The United States just redrew the map of the world. Not with armies. With chip exports.

In January 2025, the Biden administration finalized what it called the "AI Diffusion Rule" a three-tier system that determines which countries can buy the advanced AI chips that power frontier models. Tier 1: close allies. Eighteen countries, mostly NATO plus Japan, South Korea, Australia. They get unrestricted access. Tier 2: everyone who needs a license and a limit. Tier 3: arms embargo territory China, Russia, Iran, North Korea. Can't buy at all.

Most of the world is Tier 2. That sounds fine until you realize what it means in practice: capped chip imports, government-to-government approvals, quarterly reporting requirements, and enough uncertainty to make any serious AI investment a gamble. A startup in Nairobi can't just order H100s. A research lab in São Paulo has to navigate export licensing that was designed with Beijing in mind. An Indonesian cloud provider has to explain to the Commerce Department why it needs GPUs and then wait.

The framing is national security. The effect is economic zoning. Countries that were never a threat to American AI dominance are now subject to the same friction as countries that are. And friction, in technology markets, is fatal. It doesn't slow you down. It routes around you.

We built new borders. Not on land. On compute.

In Part 1 of this series, we looked at the talent divide how AI skills are concentrating wealth in a shrinking pool of workers. In Part 2, we traced the enterprise gap how implementation capacity is separating companies that adopt AI from companies that just buy it. Now zoom out further. The same vertical wall runs between countries, and the bricks are policy.

The Scoreboard That Lies

The US-China AI race makes great cable news. Two superpowers. Dueling models. Competing chip fabs. It's tidy. It's dramatic. It's also a distraction from the actual story.

Here's what the scoreboard shows: the US added 42 Chinese entities to the Entity List in March 2025. Another 23 in September. Nvidia's most advanced chips can't ship to Chinese buyers. TSMC won't fabricate certain designs for Chinese firms. ASML can't sell its best lithography machines to anyone the US flags. The squeeze is real.

Here's what the scoreboard hides: those same controls ripple outward. The licensing regime doesn't just target Chinese companies. It creates friction for every country that isn't in the inner circle. A Thai university that wants to build a small GPU cluster for natural language research faces the same bureaucratic maze that was built to slow Huawei. The controls don't distinguish between a Chinese military lab and a Brazilian agritech startup. They distinguish between countries the US trusts and countries it doesn't and "doesn't trust" is a very long list.

The Carnegie Endowment published a report in July 2025 that said the quiet part out loud: US AI strategy has "mostly overlooked" the Global South. Not as a bug. Not as an oversight they're fixing. As a structural feature of how Washington thinks about AI which is to say, Washington thinks about AI as a competition with one opponent, and everyone else is scenery.

One hundred and fifty countries aren't in the AI race. They're not even in the stadium.

Europe's Self-Inflicted Wound

If export controls are the wall America built, the EU AI Act is the wall Europe built around itself.

The numbers are brutal. Compliance with the EU AI Act costs approximately €29,277 per year per AI unit for ongoing obligations. Certification fees run €16,800 to €23,000 per system. For a mid-size company deploying AI across a few business functions, you're looking at six figures in regulatory overhead before the AI does a single useful thing.

Consider a real scenario playing out right now. A French healthcare startup builds an AI diagnostic tool. Good product. Real clinical value. To deploy it in the EU, they need to classify the system's risk tier, complete a conformity assessment, register in the EU database, implement continuous monitoring, and pay for third-party certification. Annual cost: north of €29,000 per unit deployed.

Their American competitor? Ships it. FDA has its own process, sure, but there's no equivalent of the AI Act's horizontal regulatory layer sitting on top. No per-unit compliance tax. No mandatory risk classification for every AI system regardless of domain. The American company moves faster, spends less on lawyers, and prices the French startup out of its own market.

This is not theoretical. European AI investment is already lagging. The continent that gave us CERN, the World Wide Web, and half of modern mathematics is regulating itself out of the AI era. The EU AI Act was written with good intentions transparency, safety, human rights. But the implementation creates a cost structure that functions as a wall. Not against bad AI. Against all AI that isn't backed by enough capital to absorb the compliance tax.

Big Tech can pay €29K per unit without blinking. Google's compliance team has more people than most European AI startups have employees. A ten-person company in Lyon cannot absorb that cost. They either raise prices (and lose to the American competitor), relocate (and drain European talent), or shut down. The regulation meant to protect Europeans from AI consolidation is accelerating it.

We'll dig deeper into the regulatory dimension in Part 4: The Safety Tax. For now, the point is this: Europe isn't being walled off by America or China. It's walling itself off.

Left Behind (Again)

Africa's internet penetration sits at roughly 43%. That's the baseline not AI readiness, just the ability to get online. AI adoption is a fraction of that fraction.

This matters because AI isn't a gadget. It's infrastructure. Like electricity. Like telecommunications. Every previous wave of transformative technology followed the same pattern: rich countries deployed it first, poor countries got it late and on worse terms, and the gap between them widened in ways that took decades to partially close. AI is doing the same thing, faster.

