Can Cheaper AI Outrun the Data Center Backlash?

After U.S. voters on Tuesday elected officials who are pushing to restrict data center expansion, some tech executives and investors are concerned that new delays in standing up computing capacity will make existing servers for AI more expensive to run.

And if computing becomes even pricier, there’s a question of whether declines in the cost of running AI could slow down.

So far this year, the cost of running AI has been falling, even as the price of renting servers has risen. For example, OpenAI’s GPT-6.1 Sol model, released in late September, can perform comparably to a bigger OpenAI flagship model, GPT-6 Astra, when used with a high reasoning setting. Yet the newer model costs about a fifth as much on a per-task basis as the older model, research firm Artificial Analysis said.

Anthropic has touted similar efficiencies with its latest models. And that’s happening despite the fact that buying or renting Nvidia hardware has gotten a lot costlier this year.

As AI costs become a major concern at enterprises, the price-drop trend will need to persist or else customers might pull back.

One senior executive of a private equity firm that owns and develops data centers told me this week that they believe prices of AI models will keep falling due to “token efficiencies”—the ability of AI to do more work with less compute—thanks to both hardware and software advancements.

In other words, they said, such efficiencies would help offset constraints on data center expansion. Of course, the executive is heavily incentivized to say such things, given the financial bets their company has already made.

On Wednesday, Amazon Web Services CEO Matt Garman said during an event hosted by private equity firm Blackstone that he is optimistic about the company’s data center expansion despite the election results. “We feel like we're in a pretty good spot, and I think others may have some more work to do.”

To be sure, Amazon has faced significant local resistance to some of its data center projects, and even had to drop a project in Virginia this year.

Garman didn’t comment on whether he expects server rental prices to rise further. But stock investors seemed to think so initially, based on the way they bid up shares of major cloud providers Wednesday while selling off stocks of smaller cloud providers that still need approvals for new facilities. (The major cloud providers pared back some of the gains during a selloff Thursday.)

Notably, many facilities that are supposed to come online next year have gotten some or all of the necessary state or local approvals, so the effects of the electoral backlash might have a big impact on facilities aiming to launch in 2028 and beyond.

Anthropic’s initial public offering prospectus offers another signal that the price of using its models has room to fall: the company’s gross profit margins have hovered around 85% in the first nine months of the year. (That excludes the cost of developing the models themselves.) So it may have some slack to lower prices if it needs to boost demand.