China’s AI Factories Are Rising Faster Than Washington Expected

A section of scrubland is traversed by the road leading to one of China’s newest data centers, where dust is carried by the wind in long, low waves. A facility that hums day and night is powered by solar panels that are arranged in tight rows and shimmer under a pale sky.

Racks of servers blink silently inside, processing language models that are currently in use as far away as Nairobi and São Paulo. It’s difficult to ignore how unceremonious everything seems—no spectacle, no grand unveiling, just steady construction that goes almost unnoticed. This seems to be the exact point.

Washington policymakers have long viewed artificial intelligence as a race characterized by innovations, such as larger models, more intelligent outputs, and more potent chips. However, the framing seems a little strange when strolling through China’s expanding network of what engineers refer to as “AI factories.” These locations don’t pursue media attention. They generate scale. It turns out that scale might be the longer-lasting benefit.

China’s push has been influenced more by industrial instinct than by frontier ambition. Companies are encouraged to integrate AI into manufacturing lines, logistics hubs, and transportation systems by the government’s “AI Plus” strategy, which is subtly incorporated into planning documents. It’s possible that the buildout is being accelerated by this practical, occasionally even mundane focus. China appears to be asking a different question as American companies strive for the next technological advancement: how quickly can this be used everywhere?

The investment rhythm is altered by that question. Data centers, which are frequently powered by inexpensive solar or coal-based grids, rise close to energy sources in cities like Inner Mongolia and Gansu. The electricity is subsidized, sometimes significantly, which lowers operating costs in ways that executives in Silicon Valley privately acknowledge are hard to match.

The smell of heated metal permeates the air as workers pass stacks of cooling units and shipping containers that have been converted into server housing outside one facility. Cost appears to be the deciding factor, according to investors.

Parts of Silicon Valley were shocked by the emergence of businesses like DeepSeek, not because their models were drastically different but rather because they were effective. A few million-dollar training runs, as opposed to billions, necessitate an uncomfortable recalibration. While openly supporting American companies, some venture capitalists started discreetly testing Moonshot AI’s Kimi model in production settings when it was released at a fraction of Western costs.

Those choices have a hint of tension. On the one hand, U.S. labs continue to be at the forefront thanks to Nvidia chips. However, when applications become more important than raw capability, that lead seems to get smaller. It’s still unclear if more popular models or better models prevail.

It becomes apparent when one strolls through a logistics warehouse outside of Shenzhen. AI systems that don’t need to be flawless—just dependable and affordable—direct robots as they move between shelves. Employees keep an eye on screens and occasionally intervene when a problem arises. The system has an iterative feel to it, continuously getting better without ever aiming for perfection. As it develops, there is a sense that China may have an advantage due to its readiness to implement early and improve later.

It wasn’t always the case. Analysts have long identified China’s restricted access to cutting-edge semiconductors as a structural vulnerability. Export restrictions were meant to impede, if not completely halt, progress. However, the workaround culture—stockpiling older chips, sourcing through middlemen, and optimizing software—has produced an unanticipated result: a forced efficiency.

In the meantime, Chinese AI models are becoming more widely used worldwide, frequently as a result of open-source releases. Developers who might never be able to afford high-end Western tools are adapting Alibaba’s Qwen, which has been downloaded hundreds of millions of times. These models are the norm rather than alternatives in some regions of Southeast Asia and Africa. Ecosystems are drawn along by the gravity created by that adoption.

This is a subtle change that feels more like infrastructure subtly taking hold than a sprint. With companies like OpenAI and Anthropic pushing boundaries that others haven’t, the United States continues to rule the frontier. However, dominance in usage does not always follow from dominance at the frontier.

The issue of pace is another. The building of data centers in China has taken on an almost industrial cadence that is scalable, repeatable, and a little relentless. American projects, on the other hand, frequently go through several levels of approval, funding, and public review. One model is unquestionably faster than the other, but neither is intrinsically superior.

As these AI factories proliferate, it gets more difficult to write them off as inferior to the “real” race. At least one aspect of the race is represented by them. It doesn’t look dramatic—rows of servers feeding applications that reach millions, silently training models on less expensive power. However, it builds up.

It’s too soon to predict how this will turn out. Innovations might tip the scales back in favor of those with the most sophisticated chips. Alternatively, advancement might stall, benefiting those who deploy extensively and affordably. But for the time being, something is clearly changing.

And it’s happening more quickly than anticipated.