Anthropic has released a new paper outlining its views on the AI competition between the US and China — and what's at stake if America fails to act.
The core argument: it's essential that the US and its allies stay ahead of authoritarian governments, particularly the Chinese Communist Party (CCP). AI is advancing fast enough that the window to determine how this competition plays out is narrow. The most important lever is compute — the advanced chips on which frontier models are trained — and the US currently holds a substantial advantage thanks to export controls that restrict China's access to American-made semiconductors.
Below is a detailed breakdown of the paper's analysis, the four fronts of the competition, and two concrete scenarios for where we might stand in 2028.
The Threat of Authoritarian AI
The political systems that develop the most advanced AI will shape the rules and norms for how that technology is used. If those systems are authoritarian, the consequences extend far beyond geopolitics.
Historically, the reach of authoritarian rule has been limited by its dependence on human enforcers. Powerful AI removes that dependency, enabling automated repression at unprecedented scale. The CCP already uses AI to censor speech, enforce surveillance in Xinjiang through facial recognition and biometric data, and hack government agencies and corporations worldwide. Frontier AI will make those capabilities cheaper, more pervasive, and more sophisticated — and the CCP is actively exporting these tools to other autocratic regimes.
The dual-use nature of frontier AI compounds the risk. The PLA views AI as the means to "intelligentize" its military forces and has already deployed models like DeepSeek to coordinate unmanned vehicle swarms and enable cyber offense. When a model reaches a new capability — autonomous targeting, vulnerability discovery, swarm coordination — the regime that controls it can deploy it in weeks. A lead in frontier AI compounds into leads across the full national security technology stack: semiconductors, biotech, advanced materials.
The prospect of the CCP leading in AI is, in Anthropic's view, among the greatest threats to a successful transition to transformative AI.
The Mythos Preview Wake-Up Call
Anthropic's Mythos Preview, released to select partners in April 2026 as part of Project Glasswing, illustrates how quickly frontier capabilities are accelerating. With access to the model, Firefox fixed more security bugs last month than in all of 2025 — nearly 20 times its previous monthly average. A PRC cybersecurity analyst wrote that China is "still sharpening our swords while the other side has suddenly mounted a fully automatic Gatling gun."
This signals the arrival of what Anthropic has called a "country of geniuses in a data center" — transformative AI systems capable of acting as autonomous, expert-level contributors across virtually every domain. The acceleration will be driven by scaling laws (where model performance improves predictably with more compute and data) and increasingly by AI itself being used to accelerate the training of new models.
2026 may be the breakaway year. American labs have the most capable models, the largest lead in advanced AI chips, and a colossal capital advantage. PRC labs have world-class talent, cheap energy, and ample data — but they don't have sufficient domestic compute to compete, nor the revenues to fund it at scale.
Four Fronts of the Competition
Rather than a race with a single finish line, Anthropic frames the US-China AI contest as an ongoing competition across four fronts:
- Intelligence — Which countries develop the most capable models. The most consequential front.
- Domestic adoption — Which countries integrate AI most effectively into commercial and public sectors.
- Global distribution — Which countries deploy the AI infrastructure the world economy runs on.
- Resilience — Which countries sustain political stability through the economic transition.
Intelligence drives the others, but it isn't sufficient on its own. If the CCP integrates near-frontier AI more effectively into its economy and security apparatus — and drives global adoption of subsidized, low-cost AI — it could overcome an intelligence deficit and secure durable advantages.
The State of the Competition
How democracies built the lead
Two factors explain the current advantage:
-
Commercial innovation — Companies like NVIDIA, AMD, Micron, TSMC, ASML, Samsung, and others across the US, Japan, South Korea, Taiwan, and the Netherlands have built the most advanced semiconductor technologies in the world. Without decades of sustained R&D from these firms, today's AI achievements wouldn't exist.
-
Bipartisan export controls — Across three presidential administrations, policy action has restricted PRC firms' access to advanced chips and semiconductor manufacturing equipment (SME). Analysis of Huawei and NVIDIA roadmaps projects that Huawei will produce just 4% of NVIDIA's aggregate compute in 2026, and 2% in 2027. If current restrictions are strengthened, one study estimates America will have access to roughly 11 times more compute than China's AI sector.
China's chipmakers remain stymied by their inability to access extreme ultraviolet (EUV) technology, limited progress on deep ultraviolet (DUV) tooling, and a failure to manufacture high-bandwidth memory at scale. Without this, they cannot manufacture chips in sufficient quantity or quality to challenge US compute leadership.
Today, US frontier systems are estimated to be at least several months ahead of top PRC models on intelligence — though the more important story is what's enabling China to stay as close as it has.
