Reading 2028: Two scenarios for global AI leadership

Original article: "2028: Two scenarios for global AI leadership"

The most useful thing about this article is its structure. It does not treat AI competition as a vague race between companies or a simple comparison of model benchmarks. It turns the issue into a system: compute, export controls, model distillation, domestic adoption, global distribution, and AI safety all reinforce one another.

Anthropic presents two possible futures for 2028. In the first, the United States and its allies preserve a strong compute advantage, close loopholes around chip access, deter large-scale distillation attacks, and accelerate AI adoption across democratic countries. In that world, democracies have more leverage to set norms for frontier AI. In the second, restrictions weaken or enforcement fails, Chinese AI labs remain close to the frontier, and lower-cost infrastructure spreads through global markets. In that world, authoritarian systems gain more influence over AI deployment and governance.

My first reaction is that this is not just a technology essay. It is a warning about what could happen if authoritarian countries gain access to the state of the art in AI. The danger is not only that they may build better products. The danger is that they may gain a powerful new tool for surveillance, censorship, military coercion, cyber operations, and social control.

What I agree with

I agree with the article's focus on compute. Discussions about AI often stay at the level of model quality, product design, or open-source culture. But frontier AI depends on chips, data centers, energy, engineering talent, and capital. Compute is not just another input. It determines how many experiments a lab can run, how quickly it can improve, and whether it can stay near the frontier at all.

The article also makes a strong point that AI capability does not enter society in isolation. It flows into cybersecurity, military systems, scientific research, finance, manufacturing, and government operations. A lead in AI can become a lead in many other domains. If authoritarian countries control the most advanced AI systems, the consequences could be disastrous not only for the people living under those regimes, but also for democratic countries.

That is why strict export controls on advanced compute and EUV lithography equipment matter. These are not ordinary commercial goods. They are part of the infrastructure that makes frontier AI possible. Treating them like normal products in a normal market badly misunderstands the stakes.

I also find the concern about a neck-and-neck race persuasive. If both sides believe they are only months apart, safety evaluations and release discipline become harder to defend. The pressure to move first does not only affect states. It affects companies too.

Where I am more skeptical

The article divides the world into democratic and authoritarian AI futures. I think that distinction is important, but it is not sufficient.

LLMs are not developed by "democracies" in the abstract. They are developed by companies controlled by small groups of executives, investors, and researchers. Democratic leadership does not automatically guarantee safe or humane AI. Companies still pursue profit. Governments still expand surveillance capacity. Platforms can still use AI to shape information environments.

So I agree with the article's main concern, but I also want to ask a second question: if democratic countries lead, who supervises the leaders? Winning the AI race is not enough. The winners must also be constrained.

Objections I have seen

One objection is that export restrictions may push Chinese companies to develop their own GPUs faster. Jensen Huang has made versions of this argument in public discussions, including this one: YouTube.

I understand the concern, but I do not find it decisive. It is possible that restrictions create incentives for domestic alternatives. But incentives are not the same as capability. Advanced chips require an entire ecosystem: design software, manufacturing equipment, process knowledge, supply chains, packaging, memory, data centers, and years of accumulated engineering experience. Export controls may not stop progress forever, but they can slow it down. In a fast-moving field like AI, buying time matters.

Another objection is that export restrictions violate the free market. I think this argument fails because frontier compute is not an ordinary commodity. We do not allow nuclear weapons, advanced missile systems, or dangerous biological materials to be traded freely just because markets are efficient. Some technologies are too dangerous to be governed by ordinary market logic. Frontier AI compute belongs in that category.

Free markets are valuable when the goods being exchanged are ordinary goods. But when a product can strengthen mass surveillance, autonomous cyber operations, military decision-making, and authoritarian control, the market must be limited by public responsibility.

What this changes for me

The article reminds me not to understand AI only through products. A model in a browser looks like a simple interface, but behind it are supply chains, industrial policy, energy systems, data centers, research institutions, and national strategy.

It also changes how I think about the responsibility of AI researchers. Technical work is never completely separate from politics. If a researcher helps an authoritarian state build more advanced AI, that work may later be used for censorship, repression, cyber attacks, or military power. Higher pay does not erase that responsibility.

AI researchers should not work only for money or prestige. They should ask who benefits from their work, who may be harmed by it, and what kind of world their research strengthens. If frontier AI becomes one of the most powerful technologies in history, then working on it is not morally neutral.

My conclusion

The article is valuable because it shows that AI development is now shaped by political and economic choices as much as by technical progress. Compute policy, export controls, model access, safety standards, and researcher decisions all matter.

My current takeaway is strong: frontier AI leadership should stay in the hands of societies that can still argue, criticize, investigate, vote, and impose limits on power. That does not make democratic countries perfect. It does mean they have mechanisms for correction that authoritarian systems often suppress.

The strongest AI systems should not be built for repression. They should be built under institutions that can be questioned, audited, and held accountable.