China's AI Developments

China isn't replicating AI anymore - it's defining it. A decade ago, China was labeled a tech imitator. Today it's an AI superpower setting global trends. Massive government investments have sparked innovation: - Over $140 billion targeted for AI by 2030. - Giants like Baidu, Alibaba, and Tencent lead national AI projects. China's AI innovations are transforming industries: - Healthcare: AI accurately diagnosing cancers and rare diseases. - Manufacturing: Predictive AI cutting downtime; humanoid robots redefining factories. - Autonomous Vehicles: Baidu's Apollo rivals Tesla, aiming for over 50% market penetration by 2030. - Generative AI: DeepSeek & Manus AI deliver AI solutions 13-20x cheaper than Western counterparts. China also dominates global AI research and talent: - Leading worldwide in AI publications and patents. - Home to thousands of AI companies and a booming talent pool. We aren't just witnessing an AI shift - we're experiencing a new era shaped by China. Paradigm shift.

To understand the global AI landscape you need to keep current on what's happening in China. This compact report from Artificial Analysis gives an excellent overview of the current state of AI in China. Summary highlights from the report: 🇨🇳 Chinese AI labs narrow the gap with U.S. leaders. As of Q2 2025, China's top AI labs have reduced the performance gap with U.S. labs from over a year to under three months. DeepSeek R1 (May 2025) and Alibaba's Qwen3 235B A22B are now trailing only slightly behind OpenAI's o3 model, the global intelligence leader. 🔓 China leads in open weights model performance. Chinese labs took the global lead in open weights AI models in November 2024, with Alibaba’s QwQ 32B surpassing Meta's Llama 3.1 405B. DeepSeek's R1 (May 2025) is now the top open weights reasoning model, reflecting a strategic divergence from U.S. labs which largely withhold model weights. 🚀 DeepSeek emerges as China’s premier AI lab. In less than two years, DeepSeek advanced from its first model in November 2023 to becoming the #2 AI lab globally, tied with OpenAI. The May 2025 release of R1-0528 featured a 48-point intelligence jump through post-training alone, despite no architectural changes to its 671B model. 🧠 Post-training is the new frontier in model improvement. DeepSeek's intelligence leap in R1-0528 was achieved entirely via post-training—underscoring the growing power of reinforcement learning and fine-tuning, especially for reasoning capabilities. This method enabled it to outperform models with similar architectures. 🇺🇸 U.S. AI leadership becomes more fragmented. While OpenAI’s o3 remains the most intelligent model, the U.S. frontier now includes competitive entries from Google (Gemini 2.5 Pro), xAI (Grok 3 Mini), and Anthropic (Claude Opus 4). OpenAI is no longer the unchallenged leader, even domestically. 📱 AI consumer apps in China dominate user adoption. Alibaba, ByteDance, and Huawei each operate AI apps with over 100 million monthly users—Tongyi Qianwen, Doubao, and Celia, respectively. ByteDance’s Doubao is China’s most used chatbot (~110M MAU), while Huawei's Celia leads in voice AI (~200M MAU). 🏗️ A vibrant and well-funded startup ecosystem fuels Chinese AI. Startups like DeepSeek, Moonshot, and StepFun are central to China’s AI rise, many backed by big tech. DeepSeek Chat already has ~100M monthly users, and companies like StepFun have developed models over 1 tn parameters. 🎨 China reaches parity in text-to-image generation. ByteDance's Seedream 3.0 now ranks just 4 ELO points behind OpenAI's GPT-4o in image generation, achieving effective parity in quality. The top image models globally come from nations beyond just the U.S. and China. 🎬 U.S. still leads in video generation, but China is closing in. Google’s Veo 3 dominates both text-to-video and image-to-video benchmarks, but Chinese labs like Tencent and Alibaba are highly competitive. Alibaba’s Wan 2.1 and Tencent’s Hunyuan Video trail only slightly.

