Washington Wants 35 Countries to Pick a Side It Can't Power

The State Department is drafting letters to 35 countries demanding they choose between the American-led Pax Silica alliance and China's World AI Cooperation Organization, with dual membership off the table ^[1]. House Foreign Affairs Chairman Brian Mast framed the choice in the starkest terms available.

They are trailing substantially. — Scott Bessent

The side Washington is asking them to choose is the one that promises secured AI infrastructure. Here is what that infrastructure looks like on the ground. Of the 1,066 gigawatts of data center power developers have requested, only 28 percent will materialize ^[2]. Texas has paused all data center grid approvals pending an audit ^[2]. Eric Schmidt has pointed to Microsoft holding GPUs it cannot energize because the grid is strained.

The amount of energy that we need for [AI]... [is] 1,000 times more power than we currently have. — Jensen Huang

Eaton's U.S. data center backlog stands at 307 gigawatts.

The limiting factor currently is memory. — Elon Musk

Mistral's Arthur Mensch put the infrastructure gap in historical terms.

100 years ago, if you weren't building electricity factories, you were preparing yourself to buy it from your neighbors, which, at the end of the day, isn't great because it creates some dependencies. — Arthur Mensch

You cannot ask nations to pledge to an infrastructure alliance whose infrastructure you cannot build. And the dependency Mensch warns about is already forming in the other direction: as centralized U.S. cloud infrastructure strains, Chinese open-weight models offer frontier-class capability at a fraction of the cost ^[3]. China is subsidizing AI electricity by up to 50 percent for its tech giants ^[4]. The subsidies are partly a workaround for a hardware deficit — China's own chips remain generations behind TSMC and burn 30 to 50 percent more power than Nvidia's ^[4] — but that deficit slows China's own data centers, not the spread of its models, which run on whatever hardware a buyer already has. Alibaba's Qwen models have passed 3 billion downloads in six months, against 418 million for Alphabet's ^[5]. Hugging Face now describes Qwen as part of the default workflow for developers.

Qwen has become part of the default workflow for developers deciding what models to fine-tune and deploy. — Hugging Face

The models are already inside American companies. Pinterest runs DeepSeek R-1 in its recommendation engine; Airbnb runs Alibaba's Qwen in its customer-service agents.

We’ve effectively made Pinterest an AI-powered shopping assistant. — Bill Ready

Open source techniques that we use to train our own in-house models are 30% more accurate than the leading off-the-shelf models. — Matt Madrigal

And China is handing the same tools to the developing nations Washington is pressuring — 5,000 AI training slots and a weather-warning system for 30 countries ^[6]. The open-weight models are already out. Three billion downloads, embedded in American enterprises, cannot be recalled ^[7]. The U.S. closed models remain behind paywalls, and the power to serve them at the scale promised does not yet exist. Washington is asking countries to choose the side that promises infrastructure it cannot build, over the side that is already handing them working tools for free.