How China’s AI surge defies US spending power

Creators of all types have been griping for years about the aggressive, often illegal tactics used by the big US-based AI model developers to train their models.

Whether it’s OpenAI allegedly using datasets with pirated books, Meta, X, OpenAI, and Microsoft all reportedly using copyrighted music to train their models, or Anthropic allegedly scraping Reddit, they all seem to have been implicated in some way.

Yet now these same big tech players are complaining that China’s AI industry is stealing their stuff. The irony is that while hypocritical, the complaints are also increasingly beside the point.

While US hyperscalers are spending hundreds of billions to build AI moats, China’s developers are advancing on a fraction of the budget. The data shows Chinese developers are doing more with less via capex discipline and model efficiency.

Capex explosion comes with ROI expectations

Capital expenditure (capex) by the big US-based hyperscalers has exploded in recent quarters and promises to go further skyward in 2026.

For the 12 months ended the third quarter of 2025, for instance, the hyperscale sector saw its free cash flow margins plummet to an industry-wide average of 12.8%, an all-time low.

The dip is due to soaring capex, which hit US$450 billion over 12 months that quarter and is poised to reach US$600 billion in 2026. The dip is not the end of the world, as long as it delivers growth, though so far, there are few signs of a dramatic AI-related revenue pickup that would pay for all this spending.

See also: Movers and shakers in Southeast Asia’s genAI space

Some seemingly expect there to be a “winner takes all” outcome in this market, just as Google dominates search and Apple dominates smart devices. Hyperscaler CFOs are betting that all this capex is building a moat that will protect their market position for years to come.

Those bets look more and more like losers, as there is no obvious lock-in to the current models, with little brand loyalty from users. ChatGPT owned the market early on, but Gemini is catching up. And while some models are embedded in other platforms and/or corporate processes, this is nascent.

That lack of lock-in was obvious before China’s DeepSeek rolled out its first release, a bit over a year ago. Once it premiered, the world was shocked by what could be accomplished on the cheap.

Rumors flew that the model was built somehow using copied data, but the platform also had a legitimately different approach than the incumbents.

And now, a year or so later, China’s AI ecosystem has developed dramatically, despite limitations on chip access.

Leaner, meaner, and open-sourced

The Huawei case

Billions vs. millions