The Unavoidable Force: Why #China’s #AI Momentum Can’t Be Wished Away When Jensen Huang speaks about artificial intelligence, it isn’t conjecture. His vantage point sits at the center of the global AI supply chain. He sees demand, capability, and momentum before most of the world does. What he outlined recently should make every serious technology leader pause. Not because it is dramatic. But because it is structural. Today, roughly half of the world’s AI researchers are Chinese. An even larger share of newly filed AI patents—nearly seventy percent—originate from China. These aren’t vanity numbers. They signal where sustained capability is being built, not just where headlines are generated. For years, the comfortable assumption was simple: the West would innovate, Asia would execute. That model no longer holds. Over the past decade, China has moved decisively from adoption to authorship. In several domains of applied AI, materials science, and systems engineering, it is no longer following—it is defining the tempo. What sits beneath this shift is not secrecy or shortcuts. It is education. One of the most striking observations Jensen shared is that nine of the world’s top ten science and technology universities are now in China. This did not happen quickly, and it did not happen by accident. It reflects decades of focused investment in human capability, research infrastructure, and academic rigor. Innovation ultimately follows talent density. When you combine scale, discipline, and sustained policy alignment around STEM excellence, you create an engine that compounds year after year. Science, unlike ideology, does not respond to narratives. It responds to competence. There is also a lingering myth that China’s technology ecosystem is derivative. That view is increasingly detached from reality. The AI landscape there is broad, competitive, and internally reinforcing. Research feeds industry. Industry feeds patents. Patents shape global standards. This is how durable advantage is formed. Geopolitics understandably complicates collaboration. But ignoring technological reality because it is inconvenient is not strategy—it is denial. No country or bloc can meaningfully advance AI while discounting half of the world’s intellectual output. This is not about fear. And it is not about surrender. It is about clarity. The center of gravity in AI has shifted. Recognizing that does not weaken the West; it sharpens it. Serious leadership begins with understanding the board as it actually exists, not as we wish it to be. In the AI era, the most valuable asset is not compute, capital, or even data. It is cultivated human capability—at scale. And on that dimension, China has become impossible to ignore. DC*

"As AI capabilities and governance challenges continue to grow—and Chinese models close their gap with the leading edge—understanding China’s role in AI safety and governance is more critical than ever. Since 2023, Concordia AI’s annual State of AI Safety in China reports have analyzed how China addresses general-purpose AI risks, particularly dangerous misuse, accidents, and loss of control—challenges with an outsized need for international cooperation. This year’s report provides updates from May 2024 to June 2025 across five domains: domestic governance, international governance, technical safety research, expert views on AI safety and governance, and industry governance. Domestic Governance • Domestic rhetoric and policy include increasingly prominent and specific calls for AI risk mitigation, while continuing to emphasize the complementarity between AI safety and development. • China is implementing its AI regulations through an expanding AI standards system. International Governance • China has emphasized AI safety and global AI capacity-building as key themes in its international AI diplomacy • China has launched bilateral AI dialogues with several key countries, though the outlook for these engagements remains mixed. Expert Views • Expert discourse in China is placing greater emphasis on AI safety and governance • Experts are increasingly publishing in-depth analyses of AI risks in biosecurity, cybersecurity, and open source AI. Industry Governance • Most leading Chinese foundation model developers have signed voluntary “AI Safety Commitments.” • Chinese AI developers typically implement well-known safety methods, but provide limited transparency on safety evaluations." Gabriel W., Jason Zhou, Kwan Yee Ng 吴君仪, and Brian Tse 谢旻希 at Concordia AI 安远AI

Something big recently happened in the world of AI, and it wasn’t in Silicon Valley. Earlier this month, China’s Moonshot AI released a new large language model called Kimi K2. It’s one of the most powerful open-source AI models we’ve seen anywhere in the world, with over a trillion parameters under the hood. That alone makes it noteworthy. But it’s how China is moving—and why—that deserves a closer look. Here’s what I find interesting: - Kimi K2 is huge, but efficient. While the model has a trillion parameters (the dials and switches that help it generate text, code, and more), only a small portion are used at any one time. This clever design, known as a “mixture-of-experts,” means it delivers high performance without the massive cost of running every parameter for every task. - It’s open. Really open. Moonshot released the model weights for anyone to use. That means that researchers, companies, and even government can quickly build on it and deploy it. It's a playbook we’ve seen from U.S. firms like Meta, but here it’s being applied from inside China’s Great Firewall. - Reports suggest that it performs well, especially for agents and coding. Kimi K2 reportedly outperforms most open models on tasks that require using tools, writing code, and completing complex, multi-step jobs. These are the skills which are foundational to AI agents and autonomous systems. So, what does this mean for the bigger picture? The release of Kimi K2 signals that China’s AI ecosystem is maturing quickly, despite U.S. restrictions on advanced chips. Moonshot AI is proving that it’s possible to build sophisticated models with fewer resources, especially when those models are open-source and can improve rapidly through community use. And it raises a deeper question: If frontier AI is less about the chips you control, but the models you share, how do we plan to maintain America’s lead? As someone who previously led digital innovation at the CIA, I’ve seen firsthand how emerging technologies can reshape the balance of global influence. Kimi K2 seems like more than a technical release to me, perhaps akin to a strategic move. A signal that the AI competition between the U.S. and China is accelerating, and that we need to think deeply about how we define and maintain leadership in this space. This one is not just about size. It’s about speed, openness, and the ability to shape ecosystems. And that’s why, in my view, the release of Kimi K2 matters. More broadly, I’m watching three things closely: 1.    How will open-source AI evolve when powerful models are no longer limited by geography or policy? 2.    What new risks and opportunities arise when any actor, good or bad, can build advanced AI on top of public models? 3.    How do we ensure democratic innovation keeps pace, maintaining safety, trust, and purpose at the core? What’s your take? Let me know in the comments below. #ArtificialIntelligence #OpenSourceAI #USChinaTechRace #Geopolitics #NationalSecurity