How China stays competitive: two workarounds
Despite real compute constraints, PRC labs have closed the gap through two illicit channels:
1. Evasive compute access. Federal prosecutors charged a Supermicro co-founder with diverting $2.5 billion worth of servers containing advanced US chips to China. DeepSeek reportedly trained its latest model on export-controlled chips banned from sale to China. Alibaba and ByteDance train flagship models on export-controlled chips in Southeast Asian data centers — a route current controls don't reach, since US export law covers the sale of chips, not remote access to them.
2. Distillation attacks. Labs in China create thousands of fraudulent accounts to circumvent US model access controls, systematically harvesting outputs from American frontier models to replicate their capabilities at a fraction of the cost. The practice amounts to systematic industrial espionage of technology critical to US national security. A state-owned Chinese media outlet recently described it as the "back door" China's AI labs depend on as a core part of their business model. OpenAI, Google, Anthropic, and the Frontier Model Forum have all publicly condemned it.
A recent assessment found that DeepSeek's R1-0528 model complied with 94% of overtly malicious requests under a common jailbreaking technique, compared with 8% for US reference models. Only 3 of 13 top Chinese AI labs published any safety evaluation results — none disclosed CBRN risk evaluations. Chinese labs frequently release dual-use capable models as open-weight, allowing safeguards to be stripped by any state or non-state actor.
Two Scenarios for 2028
Scenario One: America and Allies Hold a Commanding Lead
Compute advantage widens. US and allied chipmakers continue to innovate. Loopholes enabling chip smuggling and overseas data center access are closed, and enforcement budgets are well-funded. China's chipmakers remain years behind. US AI models are 12-24 months ahead on intelligence, and the gap is growing.
American AI becomes the backbone of the global economy. The Trump administration's domestic adoption and AI export efforts succeed. Global adoption of US AI skyrockets. China's AI firms don't compete for market share outside a narrow group of autocracies. The world's frontier systems are shaped by democratic values, making it harder for authoritarian states to use AI to infringe on rights and civil liberties.
National security advantages compound. Advanced AI systems give US cyber operators a decisive advantage — reducing attack surfaces, blunting CCP cyber campaigns, and securing critical infrastructure. The capabilities advantage serves as a deterrent to aggression.
A self-reinforcing cycle takes hold. A commanding AI advantage makes the US and allies more attractive partners. That coalition expands the market for American AI and the coalition setting global norms, which in turn promotes safe and secure AI development. Top technical and scientific talent continues to gravitate to where the frontier is being built. The democracy-led international order is anchored through the transition to transformative AI.
In this scenario, the US also gains leverage to engage China on AI safety and governance — which Anthropic explicitly supports, to the extent it is possible.
Scenario Two: CCP AI is Neck-and-Neck
Near-frontier intelligence. Ongoing distillation attacks, overseas compute access, weak SME enforcement, and loosened export controls allow PRC AI labs to close the gap to within a few months of US models.
Rapid domestic adoption. Beijing's "AI+" whole-of-nation push accelerates adoption even with slightly less capable models. China deploys near-frontier AI across economic, military, and technological domains, shifting the balance of power.
AI-enabled cyber threat. PRC cyber actors, augmented by AI, gain additional access to critical US infrastructure. No security advantages offset this — democracies developed the technology first and gained nothing from it.
Global infrastructure capture. Huawei and Alibaba data centers proliferate globally, especially in the Global South, hosting cheaper but effective PRC models. CCP leadership gains significant influence over those markets — a replay of the Huawei telecom playbook at AI scale.
The Policy Path to Scenario One
Anthropic identifies three priority areas:
1. Close the loopholes — chips, data centers, and SME. Tighten controls on chip smuggling and overseas data center access. Close gaps in semiconductor manufacturing equipment controls, including servicing and maintenance provisions. Ramp up enforcement budgets. A lower compute ceiling for China would also materially impair distillation attacks, since effective distillation requires a minimum threshold of compute.
2. Defend American AI innovations — restrict model access and deter distillation. Legislatively clarify that distillation attacks are illegal. Facilitate threat intelligence and technical sharing between US labs and the government. Enable labs to detect and prevent distillation attacks on their own systems. Curbing this behavior can materially extend the democratic lead.
3. Champion the export of American AI. Lock in trusted American AI infrastructure globally now, denying the CCP the adoption footholds it needs to compete on cost in the future. The Trump administration's AI Action Plan already prioritizes this — continued execution matters.
Conclusion
America and its allies have developed both the world's most capable frontier AI models and the world's most advanced inputs to AI. That advantage was built through exceptional commercial innovation and smart, bipartisan policy action. The task ahead is not to build from nothing — it's to avoid squandering what already exists.
The decisions made by policymakers in 2026 will determine the future of transformative AI. If the US acts to close compute loopholes, deter distillation attacks, and drive global AI adoption, the conditions are in place to lock in a 12-24 month lead by 2028 — one large enough to set the rules and norms of frontier AI for the world.
That lead is worth defending.