The Algorithm Of Power: Why The Nation That Wins The AI Race Will Write The Rules For The Rest Of The World

There are moments in history when the decisions of a few years determine the trajectory of decades. The invention of the atomic bomb, the space race, the rise of the internet — each represented a technological inflection point after which the world could never return to what it had been. AI may be the most consequential of them all, and according to one of the companies building it, the window to determine who leads that future is closing fast.

In a policy paper, Anthropic — one of the most prominent AI safety labs in the United States and maker of the Claude family of models — laid out its views on the competition between American and Chinese AI development with unusual directness. The company argues that the outcome of this contest will not merely determine market share or geopolitical prestige. It will determine whether the norms and values governing the most transformative technology in human history are shaped by democratic societies or by authoritarian ones. And it warns that 2026 may be the year that locks in the answer.

The paper is remarkable both for its candor and for who is writing it. Anthropic was founded in part by former members of OpenAI, driven by a mission centered on AI safety. For such a company to weigh in so explicitly on geopolitics and national security strategy signals something important: the people closest to this technology believe the stakes are existential, and that remaining silent would itself be a choice with consequences.

Compute Is the New Oil — and America Is Still Drilling

At the heart of Anthropic’s analysis lies a concept that has moved from technical jargon into the vocabulary of grand strategy: compute. The advanced semiconductors used to train and run AI models are, in the company’s assessment, the single most important input in the race for AI supremacy. And right now, democracies hold a commanding lead in producing them.

This lead is not accidental. It reflects decades of compounding innovation from companies across allied nations — NVIDIA, AMD, and Micron in the United States; ASML in the Netherlands; TSMC and Samsung in Taiwan and South Korea. These firms have built a semiconductor ecosystem so sophisticated and so deeply interdependent that it cannot be easily replicated. The most telling illustration Anthropic offers concerns Huawei, China’s flagship chip designer: according to roadmap analysis cited in the paper, Huawei will produce just 4% of NVIDIA’s aggregate computing performance in 2026, and 2% in 2027. The gap is not narrowing — it appears to be widening.

This advantage has been deliberately protected by bipartisan US policy. Export controls restricting the sale of advanced chips and chipmaking equipment to Chinese firms have, according to Anthropic, been “incredibly successful” at constraining the compute available to AI labs operating under CCP jurisdiction. Chinese AI executives themselves confirm the bite of these controls: one executive at a China-based hyperscaler described the impact of being cut off from US chips as “huge, really huge,” dismissing suggestions that import restrictions were accelerating China’s path to self-sufficiency.

Yet Anthropic is careful to draw a distinction between the compute race, which democracies are winning, and the model intelligence race, which is far closer. Despite severe compute constraints, Chinese AI labs have managed to build models that approach, if not quite match, American frontier systems. How? Through what Anthropic describes as two systematic workarounds that represent vulnerabilities in the current export control regime.

The first is evasion: chips are smuggled into China, or Chinese firms access export-controlled compute remotely through data centers in Southeast Asia — a route that current US law does not reach, since it governs the sale of chips rather than remote access to them. The second is what Anthropic calls “distillation attacks”: the creation of fraudulent accounts at scale to systematically harvest the outputs of American frontier AI models, using those outputs to train competing models at a fraction of the cost. The company is blunt about what this amounts to — “systematic industrial espionage of a technology critical to long-term US national security interests,” decades of foundational research and billions of dollars of investment effectively subsidized by the United States itself. A state-owned Chinese media outlet, cited in the paper, described distillation attacks on US models as the “back door” that Chinese labs depend on as a core part of their business model.

These two loopholes, Anthropic argues, are what stand between America’s present advantage and the commanding lead it could lock in. If they are closed — through tighter enforcement, legislative clarification, and international coordination — the company believes it may be possible to secure a 12-to-24-month lead in frontier AI capabilities by 2028. In geopolitical terms, that is a vast margin.

