 |
| The U.S. and China are engaged in an intense competition over artificial intelligence technology. On September 8, 2026, U.S. intelligence agencies accused six Chinese AI companies of systematically stealing American AI models through knowledge distillation campaigns." US-China-AI-Competition-2026.jpg |
On September 8, 2026, three major U.S. intelligence agencies made an unprecedented move. The Federal Bureau of Investigation, the National Security Agency, and the Cybersecurity and Infrastructure Security Agency jointly accused six Chinese artificial intelligence companies of systematically stealing technology from American firms. The accusation represents a dramatic escalation in what many are calling the defining technological rivalry of the century—a competition between two global superpowers over who will control the future of artificial intelligence.
The allegations are specific, detailed, and damning. The Chinese companies have allegedly extracted billions of data tokens from American AI models since late 2024, using a technique called "knowledge distillation" that allows them to replicate the capabilities of sophisticated American systems while cutting costs and development time dramatically. This isn't industrial espionage in the traditional sense. It's something more nuanced, more technical, and potentially more destabilizing—a form of intellectual property theft that operates in a legal and regulatory gray zone.
The Six Companies at the Center of the Controversy
The U.S. government's advisory named six Chinese AI companies as the primary actors in what officials describe as "aggressive, malicious, and targeted distillation activities." Understanding who these companies are and what they do is essential to grasping the scope of the accusation.
DeepSeek is perhaps the most prominent name on the list. The company has gained international attention for releasing AI models with remarkably low reported training costs—claims that U.S. officials now argue are misleading. DeepSeek allegedly conducted organized distillation campaigns targeting specific capabilities from American frontier models. The company's R1 and V3 models, which represent significant breakthroughs in AI reasoning, allegedly benefited from systematic extraction of data from Claude, ChatGPT, Gemini, and Grok.
Moonshot AI represents another significant actor in this space. According to the advisory, the Beijing-based company conducted widespread distillation operations against American models to improve its Kimi model family. The scale of these operations was substantial, involving millions of queries across multiple U.S. frontier systems.
Alibaba, the massive Chinese e-commerce and technology conglomerate, used distillation to enhance its Qwen family of AI models. Alibaba's involvement signals that this isn't just the work of smaller startups—major multinational corporations with significant resources are allegedly participating in these campaigns.
MiniMax, StepFun, and Z.AI round out the list of accused companies. While these firms are perhaps less well-known internationally, their inclusion demonstrates the breadth of the alleged operations across China's AI ecosystem.
What Is Knowledge Distillation and Why Does It Matter?
To understand the controversy, it's essential to grasp the technical practice at the center of it: knowledge distillation. Distillation is a legitimate machine learning technique where researchers train a smaller, more efficient model by using the outputs of a larger, more powerful model as training data. The smaller model "learns" from the larger one's responses, capturing much of its capability in a more compact form.
The practice itself is not new or inherently illegal. Artificial intelligence companies worldwide, including leading American firms, use distillation internally. OpenAI, Google, and Anthropic all employ distillation to create mobile-friendly versions of their models, to compress systems for specific applications, or to develop more efficient variants of their frontier models.
 |
| "Six Chinese AI companies—DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI—were named in a U.S. government advisory for allegedly extracting billions of tokens from American frontier AI models like ChatGPT, Claude, Gemini, and Grok since late 2024." DeepSeek-Alibaba-Chinese-AI-Companies-2026.jpg |
What the U.S. government alleges is different in crucial ways. The scale of the Chinese operations appears to be industrial and systematic. The companies allegedly used billions of tokens—essentially millions of queries—sent to American AI systems to extract specific capabilities. They targeted particular functionalities: reasoning abilities, specialized optimizations, domain-specific functions, and agent-like behaviors that allow models to use tools and solve complex problems.
The companies also allegedly employed evasion tactics to avoid detection. According to the advisory, they routed distillation requests through multiple pathways, including direct APIs, remote cloud providers, and third-party aggregators designed to obscure user metadata and hide the true origin of the queries.
The Technical Details of the Allegations
The federal advisory provides remarkably specific technical details about which models were targeted and how they were used. DeepSeek, for example, allegedly queried four separate versions of Claude, two versions of Google's Gemini, five versions of OpenAI's ChatGPT, and SpaceX's Grok 4 to generate synthetic training data for its models. The company focused on extracting reasoning capabilities—the ability of AI systems to work through problems step-by-step—as well as specialized optimizations and domain-specific functions.
