AIWatch.

At this point, It’s reasonable to assume that continued scaling - across training compute, model capacity, post-training, and inference - will continue to produce better results. Maybe “scale is all you need” isn’t entirely true, but “scale is much of what you need” certainly seems to be. Training larger models, serving more inference, and staying competitive requires not only the best compute and the most advanced chips, but above all one thing: electricity. By 2030, that could become the single biggest bottleneck.

This is where China holds a clear advantage over the United States, and I won’t even get into Europe. China is currently building 37 nuclear power plants. In 2025 alone, it installed as much solar capacity as the United States did over the previous 10 to 15 years combined (315GW). The US is pursuing a mix of natural gas, renewables, existing nuclear capacity, grid expansion, and, over the longer term, Small Modular Reactors (SMR). Whether that approach can scale fast enough to meet the pace of growing demand remains an open question. At the same time, ideas such as moving data centers into space are being discussed, since compute could theoretically scale more easily there and solar power would be available in virtually unlimited quantities. For now, however, these remain theoretical concepts. On top of that, the US power grid would require a comprehensive modernization to keep pace with future demand.

In short, if we focus only on who has the best models today and who will be able to train the most powerful models tomorrow, we’re missing what may ultimately be the more important factor: energy infrastructure. And this is where China is clearly ahead. That’s simply a fact. If China catches up in AI models, that is already highly significant from a competitive standpoint. But when looking at the next four to five years, the strategic challenge may lie elsewhere. In that respect, China already has a substantial structural moat.

It’s an aspect that should always be part of the discussion.

(I fact checked the numbers, they are correct).