While DeepSeek AI is getting all the attention for the last month, it's easy to forget that China has an entire ecosystem of companies making advancements in different parts of the stack. China’s progress in AI is not driven by isolated breakthroughs, but by a dense network of companies advancing every layer of the technology stack. This ecosystem spans hardware, software, and real-world applications, creating a self-reinforcing cycle of innovation. At the hardware level, companies like SMIC and Huawei are developing semiconductors to meet the computational demands of AI. Hesai Technology produces LiDAR systems critical for autonomous navigation, while CATL and BYD push battery efficiency, enabling energy-intensive AI applications in electric vehicles and robotics. Software companies such as Baidu, Inc. and Tencent refine AI algorithms for natural language processing and data analysis, while ByteDance applies machine learning at scale in consumer platforms. Autonomous vehicle startups like Pony.ai and electric vehicle companies like Xpeng are advancing AI-driven perception and decision-making systems, with some models incorporating assisted driving features. The interplay between these sectors accelerates progress. For example, improvements in battery density from CATL enable longer operational ranges for electric and autonomous vehicles. Meanwhile, autonomous vehicles equipped with sensors and onboard AI continuously collect real-world driving data, which can be used to refine machine learning models for safer and more efficient navigation. Similarly, SMIC’s chips power both data centers running large language models and edge devices in smart factories. This ecosystem thrives because companies solve immediate, practical challenges, whether in manufacturing, transportation, or consumer tech, while feeding advancements back into the broader stack. The result is a pragmatic, vertically integrated approach to AI development, where progress in one layer amplifies capabilities across others. China’s edge lies not in any single technology, but in the collective momentum of its industrial chain. Image source: Kyle Chan from High Capacity. Link: https://lnkd.in/gpuqhhWs

Chinese Open-Source AI Matches OpenAI’s Best at a Fraction of the Cost Chinese researchers have unveiled DeepSeek R1, an open-source AI model that matches the performance of OpenAI’s leading system, o1, while operating at 98% lower costs. This remarkable achievement was made possible through a unique training method that relies solely on reinforcement learning (RL), bypassing the need for costly supervised fine-tuning. Revolutionary Training Method DeepSeek R1 employs large-scale reinforcement learning without an initial phase of supervised fine-tuning, a departure from traditional AI training methods. Reinforcement learning allows the model to learn through trial and error, mirroring human learning processes. Here’s how it works: 1. Trial and Error: The model experiments with decisions and is rewarded for good outcomes and penalized for poor ones, gradually improving its performance. 2. No Pre-Tuning Needed: Unlike OpenAI’s systems, which require human input during the supervised fine-tuning phase to guide initial learning, DeepSeek R1-Zero starts learning without any pre-defined context, saving both time and resources. This approach not only lowers development costs but also enables broader accessibility for researchers and organizations. Performance Metrics Despite its budget-friendly training model, DeepSeek R1 demonstrates reasoning capabilities on par with OpenAI’s o1. It has also been ranked among the top 10 AI models globally, solidifying its place as a formidable competitor in the AI space. Implications for AI Development 1. Cost-Effective AI: By eliminating expensive supervised fine-tuning, this open-source model paves the way for democratizing AI research. 2. Reinforcement Learning Potential: The success of DeepSeek R1 underscores the growing potential of RL as a primary training method for AI. 3. Geopolitical Implications: With this achievement, China reinforces its position as a major player in the global AI race, challenging the dominance of Western tech giants like OpenAI. DeepSeek R1’s breakthrough demonstrates that high-performing AI systems can be developed with innovative methods at a fraction of the cost, pushing the boundaries of what’s possible in AI research and accessibility.

