What are the U.S. and China offering other countries in the AI race, and what does it do to sovereignty?

Most talk of sovereign AI imagines a country deciding alone. In practice, almost every country is choosing among options offered by two powers. The most useful question is often less "can we be independent?" than "whose dependency are we accepting, and on what terms?"

A warning about the term itself. Nvidia, which sells the chips, publishes its own explainer on sovereign AI and promotes national AI factories as "the bedrock of modern economies". Selling independence is good business, and buyers should read the pitch with that in mind.

The American offer: a full stack, with conditions

Executive Order 14320, signed July 23, 2025, sets out the policy: extend American leadership and "decrease international dependence on AI technologies developed by our adversaries". It created the American AI Exports Program, and Commerce ran its first call for proposals from April 1 to June 30, 2026.

Consortia of companies can propose a full-stack package, meaning everything from chips and data centres through models to applications. Designated packages may get government advocacy and other federal help. The vocabulary is striking. White House science adviser Michael Kratsios spoke of "sharing a trusted, sovereign AI stack with the world". A foreign stack that you call sovereign is exactly what critics in Canada and Europe object to, while supporters would say it is simply the fastest way to get capability.

The conditions show in chip rules. In the Gulf, approvals for chips came with security and reporting duties, and the UAE's July 2026 upgrade applies only to approved companies, with G42's authorization set to lapse after 270 days without further notice. Our article on the Gulf buildout covers how this works.

The American stance also reaches allies' own policies. RBC reports that in March 2026, the U.S. Trade Representative listed Canada's sovereign computing initiative among trade barriers. So one arm of Washington promotes sovereign AI abroad as an export, while another objects when a partner's version favours its own suppliers.

The Chinese offer: open models and a seat at the table

China's approach mixes technology and institutions. At the World AI Conference in Shanghai in July 2026, President Xi Jinping highlighted the World Artificial Intelligence Cooperation Organization (WAICO). Representatives of 29 countries signed the agreement establishing it on July 16, reports CSIS, an American think tank. It sits outside the UN system and no G7 economy or EU member joined. Xi also promised 5,000 AI training places for developing countries over five years and cooperation centres with regional groups.

The technology side matters at least as much. CSIS describes China offering its regulatory model "paired with open-weight models, AI training programs, infrastructure, and turnkey technological solutions". An open-weight model is one whose trained numbers you can download and run on your own hardware, which is attractive to anyone worried about being cut off. It need not come with a monthly bill or a foreign server.

That attraction has limits. Running a Chinese open model removes some dependency, but its training choices, including whatever it was taught to avoid saying, come with it, and you still need chips. Beijing's stated aim of equity for the Global South is also a bid for influence, as CSIS argues in comparing WAICO with the Belt and Road Initiative.

Side by side

Some buyers mix the two where permitted. Humain's CEO says it will not let Chinese frontier labs train on its systems, but it will run Chinese open-source models for inference within its American framework.

The middle-power response

Between the two sit Canada, Germany, France, Japan, India and others. Their answer is increasingly to band together. Canada's AI for All puts it as "build, partner, buy", and Prime Minister Carney speaks of partnering with reliable allies. Canada and Germany launched a Sovereign Technology Alliance, and Cohere's merger with Aleph Alpha turns that into a company. RBC's analysts define sovereignty in this setting as freedom from coercion, including the ability to substitute providers under pressure, rather than isolation.

Even so, realism is widespread. An academic review in Natural Language Processing notes that self-sufficiency across the stack is hard for even the two superpowers: Nvidia and Huawei both outsource fabrication. And the Tony Blair Institute's work on the topic, as summarized by the Asian Development Bank's knowledge hub, calls for "deliberate interdependence". On that view, the real skill is choosing dependencies you can live with and can leave.

How should a mid-sized country or company weigh these offers?

No one answer fits, and this guide does not tell governments which side to pick. A useful way to think is to ask four questions.

What are you afraid of? A cut-off by a foreign government, a legal order for your data, a sudden price rise and a model that reflects foreign politics all call for different fixes.

Which layer can you actually control? For most organizations, it is the model and the application. Downloading open weights is within reach, while making chips is not.

How easily could you leave? Prefer open formats and contracts that let you take your data and swap models. A company can test this now by running one workflow on a second model.

Who else is checking the work? Test any model, American, Chinese or domestic, on the questions that matter to you, including sensitive ones. Compare answers before you rely on it.

For a small business, the answer may simply be a mainstream assistant with sensible settings, plus a tested open-weight fallback. For a government, it may be a portfolio of suppliers, reviewed as the offers change.