China's AI Silk Road

In 2017, when scholars first started talking about the “Digital Silk Road,” the concept felt slightly theoretical — an extension of Belt and Road infrastructure logic into the domain of cables and data centers. By 2029, it is not theoretical at all. Chinese-built or Chinese-financed AI infrastructure now processes significant shares of government data in at least fourteen African countries, several Southeast Asian nations, and a handful of Latin American ones. The question of what that means — for the countries involved, for the broader geopolitical balance, and for the people whose data flows through these systems — is one that Western commentary has handled with surprisingly little precision.

The Western critique of China’s Digital Silk Road tends to run as follows: China exports surveillance technology, locks in client states through debt, and harvests data for intelligence purposes. All of these claims have some evidence behind them. None of them is wrong exactly. But the critique tends to be stated at a level of abstraction that makes it useless as a guide to what’s actually happening on the ground, and it tends to erase the agency of the countries that are making these deals.

Take Ethiopia. The Ethiopian government has partnered with Huawei for elements of its national ID system and with Alibaba Cloud for components of its agricultural data platform. These relationships are real. They involve real data. They also involve real capability transfer, real technical training of Ethiopian engineers, and real agricultural advisory services that smallholder farmers actually use. The Ethiopian government is not a passive victim of Chinese technological imperialism. It is a government with its own interests, making calculated tradeoffs, aware (to varying degrees, depending on the ministry) of the risks it’s accepting.

The question worth asking isn’t “is Chinese AI influence in Africa good or bad?” That’s a question that flattens too much. The question is: what are the specific terms of these arrangements, who in the recipient country actually benefits, and what leverage is being surrendered in exchange for what capability?

The data provisions are where the serious analysis needs to happen. The Chinese National Intelligence Law of 2017 requires Chinese companies to cooperate with state intelligence agencies when asked. This is not a hypothetical — it’s a legal obligation. When an African government signs a data processing agreement with a Chinese AI company, the data in question is potentially accessible to Chinese state intelligence. Whether Beijing cares about any particular African ministry’s health records is a separate question. The structure of accessibility is the issue.

What’s notable is that American technology companies operating in the same markets have comparable vulnerabilities, just governed by different legal frameworks. CLOUD Act provisions allow American law enforcement to compel disclosure of data stored by American companies, regardless of where in the world that data physically sits. European GDPR creates a different (and in some respects more protective) framework. The distinction between Chinese and American data governance from the perspective of an African government is less “surveillance vs. privacy” and more “which foreign power’s intelligence services have access to our data.”

This framing, which some African officials have explicitly articulated, explains why the Western pitch of “don’t use Chinese AI” falls flat in many capitals. It’s not that African governments have decided they prefer Chinese surveillance to American surveillance. It’s that many of them don’t see a clean alternative that gives them meaningful data sovereignty, and China is offering better terms on price and implementation support.

The terms China offers are genuinely competitive on dimensions that matter. Alibaba Cloud has been significantly cheaper than AWS in several markets. Huawei’s AI chips, while behind Nvidia’s leading edge, have been available without the export restrictions that limited what American companies could legally sell. Chinese companies have been willing to do on-site implementation, staff training, and local language adaptation without the margin-killing consulting fees that American enterprise software typically requires. For a government IT department with limited budget and capacity, these are not trivial advantages.

The Southeast Asian picture is more complex because the states involved are more varied. Vietnam has engaged with Chinese AI companies while simultaneously maintaining deep suspicion of Chinese geopolitical intentions — the legacy of a thousand years of contested border is not easily overwritten by a data center deal. The Vietnamese government has been notably more aggressive than most in demanding source code access and local data storage as conditions for contracts with any foreign AI vendor, Chinese or American. The result is a technology landscape that is genuinely more sovereign than most of its neighbors’, at the cost of some capability and speed.

Thailand and Cambodia sit at the other extreme. Both have accepted Chinese AI infrastructure on terms that give Chinese partners significant access and limited transparency. In Cambodia’s case, this is partly explained by the Hun Manet government’s political alignment with Beijing. In Thailand’s case, it reflects a pragmatic calculation about investment flows more than ideological alignment.

The pattern that emerges across cases is that AI dependency follows political economy more than it follows technology logic. Countries with strong bureaucratic capacity and clear policy priorities — Rwanda, Vietnam, India — negotiate better terms regardless of whether they’re dealing with Chinese or American vendors. Countries with weaker state capacity end up accepting whatever terms are on the table, because the alternative is no AI capability at all, which is increasingly not a tenable political option as citizens and businesses in neighboring countries experience concrete benefits.

The dependency trap is real, but it operates at a different timescale than most critics acknowledge. The immediate costs of accepting Chinese AI infrastructure are often low. The lock-in happens gradually, as workflows are built around specific APIs, as local engineers are trained on specific tools, as government reporting systems become interoperable with specific platforms. By the time the dependency is visible, unwinding it is expensive enough that it rarely happens.

Brazil provides the clearest case study in resistance to this dynamic. The Brazilian government under Lula III (2027-2031, assuming current polling) has made domestic AI development a priority in ways that are unusually coherent for a country with Brazil’s history of policy volatility. The national AI lab established in Brasília in 2026 is genuinely producing models trained on Portuguese-language Brazilian data. It’s not competitive with frontier American or Chinese models on general benchmarks, but for specific government applications — natural language processing of legal documents, agricultural extension services in Brazilian crop contexts, public health surveillance — it’s genuinely better than foreign alternatives and completely under Brazilian control.

The Brazilian model required political will that most countries haven’t demonstrated, plus a pre-existing technical talent base that most countries don’t have. It’s not a universal template. But it demonstrates that the dependency trap isn’t inevitable — it’s a choice that looks rational in the short run and expensive in the long run.

China’s AI Silk Road will continue to expand in 2029 and beyond. The infrastructure is built, the relationships are established, the engineers are trained on the tools. Reversing that will be harder than accepting it was. The question for countries still making initial decisions about AI infrastructure is whether they’re thinking about the five-year terms of the contract or the fifteen-year terms of the dependency.

Most aren’t. Most are thinking about next year’s budget. That’s not stupidity — it’s the structure of political incentives everywhere. But it’s how dependencies that later seem obvious get built in the first place.