Moonshot AI reportedly conducted millions of exchanges with American frontier models to train its Kimi systems. Alibaba used distillation across multiple U.S. models to improve the Qwen family. Each company allegedly had specific targets and strategies, suggesting coordination or at minimum, knowledge of what competitors were doing.
The advisory emphasizes that these weren't accidents or minor violations of terms of service. The operations represent systematic, organized campaigns to extract proprietary technology. The agencies note that these activities likely occurred "with Chinese government awareness," though they stopped short of claiming direct government involvement or orchestration.
The Cost Argument: Why This Matters Economically
One of the most contentious aspects of the accusation concerns cost. DeepSeek has claimed that it trained its advanced models for only $5.6 million—a figure that stunned the AI industry because comparable American models cost hundreds of millions of dollars to develop. The company's efficiency seemed to validate China's approach to AI development.
U.S. officials and American AI companies argue that DeepSeek's claimed costs are misleading because they don't account for the value of data acquired through distillation. If DeepSeek obtained years of expensive research, compute investment, safety engineering, and evaluation work from American firms without paying for it, then the true cost of its models is substantially higher than claimed. This isn't just about accurate accounting—it's about whether Chinese companies are winning an unfair competition through technological theft.
The financial implications are staggering. American AI companies have invested tens of billions of dollars in developing frontier models. Training a cutting-edge large language model requires massive computational resources, thousands of engineering hours, and millions of dollars in evaluation and safety work. If Chinese competitors can replicate these capabilities through distillation without bearing the full development cost, it fundamentally distorts competition.
China's Response and the Gray Area of Legality
The Chinese government and the accused companies have pushed back against the allegations. They argue that distillation is a standard, legal practice in the AI industry. They point out that American companies use the same technique. They suggest that the U.S. is trying to maintain technological dominance through accusations rather than through superior innovation.
This argument contains a grain of truth. Distillation itself is legal. The technique doesn't involve hacking, doesn't breach security, and doesn't violate laws in most jurisdictions. It works within the terms of what's publicly available—the outputs of AI systems that customers can access through APIs and web interfaces.
But the scale, systematicity, and intent create legal and ethical complications. When companies pay for API access, they agree to terms of service. Those terms typically prohibit using model outputs to train competing systems. The allegation isn't that distillation is illegal—it's that the Chinese companies violated terms of service at massive scale while using obfuscation tactics to avoid detection.
The gray area persists because distillation sits at an intersection of legitimate technique, terms-of-service violation, and potential intellectual property infringement. It's not theft in the traditional sense because nothing is stolen through unauthorized access. Yet it's not straightforward competition either, because it relies on terms-of-service violations and evasion tactics.
Why This Escalation Matters for Global AI Competition
The September 8 advisory represents a significant moment in U.S.-China relations over technology. For the first time, three major U.S. intelligence agencies formally and publicly accused Chinese companies of systematic AI technology theft. Previous accusations came from companies themselves or from individual officials speaking on background. This joint advisory carries the weight of official U.S. government institutional authority.
The timing also matters. The accusation comes as Chinese AI companies are advancing rapidly and challenging American technological supremacy. DeepSeek has gained millions of users globally with its R1 model, which showed remarkable reasoning capabilities. Alibaba's Qwen models are highly competitive. The American AI industry's concerns about Chinese competition have shifted from theoretical to immediate.
The advisory also reflects growing concern about export controls and sanctions evasion. The U.S. has imposed restrictions on selling advanced AI chips to China, attempting to slow the development of Chinese frontier models. Knowledge distillation represents a way to circumvent those restrictions—by querying American models through standard APIs, Chinese companies can acquire capabilities without needing access to American technology, training compute, or proprietary data.
For American policymakers, this is a national security issue. The competition over AI is viewed not just as a commercial matter but as a competition for technological and geopolitical dominance. The concern is that if China's AI capabilities advance too rapidly, it could affect military capabilities, economic competitiveness, and global technological standards.
The Broader Implications for the AI Industry
This accusation will likely reshape how American AI companies operate. Companies may begin implementing stricter monitoring of API usage, looking for signs of large-scale distillation attempts. They may modify their terms of service or pricing models to discourage bulk querying for distillation purposes. Some may restrict API access by geography or usage pattern.
The accusation also raises questions about the long-term sustainability of American AI dominance. If Chinese companies can achieve competitive capabilities through distillation of American models, does this validate a different development strategy—one focused on efficient extraction and adaptation rather than from-scratch innovation? Or will American companies continue to innovate faster than competitors can distill?
