Neutrality in AI is turning into a luxury few countries can afford

Can you be a friend of both AI superpowers forever? The honest answer, in our view, is no. Hedging is a fine strategy until someone sends the bill, and that bill is starting to arrive.

AP reports that the United States and China are each assembling an AI bloc: the U.S.-led Pax Silica and the China-inspired World AI Cooperation Organization. Experts quoted say the window for choosing freely may be narrowing. The U.S. State Department has already said that countries "can't be trusted partners in Pax Silica while substantively engaging in initiatives that advance a competing vision for super intelligence."

Kazakhstan is the case study. Its AI minister visited Washington in June to join the U.S. group, then headed to Shanghai about three weeks later to make the country a founding member of the Chinese body. It also has an investment deal involving Nvidia worth up to $10 billion for a "data center valley," while Chinese investment in the country hit a record $2.8 billion in 2025, according to local officials. A former Kazakh diplomat summed it up: "Everyone is hedging."

Hedging works only while nobody enforces the rules

Bruegel notes that the original Pax Silica declaration, signed by seven countries in December 2025, said nothing about choosing between Washington and Beijing. By August, the State Department was reportedly telling members they could not also back "duplicative initiatives." The terms changed after people signed. Reuters headlined the same shift: the U.S. will tell partners they must pick sides.

That is the pattern we expect to repeat. Neutrality survives while the superpowers need your goodwill. Once supply is scarce, it becomes a negotiating chip against you. The Carnegie researchers cited by AP put it bluntly: for countries like the U.K. and Germany, access to advanced AI depends on "the whims of policymakers in Washington and Beijing."

The real choice is who holds the off switch

"Picking a stack" sounds like a procurement decision. It isn't. Chips, models, cloud and standards tend to come bundled, and each layer is a point where someone else can say no. Rest of World quotes Carnegie's Sam Winter-Levy saying the U.S. and China control 90% of global computing power, and that both have shown willingness to use access as leverage. He also points to U.S. labs moving toward managed access, with companies picking who gets a model first.

Seen that way, the US offer is the most advanced chips and frontier systems, with strings attached. China's offer is cheaper, open-source models and a seat at the table where standards get written. Those are different kinds of dependency. A closed service can be cut off. A downloaded open model can't be recalled, as Bruegel points out about DeepSeek, Qwen and Kimi. For a cash-strapped government, that difference is not academic.

The line between the blocs is blurrier than the rhetoric

Here we part ways with the doom narrative. The AEI describes a "messy, uneven bipolarity" and argues the two ecosystems aren't self-contained: a Chinese rare-earth embargo would throttle US data-center construction, and Chinese open models are attractive even in advanced economies. Saudi Arabia's Humain unveiled a platform built on a Chinese open model, while the kingdom also has a deal with Nvidia to expand its AI computing, per AP. And a Kyoto professor told AP that Japan, closely aligned with Washington, will not reject cheaper Chinese models for "lower-risk applications."

Our reading: the winners will not be the neutral, but the deliberate. Use cheap open models where the stakes are low, keep critical workloads on stacks you can contractually and legally rely on, and avoid being locked into a single vendor's roadmap. France's attempt with Mistral, whose new model AP says outperforms some advanced Chinese models, is the costly but coherent version of that approach. Most countries can't afford it.

Companies face a smaller version of the same problem. If your product depends on one bloc's models, chips or cloud, you've inherited that bloc's politics. We think it's wise to design for swappable models now, before a sanctions notice forces the rewrite.

Wanting both without a plan is how you end up with neither

Neutrality isn't dead, but it's becoming expensive, and it needs capacity most middle powers lack. Europe looks like the cautionary tale: Bruegel says the Commission proposed an autonomy-minded Cloud and AI Development Act in June 2026, then signed the Pax Silica declaration the same month. Wanting both is understandable. Doing both without a plan is how you end up with neither.

Our advice: don't ask "which side?" Ask "which layers can I afford to lose?" and build outward from the answer. Washington's pressure may even backfire, since Bruegel reports the Chinese-led body has grown from 29 members to 38. Coercion is a poor way to sell a standard.