The AI Race to Regression

Summary of Zeihan’s Argument

Zeihan argues that although AI seems to be advancing very quickly right now, this progress will soon slow and then reverse for most people. His core point is that hardware—not software—is the true bottleneck for AI.

He notes that newer models like GPT-5 may feel worse for casual users, but they were designed mainly for researchers, coders, and industrial uses. In other words, they are “frontier models”—the most advanced AIs at any given moment, trained on huge datasets with thousands of specialized chips. Frontier models require extremely powerful hardware to train and run, far beyond what consumer devices can handle. They are expensive to build and rely on complex global supply chains.

Zeihan’s overarching concern is that the chip supply chain needed for frontier AI is fragile. Advanced semiconductors require over 100,000 production steps, thousands of companies, and highly specialized components. Many of these steps have single points of failure across multiple countries. He believes geopolitical fragmentation will disrupt this system.

He also mentions that chips in data centers only last a few years. If the world loses the ability to produce cutting-edge chips, AI capabilities will erode as old chips fail. He argues that government attempts to “re-shore” semiconductor manufacturing focus too narrowly on fabrication plants, ignoring the upstream and downstream components that make the whole system work.

His conclusion is that we will likely see a decline in advanced AI availability within the decade, with access narrowing to only governments and mega-corporations that can secure their own hardware. He sees this as a coming “technological regression.”

ChatCPT’s Analysis: Where Zeihan Is Right and Where He Overstates the Case

Hardware as the bottleneck

Zeihan is right that hardware is currently the limiting factor. Frontier models—those at the very leading edge—need enormous computing power. They consist of huge neural networks trained on trillions of words and images, using thousands of power-hungry GPUs. These systems can do advanced reasoning, long-context analysis, multimodal tasks, and scientific problem solving. They are far beyond the capabilities of local open-source models that can run on a laptop or phone. Because frontier models are so large and specialized, only a handful of organizations can build them, and they do need very advanced chips.

Supply chain fragility

He is also correct that semiconductor production is globally distributed and complex. Lithography machines from the Netherlands, ultra-pure materials from Japan, design tools from the US, fabs in Taiwan and Korea, and packaging in Southeast Asia all tie together. Disruption in any of these areas could affect production. Some components truly are single-source.

Where the argument becomes overstated

Where Zeihan overreaches is in assuming that geopolitical fragmentation will cause a near-total collapse of advanced chipmaking. He treats globalization as something that might disappear entirely. In practice, the semiconductor ecosystem is consolidating into stable blocs rather than dissolving. The US, Taiwan, Japan, South Korea, the Netherlands, and parts of Europe already form a highly coordinated network with enormous investment in redundancy. These countries account for the vast majority of advanced semiconductor capability and are actively expanding it.

Government policies in the US, EU, Japan, and South Korea are not just building fabs—they are pushing upstream chemicals, photoresists, power-electronics plants, packaging facilities, and technician training. No one is trying to put the entire ecosystem “under one roof” because that is not how the industry has ever worked, even in high-tension geopolitical eras. The system is diversifying, not collapsing.

AI regression is unlikely

Zeihan’s prediction that AI access will shrink dramatically is also not consistent with current trends. AI access is becoming stratified rather than endangered. Frontier models will remain expensive and run in specialized data centers. But mid-sized models are rapidly becoming cheaper and more powerful, and small models running locally on phones or laptops are improving very quickly.

Instead of regression, we are likely to see a tiered AI ecosystem: extremely advanced frontier systems for large institutions, powerful enterprise models for businesses, and efficient open-source models for everyone else.

The real constraints and threats to AI progress:

1. Energy constraints: Data centers need absurd amounts of power; this is the most urgent bottleneck.

2. Fabrication capacity: Even with allies intact, the world simply isn’t building fabs fast enough.

3. Cooling, real estate, and power density: Massive practical limits in hot regions or overloaded grids.

4. Skilled labor shortages: Chip fabs need thousands of highly trained technicians.

5. Geopolitical disruption in Taiwan: This is the true high-impact scenario—but the U.S., Japan, and Taiwan are spending billions to build redundancy because they know this.

These risks could slow AI progress and shape who can build frontier models, but they do not point to a collapse.

Overall View

AI progress will slow in some ways as hardware and energy constraints tighten, but it will not reverse. Frontier models will remain large, expensive, and centralized, while open-source and enterprise models become more efficient and widely available. The world is not heading toward an AI regression; rather, AI will continue to advance but will be unevenly distributed, with different tiers of capability emerging for different users.