US-China AI technology debate

The buzz over DeepSeek this week crystallized, for many people, a few important trends that have been happening in plain sight: (i) China is catching up to the U.S. in generative AI, with implications for the AI supply chain. (ii) Open weight models are commoditizing the foundation-model layer, which creates opportunities for application builders. (iii) Scaling up isn’t the only path to AI progress. Despite the massive focus on and hype around processing power, algorithmic innovations are rapidly pushing down training costs. About a week ago, DeepSeek, a company based in China, released DeepSeek-R1, a remarkable model whose performance on benchmarks is comparable to OpenAI’s o1. Further, it was released as an open weight model with a permissive MIT license. At Davos last week, I got a lot of questions about it from non-technical business leaders. And on Monday, the stock market saw a “DeepSeek selloff”: The share prices of Nvidia and a number of other U.S. tech companies plunged. (As of the time of writing, some have recovered somewhat.) Here’s what I think DeepSeek has caused many people to realize: China is catching up to the U.S. in generative AI. When ChatGPT was launched in November 2022, the U.S. was significantly ahead of China in generative AI. Impressions change slowly, and so even recently I heard friends in both the U.S. and China say they thought China was behind. But in reality, this gap has rapidly eroded over the past two years. With models from China such as Qwen (which my teams have used for months), Kimi, InternVL, and DeepSeek, China had clearly been closing the gap, and in areas such as video generation there were already moments where China seemed to be in the lead. I’m thrilled that DeepSeek-R1 was released as an open weight model, with a technical report that shares many details. In contrast, a number of U.S. companies have pushed for regulation to stifle open source by hyping up hypothetical AI dangers such as human extinction. It is now clear that open source/open weight models are a key part of the AI supply chain: Many companies will use them. If the U.S. continues to stymie open source, China will come to dominate this part of the supply chain and many businesses will end up using models that reflect China’s values much more than America’s. Open weight models are commoditizing the foundation-model layer. As I wrote previously, LLM token prices have been falling rapidly, and open weights have contributed to this trend and given developers more choice. OpenAI’s o1 costs $60 per million output tokens; DeepSeek R1 costs $2.19. This nearly 30x difference brought the trend of falling prices to the attention of many people. [...] [Reached length limit. Full text: https://lnkd.in/grbFH4D6 ]

Today’s AI model release by Baidu underscores what we’ve been saying for the past year—that global AI competition is real; the competition is formidable; and the stakes are of enormous consequence. Do we want the world building its AI future on US-led democratic AI, or on autocratic AI? The Baidu release underscores the importance of a small-“d” democratic vision for AI based on three pillars: 1. The US needs to appreciate that AI is a core technology—akin to electricity. Virtually every country in the world is going to want AI but only two are capable of building it at scale: the US and the Communist Party-led People’s Republic of China. Most nations would prefer to build on US-led, democratic AI. A few weeks ago, we launched OpenAI For Countries to help support a commercial diplomacy approach to advancing democratic AI. We’ve detailed ways in which China is pursuing its own global approach. The competition to determine whether the world is built on democratic AI or autocratic AI is happening now and we need to seize this moment. 2. Infrastructure is destiny. That means the AI competition revolves around chips, data, energy, and talent. The acceleration of the AI infrastructure build-out will create jobs and support re-industrialization; serve as a catalyst for modernizing our energy grid; and ensure that we are making the economic opportunities of AI widely available and accessible. China brought 10 nuclear facilities online last year and is scheduled to bring on another 10 this year. Advances in various Chinese AI models make clear that their infrastructure commitment is real and meaningful. The US knows how to think big, act big and build big—and we need to do so when it comes to AI infrastructure. 3. We need a federally-led approach to regulating frontier AI systems that implicate strategic national security and economic competitiveness imperatives. We have been making the case that a mosaic of state laws—which builders of autocratic AI don’t need to follow—is inconsistent with a cohered national strategy (inclusive of the Center for AI Standards and Innovation) organized to optimize for responsible, safe, and democratic AI. We have also made clear that on issues like child safety and deepfakes, these are areas where states can act and have supported such proposals at both the federal and state levels. The Race to the Moon was not won by 50 different states pursuing 50 different plans—but by the United States putting its shoulder into one. https://lnkd.in/gXv3z5tH

