The tone is very different. In the US, you hear a lot of existential-risk talk. In China, the focus is more pragmatic. AI is seen as a way to squeeze growth and efficiency out of a tough economy — weak consumer sentiment, tight job markets, pressure on productivity.
Regulators like the Cyberspace Administration of China absolutely discuss AI agent risks, but it’s a dry, policy‑driven debate, not a public moral panic. The dominant question is still ‘how do we deploy this to help the economy?’ rather than ‘should we slow this down?’
Q: How big is China's AI data center build-out?
Eddie: Much smaller than many people think. When we looked through thousands of Chinese government documents on AI data centers, we found that OpenAI’s Stargate alone — a proposed half‑trillion‑dollar mega‑infrastructure plan — would cost about five times more than everything China has spent on AI data centers in the past five years. In China, typical AI data center projects cost hundreds of millions of dollars each and are spread across hundreds of smaller sites.
Q: With Nvidia constrained by US export controls, what chips is China using?
Eddie: Until 2023-24, Nvidia NVDA.O dominated a China AI chip market Nvidia’s CEO Jensen Huang has called a $50 billion opportunity. Export controls have effectively forced Nvidia to retreat from the highest end of that market.
That’s opened space for Huawei, which has become the de facto primary supplier for advanced AI workloads in China. Huawei knows it’s unlikely to match Nvidia’s single‑chip performance anytime soon, so its strategy is to build the best chip it can under sanctions, and use advanced packaging and network design to link thousands of chips into large systems, or “super nodes.” The goal is to try to compete at the system level, not the chip level.
The big questions are if Huawei's system architecture meaningfully narrows the gap with Nvidia, and if Huawei can scale fabrication and improve yields enough to meet domestic demand.
Until both answers are clearly yes, China’s AI compute capacity will remain constrained.
Q: Why is China so strong on open-source models?
Laurie: Initially, this wasn’t a political directive. Chinese labs recognized they were behind US frontier labs, so opening weights was a pragmatic way to accelerate collective learning and let researchers build on each other’s work.
Later, Beijing embraced open-weight models as a diplomatic tool, pitching them as a global public good for countries that lack their own frontier AI.
The problem is monetization. Many open-weight labs struggle to turn popularity into revenue.
Q: What about robotics and humanoids? Is China really ahead of the US?
Eddie: On hardware, China is very strong. Companies like Unitree, UBTECH, and others have made impressive gains in mobility and balance, dynamic control, and visually striking performance demos — running, jumping, kung fu routines.
But what really matters is embodied AI software: models that let robots truly understand and operate in the physical world. That’s where the bottleneck is. Timelines are highly uncertain. As Unitree’s founder Wang Xingxing put it, a “ChatGPT moment” for embodied AI could be 2 to 3 years away if things go fast, or 5 to 10 years away if progress is slower.
If that software breakthrough doesn’t come, we could see failed business models, bankruptcies, and a sharp reset in robotics valuations.
Laurie: On cutting-edge embodied AI models, US startups are still slightly ahead. A recent example is a US company that showed a humanoid learning a new task from a single short video, with no lengthy retraining. When I asked a Chinese founder about it, he acknowledged that no Chinese team has matched that yet.
The race now is to build robots that can do general-purpose tasks, autonomously, with minimal data and training, and still hit high reliability. That’s where the real gap lies.
CHART OF THE WEEK
Here's a number that could put Silicon Valley on notice: 73% of Americans don't think AI companies have done enough to prevent their technology from causing serious harm.
That's according to a new Reuters/Ipsos poll of 1,277 adults conducted September 17-20. Only 24% think the safeguards are adequate. For an industry that spends a lot of energy talking about "responsible AI," that's a pretty resounding vote of no confidence.
The skepticism doesn't stop there. More than half of respondents, 55%, said they'd welcome a slowdown in AI development, a striking number given how much the industry's narrative leans on speed as a competitive necessity. And this isn't a one-off blip: the same survey found 39% now see AI's impact on society as negative, the highest share since Reuters/Ipsos started asking in March, while just 11% call it a net positive. While the technology keeps improving, trust in it doesn't. Turns out those are two different problems.
(Reporting by Krystal Hu, Editing by Rosalba O'Brien)