January has been notable for the number of important announcements in AI. For me, two stand out: the US government’s support for the Stargate Project, a giant data center costing $500 billion, with investments coming from Oracle, Softbank, and OpenAI; and DeepSeek’s release of its R1 reasoning model, trained at an estimated cost of roughly $5 million—a large number but a fraction of what it cost OpenAI to train its o1 models.
US culture has long assumed that bigger is better, and that more expensive is better. That’s certainly part of what’s behind the most expensive data center ever conceived. But we have to ask a very different question. If DeepSeek was indeed trained for roughly a tenth of what it cost to train o1, and if inference (generating answers) on DeepSeek costs roughly one-thirtieth what it costs on o1 ($2.19 per million output tokens versus $60 per million output tokens), is the US technology sector headed in the right direction?
It clearly isn’t. Our “bigger is better” mentality is failing us.
I’ve long believed that the key to AI’s success would be minimizing the cost of training and inference. I don’t believe there’s really a race between the US and Chinese AI communities. But if we accept that metaphor, the US—and OpenAI in particular—is clearly behind. And a half-trillion-dollar data center is part of the problem, not the solution. Better engineering beats “supersize it.” Technologists in the US need to learn that lesson.