Alibaba recently announced that its enterprise AI agent platform, Wukong, will integrate directly with Slack and Microsoft Teams. Read that again. Slowly this time. The workflows most U.S. companies use to coordinate sensitive internal decisions (think approvals, document editing, research, meeting transcripts) are now being offered a Chinese AI infrastructure layer, packaged as a productivity tool. That same week, Tencent launched ClawBot, a WeChat-integrated AI agent reaching over one billion users. The U.S.-China Economic and Security Review Commission published a report on March 23 explicitly warning that China’s strategy of releasing powerful open-source models creates a self-reinforcing cycle: widespread adoption drives faster improvement, which drives more adoption, which erodes U.S. proprietary model advantages over time. The product launches are not the whole story. The vertical integration is. Alibaba now controls chips, models, and enterprise agents in a single stack. The Slack and Teams integration is a distribution strategy designed to make Chinese AI infrastructure the path of least resistance for enterprises building agent workflows. Every organization that routes sensitive internal processes through Wukong creates a dependency. Dependencies, in my former world, are called vulnerabilities. There is a legal dimension here that does not get enough attention in discussions about Chinese technology adoption. Article 7 of China’s National Intelligence Law (2017) requires that all organizations and citizens “support, assist, and cooperate with national intelligence efforts in accordance with law.” Article 14 gives intelligence agencies the authority to formally require that cooperation from any organization, including technology companies. This is not a hypothetical. It is the legal architecture under which every Chinese tech company operates, and it applies to data they process on behalf of their enterprise customers. The question boards and leadership teams need to ask is not just whether an AI tool is useful and cost-effective, or whether the vendor is trustworthy in a commercial sense. It is whether the legal environment governing that vendor creates an obligation to provide access to your organization’s data that your procurement process never accounted for. The DeepSeek moment showed how quickly adoption can accelerate when the price is right and the capability is real. And it is real. It’s a good product. The question for enterprise leaders is not whether Chinese AI tools are useful. It is whether the combination of dependency risk and potential data exposure has been assessed, or whether it has been accepted by default. Has your organization evaluated the data flows and legal obligations of the AI infrastructure your teams are adopting right now? #StrategicCompetition #AI #NatSec #AIGovernance #NationalSecurity #TechPolicy #DigitalResilience

The US raised $250 billion in AI funding last quarter. Every frontier lab is American. And yet China now has roughly double the users of the most viral AI agent framework on the planet. Something does not add up. Or maybe it adds up perfectly. If you have spent time in Chinese tech, you have seen this before. Early 2000s. Shenzhen factories cranking out handsets. Every startup building a messaging app. Chaotic, redundant, mostly doomed. Superior technology did not capture the largest market. Distribution and proximity to daily life did. That chaos produced WeChat. OpenClaw, a framework for building personal AI agents, went viral in early 2026. In the US, Google disabled accounts of subscribers routing tokens through it. Anthropic killed OAuth access. No major cloud provider launched a hosted service. The security concerns were real (280+ advisories, thousands in compute burned per user) but the net effect was containment. In China, every major cloud provider launched a hosted service within weeks. Tencent shipped more than ten agent products in a single month. ByteDance built the official skill marketplace mirror. Baidu hosted installation events that drew thousands. When Tencent integrates OpenClaw into WeChat, the agent shows up as a regular chat contact. You message it. It controls your computer remotely. No separate download. No API keys. For 1.4 billion people who already pay bills, book rides, and file taxes inside one app, talking to an agent feels like talking to another contact. The infrastructure was already there. Then there is pricing. Chinese AI models run 5 to 24 times cheaper than American equivalents. A chatbot query might burn 30,000 tokens. An agent doing a complex task burns 20 million. At that scale, cost per token decides what gets built and what stays a demo. The US is building the best engines. China is building the roads. PS - I wrote about how WeChat became the playbook for this kind of distribution dominance. The panic, the four-character email, and the reclusive engineer who built a $700 billion super-app. Link in comments.

AI moves fast because there is no loyalty in this market. Customers switch models. Developers switch APIs. Investors switch narratives. Yesterday’s genius becomes today’s laggard after 1 benchmark and 2 tweets. Through this, one belief survived: open models would get good, but always remain a tier below the frontier. Last week, that belief took a body blow. Moonshot AI unveiled Kimi K3, a 2.8T parameter model built for coding, long-context reasoning and agentic work. Alibaba followed with Qwen3.8, a 2.4T parameter flagship whose weights are expected to open after preview. 2 Chinese labs introduced models in the ~2.5-to-3T-parameter range. Both use openness as distribution and target coding, reasoning and agentic work. Open models are no longer competing for leftovers. They are coming for the workloads with the strongest economics. The clearest evidence of Kimi’s success: Moonshot had to stop accepting new subscriptions because demand overwhelmed available compute. Champagne problems. The open source resurgence is uncomfortable for closed model labs because scarcity underpins their economics. Open weights reduce switching costs, strengthen buyer leverage and make it harder to charge a permanent premium for temporary technical leadership. Washington’s growing anxiety is revealing. The administration is considering, among other restrictions, a ban on Chinese models. The risks are real: Chinese-hosted APIs can expose sensitive data to foreign jurisdiction, models may encode censorship or political bias, and highly capable open weights can lower the barriers to cyber, military or biological misuse. Yet a blanket ban would be porous, economically distortionary and geopolitically self-defeating. Once model weights are released, they are hard to contain. Restrictions would bind compliant U.S. companies more than determined adversaries, raise costs for startups, and reduce competition. It would help China cast itself as the champion of accessible AI while the US becomes identified with expensive, permissioned intelligence. The smarter framework regulates risk rather than nationality: restrict foreign-hosted services in sensitive environments, and require rigorous provenance, security and behavioral testing for locally deployed models. Hosted APIs and downloaded weights should be treated differently. This approach is less emotionally satisfying than banning the scary foreign model. It has the minor advantage of addressing the actual risk. The strategic answer is not to wall off Chinese models, but to build a stronger American open-model ecosystem that competes on capability, trust, cost and distribution. If the frontier is becoming more open, America should compete to shape it, not retreat behind it.