It is critical that we get reporting live, on the ground, from China. Thanks, Bloomberg, Stephen Engle, and Adrian Wong for your commitment to covering all the important happenings in our geography.   I have attended the China Development Forum for the past few years, and I must say that this year the sentiment was quite different. Attendance was at a record high, with a lot of global investors whom I have not seen in the past couple of years. But here’s what caught my attention: Geopolitics is always a topic, but the difference this year is that half the room thinks that geopolitics has some upside for Chinese businesses and multinationals operating in the country. Coincidentally, inbound tourism to China in 2025 reached an all-time high of 150 million, with over 30 million entering visa-free. Amidst all the geopolitics and uncertainty in the world, the surprising sentiment? Interest is back in a big way, in both business and personal tourism.   Everyone’s attention is on the EVs, solar panels, and robots. But watch this too: the rapid scaling of China’s latest export – the export of AI tokens. China is already the world’s largest AI “token economy,” and at this rate of acceleration, it’s growing exponentially by the minute. This is the latest transformation of Chinese exports – from hard goods to soft goods – IP, patents, brands, and now tokens.   This trend isn’t slowing down. Here’s why: With electricity costs more than 30-40% lower than in other developed countries, tokens have become a derivative of electricity. Chinese models are often priced at $0.30 per million input tokens (or lower), while comparable US models can cost $5 or more. We are effectively exporting the Chinese electricity grid. It’s intangible and hard to tariff. It’s not even counted in the trade balances. Yes, there are data sovereignty and latency challenges – and these apply to every country’s token economy – but the overall trend is continuing.   Like the Chinese hard goods export phenomenon before, affordability leads to more adoption, which in turn leads to more scale and drives down costs. That pattern doesn't break.   Here’s another observation: alternative sources of energy are becoming even more important. The Chinese solar, wind, and battery companies are sensing a moment where they could be seen not only as a sustainability play, but also as an alternative and resilient play. A year ago? That wasn’t on anyone’s radar.

I recently spent 10 days traveling through Beijing, Shenzhen, Hangzhou, and Shanghai to understand the state of China's AI ecosystem. I spoke to the cofounders of Deepseek, the team behind Qwen (Alibaba Cloud), partners at Hongshan (prev. Sequoia China), the founder of Sinosoft, 50+ AI x Robotics startups at all stages, and attended the Miracleplus (prev. YC China) F25 Demo Day. Here are a few things I learned. 1. China’s B2B ecosystem is split into two markets: a small, competitive free market and a dominant, connection-driven state sector. While private startups operate like US firms (buying software to fix inefficiencies), the market is overwhelmingly controlled by State-Owned Enterprises (SOEs) that buy based on social connections. Because these giant SOEs demand heavy customization and prioritize stability over optimization, cloud and enterprise AI adoption remains surprisingly low. This condition creates a 10% ceiling that prevents even market leaders from achieving the 40–50% dominance seen in the US. To survive, Chinese tech companies are forced to abandon pure SaaS models and verticalize. Large software companies eventually all converge to building the entire stack and become project-based consultancies that sell low-margin consulting hours rather than high-margin software licenses. With SMBs too volatile and unwilling to pay, the result is a significantly underdeveloped B2B market. 2. I think the prevailing Western narrative that Chinese AI is simply "six months behind" is an oversimplification. China’s top foundational labs have achieved near-parity with US frontier models. However, there are virtually no independent winners in critical infrastructure. Unlike the US, where categories like data labeling, evaluation, and neo-clouds have birthed multi-billion-dollar giants (e.g., Scale AI, CoreWeave), China’s ecosystem has evolved through vertical integration, leaving these industries virtually empty. 3. Seed funding in China has shifted away from private capital toward a model where City/District Guidance Funds (e.g. Shenzhen Innovation Fund) are the dominant players. Unlike US VCs chasing 100x financial returns, these funds are motivated by boosting local GDP and tax revenue; as a result, this capital comes with strict golden handcuffs. To receive funding, startups are obligated to register, pay taxes, and physically operate within that district, effectively turning venture capital into a tool for regional urban planning rather than free-market allocation. Insanely grateful to the people who made the trip possible. Daniel Tian, let me crash their demo day. Anyone building agents within iOS or WhatsApp should check out https://photon.codes/ Kevin Liu, made this trip possible Aili Liu, let me crash Future Factory for a day in Shenzhen. Insanely cracked group and would highly recommend anyone who want a more structure path to explore China's tech ecosystem. And many many more people I met along the way that are not on Linkedin.