But this time there's a twist. The previous technology gaps were mostly about money and logistics laying cable, building towers, manufacturing devices. AI has all of those barriers plus a new one: policy. Export controls mean you can't just buy your way to compute. The three-tier system means your country's geopolitical alignment determines your AI ceiling. A wealthy Nigerian entrepreneur with the capital to build a data center still can't buy the same chips as a wealthy Canadian entrepreneur. Not because of money. Because of passport.

So what do developers in Lagos do? They build on what's available. Right now, that means Meta's Llama models and increasingly DeepSeek. Open-source, open-weight, and most critically, open-access. No export license required. No per-unit compliance fee. Download, fine-tune, deploy.

A Nigerian developer building a Yoruba language model doesn't need H100s. They need a decent GPU and an open model they can fine-tune. Claude and GPT-4 APIs cost real money money that goes further in Lagos than San Francisco, which means it hurts more. Llama is free. DeepSeek is free. The decision isn't ideological. It's arithmetic.

The Global South isn't choosing Chinese AI because of Chinese soft power. They're choosing it because America and Europe priced themselves out of the conversation.

DeepSeek Changes the Math

On January 27, 2025, DeepSeek released R1. Open-source. Competitive with the best Western models on major benchmarks. Built at a fraction of the cost that OpenAI and Anthropic spend on training runs.

The reaction in Silicon Valley ranged from denial to panic. Nvidia's stock dropped. The White House scrambled. The entire thesis of the export control regime that restricting chips would maintain American AI dominance took a direct hit.

DeepSeek didn't just build a good model. They proved that compute constraints breed innovation. The Chinese AI ecosystem, squeezed by export controls, didn't wither. It adapted. It found efficiencies. It invented new training techniques mixture of experts, novel attention mechanisms that squeezed more capability out of less hardware. It built something competitive using fewer and older chips. And then it open-sourced it, which meant everyone else could use it too.

Think about what that signals to a government in Jakarta or Accra or Bogotá. The Americans say: buy our chips (if we let you), use our cloud (at our prices), on our terms. The Chinese say: here's a model. It's free. Do what you want with it. One pitch requires navigating a bureaucracy. The other requires a download link.

By August 2025, Chinese AI models had surpassed Western counterparts in cumulative open-source downloads, according to the Australian Strategic Policy Institute's Critical Technology Tracker. Not in one category. Across the board. The models that the Global South is actually downloading and deploying are increasingly Chinese.

Look at Southeast Asia. Vietnam and Thailand are adopting DeepSeek for local language processing Vietnamese, Thai, Lao languages that Western models handle poorly because the training data skews English. DeepSeek's multilingual capabilities, combined with the fact that it's free and open, make it the obvious choice. Not the ideologically convenient choice. The practical one.

This is the part that should keep Washington up at night. Export controls were supposed to create dependency on American AI. Instead, they're creating dependency on Chinese AI not through coercion, but through access. America is making its AI products harder to get. China is making its AI products easier to get. The market is responding exactly the way markets respond.

The Crack in the Wall

Here's the take, and it's not comfortable for anyone.

The US built export controls to slow China's AI development. Those controls also walled off the Global South not as a target, but as collateral damage that nobody in Washington thought much about. The Carnegie Endowment noticed. The countries affected definitely noticed. Washington, mostly, has not.

Europe built the AI Act to protect its citizens. Those regulations also created a compliance wall that's pricing European startups out of their own market while American and Chinese competitors operate under lighter frameworks. The intent was safety. The effect is a competitive moat for incumbents.

And then DeepSeek showed up.

Open-source. High-quality. Free. Built by the very country the export controls were designed to contain. Using fewer chips than the controls were designed to restrict. Available to every developer in every country regardless of tier classification.

The first crack in the vertical wall came from inside the wall. That's not irony. That's the predictable result of building a wall in a world where information wants to move.

The US-China AI rivalry gets the headlines. But the real story of AI geopolitics in 2025 isn't the race between two superpowers. It's the 150 countries watching from outside the stadium, finding their own way in through whatever door is open. Right now, that door has Chinese characters on it.

This doesn't mean Chinese AI is better. It doesn't mean export controls are pointless. It means the strategy has a massive blind spot: you can't maintain AI leadership by restricting access to your own products while your competitor gives theirs away. That's not a technology strategy. That's a market evacuation.

The talent divide from Part 1 splits workers within countries. The enterprise gap from Part 2 splits companies within markets. The country divide splits the entire world into haves and have-nots and the line isn't drawn by capability or ambition. It's drawn by export policy and regulatory overhead.

Next week in Part 4, we'll look at the safety tax how the push for AI regulation, from the EU to California, is creating new costs that fall hardest on the smallest players. The wall keeps getting taller. The question is whether anyone's building a ladder.

If you work in policy: look at the three-tier map and ask who it actually serves. If you run a company competing globally: the AI toolkit available in Tier 2 and Tier 3 countries is about to look very different from yours and it won't be American. If you build AI: the next billion users aren't coming through your API. They're coming through open-source models you don't control.

The wall is vertical. But the cracks run horizontal. And they're spreading.