Two scenarios for global AI leadership \ Anthropic

The global race for Artificial Intelligence is often framed as a sprint toward a finish line. But for those inside the industry, it looks less like a track meet and more like a high-stakes game of geopolitical chess. At the heart of this struggle isn’t just who has the smartest chatbot, but who controls the “compute”—the raw processing power that fuels the next generation of transformative AI.

We are approaching a critical inflection point. The decisions made today regarding export controls, hardware access, and model security will determine whether the future of AI is shaped by democratic values or authoritarian control.

The Compute Moat: Why Hardware is the Real Battlefield

In the world of frontier AI, intelligence is a function of scale. While algorithmic breakthroughs are vital, they are essentially multipliers of compute. More chips mean more experiments, which lead to better algorithms, which in turn require even more chips to implement. This creates a self-reinforcing cycle of dominance.

Currently, democracies hold a significant lead in the semiconductor supply chain. The most capable chips—the “engines” of AI—are developed by American companies. However, this lead is under constant pressure. While state-backed initiatives like “Made in China 2025” have poured billions into domestic chip production, the technical gap remains wide.

Without access to Extreme Ultraviolet (EUV) lithography and high-bandwidth memory, building a world-class AI ecosystem from scratch is nearly impossible. This makes export controls the most powerful tool in the democratic arsenal to prevent the “intelligentization” of authoritarian military forces.

Shadow Tactics: Distillation Attacks and Policy Loopholes

If the hardware door is locked, how do competitors keep up? The answer lies in a practice known as distillation attacks. This is essentially industrial espionage for the AI age.

In a distillation attack, a lab creates thousands of fraudulent accounts to query a top-tier model (like those from Anthropic or OpenAI). They harvest the high-quality outputs and use that data to train their own smaller, cheaper models. Effectively, they “distill” the intelligence of a billion-dollar American model into a fraction of the cost.

Beyond distillation, the “shadow” supply chain persists. Advanced chips are frequently smuggled into restricted regions or accessed via remote data centers in Southeast Asia, bypassing traditional sales bans. This “leakage” allows near-frontier models to emerge even when official hardware access is restricted.

The 2028 Fork in the Road: Two Possible Futures

As we look toward the end of the decade, we see two distinct scenarios for global AI leadership. The path we take depends entirely on whether we treat AI as a commercial product or a strategic national security asset.

Scenario A: The Democratic Lead

In this future, policymakers successfully close the loopholes. Smuggling is curtailed, and distillation attacks are treated as illegal industrial espionage. The result is a 12-to-24-month lead in model intelligence.

This “breathing room” allows democracies to set the global norms for AI safety and ethics. In this world, AI becomes a tool for unprecedented economic growth and scientific breakthroughs in medicine and energy, while acting as a deterrent against aggression.

Scenario B: The Neck-and-Neck Race

If current loopholes remain open or export controls are loosened, we enter a “destabilizing race.” In this scenario, AI capabilities in authoritarian regimes reach parity with the West.

The danger here isn’t just economic; it’s existential. AI can remove the need for human enforcers in surveillance states, enabling automated repression at scale. From facial recognition in Xinjiang to AI-coordinated cyber-attacks on critical infrastructure, the tools of control become pervasive and invisible.

Beyond Intelligence: The Four Fronts of AI Competition

It is a mistake to think this is only about who has the “smartest” AI. True dominance is fought across four distinct fronts:

• Intelligence: Who develops the most capable frontier models?
• Domestic Adoption: Who integrates AI most effectively into their economy and public sector? (e.g., China’s “AI+” initiative).
• Global Distribution: Who provides the infrastructure (the “AI stack”) that the rest of the world runs on?
• Resilience: Which societies can maintain political stability during the massive economic displacement AI will cause?

Intelligence is the primary driver, but adoption is the accelerant. A slightly less intelligent model that is integrated into every factory and government office can still provide a strategic advantage over a superior model that remains locked in a research lab.

The “Gatling Gun” Moment: AI as a Force Multiplier

The speed of AI advancement is no longer linear; it is exponential. We are seeing a transition toward “transformative AI”—systems that can act as a “country of geniuses in a data center.”

Consider the impact on cybersecurity. Recent previews of advanced models have shown the ability to fix security bugs at rates 20 times higher than human averages. In a military context, this is the difference between a bolt-action rifle and a fully automatic Gatling gun. When a model discovers a new software vulnerability, the regime that controls it can deploy that exploit in weeks, not years.

This acceleration makes the “window of opportunity” to secure a lead incredibly small. Once a certain threshold of intelligence is reached, AI will begin to assist in the design of its own successors, potentially locking in a lead that can never be overcome.

Frequently Asked Questions