"We can’t lose to China. When discussing AI governance, this one line is often produced as a conversation-ending trump card. The rationale goes that AI safety requirements will undermine the US’s efforts in technology competition with China. Common wisdom holds that regulation will stymie market competition and innovation through exorbitant compliance costs. Innovative technologies drive productivity, and subsequent economic growth manifests as national power. Advanced technologies are also crucial to ensure that the US has a military advantage. This paper analyzes the claim that AI safety and US primacy are direct trade-offs. Zerosum competition is a dangerous framework and one to navigate carefully. However, for the sake of scope, we place these concerns to the side and ask the question: Can the US still “win” against China if it introduces AI safety regulation? We interpret “winning” as the US leading in AI innovation, translating such innovation into economic growth, and developing and fielding superior military technologies. We find that AI safety measures will not hinder the US in an “AI race” because: 1. Inexpensive AI safety costs are unlikely to impact AI innovation. 2. Safer AI can drive trust and thus adoption, which is crucial for economic growth. 3. Safety requirements on public models will not impede military capabilities. Our analysis focuses on mandatory pre-deployment evaluations, which are a series of techniques to assess risk prior to releasing a model to the broader public. By limiting these requirements to the developers of the most powerful models, such evaluations would likely cost less than 0.5% of upfront training costs. The relative compliance cost would be even smaller if training costs reach the predicted one billion dollars by 2027. Expensive training costs and the current market landscape also suggest that small companies will not be burdened by these compliance costs since they cannot afford to operate in this market, regardless of regulation. We also find that introducing pre-deployment evaluations could increase public trust in AI, drive greater adoption in business settings, and thus spur economic growth. Finally, we conclude that pre-deployment evaluations have no effect on the government's ability to develop and deploy military and intelligence AI models, since the government could feasibly develop partnerships to access models prior to them being released to the public." By Claudia Wilson at the Center for AI Policy

China does not need to beat America at AI. It only needs to make intelligence cheap enough that using frontier models for everything becomes economically indefensible. That is the part of the East versus West AI race I think most boards are missing. The performance gap between the leading US and Chinese models has fallen to around 2.7 percent, while US private AI investment is more than 23 times China's. So the real battle may not be who builds the smartest model. It may be who makes intelligence cheap enough to use everywhere. I can speak unusually confidently about this because I am one of the very few independent people in AI who has had close access to all the major AI labs across the US, Europe and Asia, including spending time inside China's AI ecosystem. I am not tied to Microsoft, Google, OpenAI or any single model vendor. My new newsletter explains why tokenomics, model routing and open models could become far more important to enterprise AI than the model leaderboard itself. The winners will not choose one model. They will make models compete for every task.

America spent a trillion dollars building the best artificial intelligence on earth. Last week it gave the rest of the world a reason never to depend on it. Three moves in a single week tell the story. The United States ordered Anthropic to cut its two most powerful models off from every foreign national, days after launch. India included. China spent five years quietly deleting 12,200 university degrees and rebuilding them around AI and robotics, just as a record 12.7 million graduates reach its workforce. India stopped waiting. Globalisation is dead, said Zoho's Sridhar Vembu. Bharat must find her own way. A new JPMorgan geopolitics report explains why these three belong in the same breath. The AI race, it argues, will not be won at the frontier. The model leaderboard is converging. It is being decided one layer down, in energy, supply chains and diffusion. Start with energy. China generates more than twice the electricity the United States does, at 8 cents a kilowatt hour against America's 18. Since 2019 it added 2,500 terawatt hours of new power. America added 221. Every model runs on that power, and cheap power is the ceiling on how much intelligence a country can deploy. Then reach. America builds the most advanced systems and keeps them closed. China builds good enough systems, makes them open and cheap, and exports them across Southeast Asia, Africa and Latin America. Downloads of Chinese and American models reached parity last year. Here is the number that should stop you. America holds three quarters of the world's AI venture capital and outspent China five to one on AI last year. It is winning the money race and losing the layers that compound. Two decades in payments, at PayPal, at Mastercard, at Gojek, across five markets, taught me the same thing every time. The better product rarely wins. The one that reaches everyone does. A model is a product. Diffusion is the infrastructure. Only one of them can be taken away from you. For every country and every company leaning on borrowed intelligence, the choice is narrowing. Build the rails, or keep paying rent on them. Source: JPMorganChase Center for Geopolitics, Beyond the Benchmarks, May 2026. Report in the first comment. #AI #Infrastructure #Geopolitics #DigitalSovereignty