AI’s energy obsession gets a reality check

Still, the model poses a threat to the bottom line of certain players in Big Tech. Why pay for an expensive model from OpenAI when you can get access to DeepSeek for free? Even other makers of open-source models, especially Meta, are panicking about the competition, according to The Information. The company has set up a number of “war rooms” to figure out how DeepSeek was made so efficient. (A couple of days after the Stargate announcement, Meta said it would increase its own capital investments by 70% to build more AI infrastructure.)

What does this all mean for the Stargate project? Let’s think about why OpenAI and its partners are willing to spend $500 billion on data centers to begin with. They believe that AI in its various forms—not just chatbots or generative video or even new AI agents, but also developments yet to be unveiled—will be the most lucrative tool humanity has ever built. They also believe that access to powerful chips inside massive data centers is the key to getting there. 

DeepSeek poked some holes in that approach. It didn’t train on yet-unreleased chips that are light-years ahead. It didn’t, to our knowledge, require the eye-watering amounts of computing power and energy behind the models from US companies that have made headlines. Its designers made clever decisions in the name of efficiency.

In theory, it could make a project like Stargate seem less urgent and less necessary. If, in dissecting DeepSeek, AI companies discover some lessons about how to make models use existing resources more effectively, perhaps constructing more and more data centers won’t be the only winning formula for better AI. That would be welcome to the many people affected by the problems data centers can bring, like lots of emissions, the loss of fresh, drinkable water used to cool them, and the strain on local power grids. 

Thus far, DeepSeek doesn’t seem to have sparked such a change in approach. OpenAI researcher Noam Brown wrote on X, “I have no doubt that with even more compute it would be an even more powerful model.”

If his logic wins out, the players with the most computing power will win, and getting it is apparently worth at least $500 billion to AI’s biggest companies. But let’s remember—announcing it is the easiest part.

Deeper Learning

What’s next for robots

Many of the big questions about AI–-how it learns, how well it works, and where it should be deployed—are now applicable to robotics. In the year ahead, we will see humanoid robots being put to the test in warehouses and factories, robots learning in simulated worlds, and a rapid increase in the military’s adoption of autonomous drones, submarines, and more.