Billions

Yesterday, in the Oval Office, OpenAI, SoftBank and Oracle announced Stargate: $500 billion for AI infrastructure over four years, $100 billion of it upfront, starting with a data centre campus in Texas. That’s a lot of zeroes, and it isn’t even the whole bill. Microsoft is putting $80 billion into AI data centres this fiscal year on its own. Amazon added another $4 billion to Anthropic in November, and Google topped that up with a billion more this week. Add up the hyperscalers’ capex plans for 2025 and you’re somewhere in the hundreds of billions before anyone has trained a single model.

Ten months ago I wrote about the first couple of hundred billion and asked where the returns were going to come from. The answer, it turns out, is to spend another half a trillion.

The American approach to AI is now unambiguous: more. More money, more chips, more power, bigger models. It’s a brute-force attack on the problem, and it’s a very American way to attack a problem - outspend, outmuscle, dominate, and sort out the economics later. If you believe, as I do, that AI and its infrastructure is the Manhattan Project of our time, then half a trillion dollars seems adequate.

But is it the right way?

The other game

Look across the ocean. Two days before Stargate, DeepSeek released R1, a reasoning model that matches OpenAI’s o1 on the benchmarks. Open weights, MIT licence, and an API price that reads like a typo - about a thirtieth of what OpenAI charges per token for o1.

DeepSeek is not throwing money at the problem, mostly because they can’t. They train on H800s, the deliberately hobbled chips that export controls allow Nvidia to sell into China. The final training run for V3, the base model under R1, reportedly cost $5.6 million. You can argue about what that number leaves out (prior research, the hardware itself, the salaries), and people are, but even if you’re generous and multiply it by ten, it’s a rounding error on a single Stargate site.

So instead of more, they squeezed. A mixture-of-experts architecture so that only a fraction of the model does any work per token. FP8 training to halve the memory and the bandwidth. Reinforcement learning applied straight to the base model, skipping the expensive supervised fine-tuning stage that everyone assumed was necessary. None of this is magic, and none of it is unknown to the labs in San Francisco. The difference is that in Hangzhou each of these tricks was a necessity, and in San Francisco each was a nice-to-have that you could always substitute with another ten thousand GPUs.

That’s the whole story in one sentence: America is playing the “more” game, and China, by force of circumstance, is playing the “do more with less” game.

Which one wins?

I don’t know, and neither does anyone announcing numbers in the Oval Office. But I’d point out two things.

First, this isn’t only about who gets there first. It’s about who gets there sustainably - who ends up with the foundation for the next revolution rather than just the current one. Every dollar of capex has to be depreciated, and every GPU bought today is obsolete in three years. The bill for brute force comes due on a schedule; the bill for efficiency doesn’t.

Second, innovation is quite often about less. Less waste, less bloat, less brute force. The elegant solution, the efficient path. The history of computing is full of moments where somebody with fewer resources was forced to find the clever way, and the clever way turned out to be the way. It would be a strange irony if the export controls designed to slow China down were the thing that taught them to build better.

So the question I keep coming back to is whether we’re so blinded by the dollar signs that we’re missing the real innovation happening elsewhere. Whether we’re building an empire on sand while other people are building on Bedrock (pun intended).

The race is on, and it’s a real race. But it’s not only a race to spend the most. It’s a race to think differently, and at the moment only one side is being forced to. Will we learn to squeeze, or will we keep throwing money into the pit? The answer will shape the future of AI - and, given what AI is going to touch, the future of quite a lot else.

P.S. The EU is still sitting in the corner, eating sand.