The international AI research community faces questions about norms and standards. Is large-scale distillation acceptable practice in an emerging technology field? Should it be regulated? How do companies balance openness and accessibility with protecting intellectual property and competitive advantages?
FAQ: Understanding the U.S.-China AI Dispute
Q: Is knowledge distillation illegal?
A: Knowledge distillation as a technique is legal and is used by AI companies worldwide, including American firms. However, when companies use distillation to violate terms of service or evade monitoring, it enters a gray legal area. The U.S. alleges that the Chinese companies violated terms of service at scale, which is not strictly illegal but is a contractual violation.
Q: Why would Chinese companies need to distill American models if they're developing their own?
A: Distillation allows companies to acquire capabilities without bearing full development costs. Instead of spending hundreds of millions on compute and research, they can extract what they need from existing systems. This accelerates development and reduces expenses—allowing companies to compete on efficiency rather than raw innovation.
Q: Could this be considered industrial espionage?
A: It's not traditional espionage because no hacking or unauthorized access is involved. It operates within publicly available systems. However, it involves violations of terms of service and evasion tactics, which creates an ethical gray area between legitimate practice and intellectual property theft.
Q: How can American companies prevent distillation?
A: Companies can implement monitoring to detect large-scale querying patterns, rate limit suspicious activity, modify pricing to discourage bulk usage, or restrict API access by geography or user type. However, completely preventing distillation is difficult without restricting legitimate use.
Q: Does this affect consumers using these models?
A: Potentially. If companies implement stricter access controls or pricing changes to prevent distillation, users might face higher costs or more limited access. The dispute could also influence which models are available in different regions.
Q: Is China the only country engaging in this practice?
A: The advisory names Chinese companies, but distillation is a global practice. Other countries' companies may also be engaging in similar activities, though they haven't been formally accused by U.S. intelligence agencies.
Q: What are DeepSeek's actual training costs if distillation is factored in?
A: The advisory suggests DeepSeek's $5.6 million claimed cost significantly underestimates true development expenses when distilled data value is included, but doesn't provide a specific alternative figure. The true cost would depend on valuing the extracted capabilities, which is difficult to quantify.
Q: Could this lead to trade restrictions or new regulations?
A: Yes. The accusation may prompt new export controls on AI technology, restrictions on API access for foreign entities, or new regulations governing how AI companies monitor and restrict their systems' use.
Q: How does this compare to previous technology competitions between the U.S. and China?
A: Like previous disputes over semiconductors, software, and telecommunications, this reflects fundamental competition over who controls advanced technology. AI's importance to economic and military power makes this competition particularly intense.
Q: Are American companies also accused of unfair practices in AI development?
A: The advisory focuses on Chinese companies, but American companies have faced scrutiny for data practices, labor issues in training data collection, and acquisition of smaller competitors' technology. The focus on distillation reflects what U.S. officials view as a specific Chinese strategy.
Sources
Federal Bureau of Investigation, National Security Agency, and Cybersecurity and Infrastructure Security Agency. "China-Based Artificial Intelligence Companies Conducting Industrial-Scale Distillation Campaigns Against U.S. AI Companies." Cybersecurity Advisory AA26-251A, September 8, 2026.
NBC News. "U.S. agencies say top Chinese AI companies systematically copied American models." September 8, 2026.
CNN. "US accuses Chinese AI firms of 'malicious' copying of AI technology." September 9, 2026.
Reuters. "U.S. accuses Chinese AI firms of 'industrial-scale' theft of AI technology." September 8, 2026.
CyberScoop. "Feds accuse China of 'systematic' distillation of U.S. AI models." September 8, 2026.
Bloomberg. "DeepSeek's low training costs questioned as U.S. alleges model distillation." February 2025.
Inside AI. "US Accuses Chinese AI Firms of Industrial-Scale Theft of AI Technology." September 9, 2026.
StartupFortune. "FBI, NSA and CISA Accuse Six Chinese AI Firms of Stealing US Models." September 9, 2026.
GBHackers. "DeepSeek, Alibaba and Chinese AI Firms Extract Billions of Tokens From U.S. AI Models." September 9, 2026.
Nextgov/FCW. "Intelligence agencies warn of China's large-scale AI model distillation efforts." September 8, 2026.
Recode China AI. "Inside the OpenAI-DeepSeek Distillation Saga & Alibaba's Most Powerful AI Model Qwen2.5-Max." January 2025.