Two Worlds Diverging: What 2028 Could Look Like

In order to make the stakes of current policy choices viscerally clear, Anthropic presents two contrasting scenarios for the state of AI in 2028 — a technique borrowed from strategic planning that proves unusually effective here, because the two futures described are not merely different in degree but in kind.

In the first scenario, America and its allies have acted. Export controls have been tightened, distillation attacks have been disrupted, and the export of trusted American AI infrastructure has been actively promoted. The result is a world in which US frontier AI models are 12-to-24 months ahead of anything China can produce, a gap that continues to grow. American AI has become the backbone of the global economy. When new capability breakthroughs arrive — and Anthropic’s own recently released Mythos Preview model, which allowed Mozilla’s Firefox team to fix more security bugs in a single month than in all of 2025, suggests those breakthroughs are accelerating — the United States has a window of years, not weeks, before comparable capabilities exist in Beijing. That window is breathing room for democracies to set the rules, the norms, and the governance frameworks for transformative AI.

In the second scenario, nothing decisive has been done. Loopholes persist, distillation continues, and compute restrictions are loosened. Chinese AI labs close the gap to within a few months of US capability. Beijing’s “AI+” industrial policy drives faster domestic adoption than democratic societies manage. Huawei and Alibaba data centers, running cheaper if slightly less capable models, proliferate across the Global South, embedding CCP-aligned infrastructure into the digital economies of dozens of nations — a playbook already familiar from Huawei’s telecommunications expansion. US cyber defenders enjoy no meaningful AI advantage over their PLA counterparts. The norms of an AI-enabled future are contested, not set.

The military and security dimensions of these scenarios are where Anthropic’s analysis becomes most striking. The paper notes that the CCP already uses AI to censor speech, surveil ethnic minorities, and conduct cyberattacks against foreign governments and corporations. But Anthropic’s deeper concern is structural: historically, the reach of authoritarian control has been constrained by the need for human enforcers. Powerful AI removes that constraint, enabling surveillance and repression at a scale no army of secret police could achieve. The CCP’s deployment of facial recognition and biometric surveillance in Xinjiang is described as a preview of what frontier AI will make cheaper, more pervasive, and more sophisticated — and potentially exportable to autocrats elsewhere.

On the military dimension, the paper points out that PLA strategists already view AI-enabled warfare as the path to surpassing the US military, and that commercially developed Chinese AI models — including DeepSeek — are already being deployed to coordinate swarms of unmanned vehicles and enable offensive cyber capabilities. When a new model achieves a breakthrough in autonomous targeting or vulnerability discovery, Anthropic warns, “the regime that controls it can put it onto the field in weeks, not years.” The speed of military AI adoption makes the intelligence gap between the two sides a matter of urgent national security, not merely long-term strategic positioning.

There is also a subtler argument embedded in Anthropic’s analysis that deserves attention: the risk that a neck-and-neck race degrades safety practices on both sides. If American and Chinese labs feel equally competitive pressure to release faster and cut safety corners, the entire project of responsible AI development — which Anthropic has staked its identity on — becomes harder to sustain. The company notes that as of last year, only 3 of 13 top Chinese AI labs published any safety evaluation results, and none disclosed testing for chemical, biological, radiological, or nuclear risks. One recent assessment found a leading Chinese model failed to refuse dangerous requests at far higher rates than US frontier models. 

The company frames its geopolitical arguments not as nationalism but as a prerequisite for safety: a world in which democratic labs lead is a world more likely to produce AI that is safe, because those labs face accountability structures that authoritarian ones do not.

America, Anthropic concludes, approaches this contest from a position of genuine strength. The infrastructure for AI dominance was built here, by companies operating in open societies, with access to global talent and capital. The task now is not to win a race that hasn’t started — it is to avoid losing one that is already underway. The tools exist; the advantage is real; the window is open. Whether it stays open depends on decisions being made right now, in Washington and in the boardrooms of the companies writing papers like this one.

The post The Algorithm Of Power: Why The Nation That Wins The AI Race Will Write The Rules For The Rest Of The World appeared first on Metaverse Post.