AI’s costly buildout complicates the Fed’s inflation fight

Silicon Valley leaders from Elon Musk to OpenAI CEO Sam Altman have hyped the deflationary effects of the artificial intelligence boom.

“Intelligence too cheap to meter is well within grasp,” Altman wrote recently.

Musk, the CEO of Tesla and SpaceX and the world’s richest person, has argued that AI and robotics will create extreme abundance and drive down cost. SoftBank’s Masayoshi Son said he expected a 40% drop in prices and that “unnecessarily hard work, sweating work, would no longer be needed.”

None of those dreams are close to being realized.

Instead, AI is hitting a wall of corporate inertia as it spreads out into the economy — causing some near-term inflation and producing little evidence of a sustained productivity boom. 

Company adoption has proved slower than some of the boosters promised. Meanwhile, the tech industry’s multi-trillion-dollar spending spree on data centers and AI infrastructure has snarled supply chains. Spending to build out AI is raising prices in sectors like electricity. Costs are piling up before the full-scale payoff arrives. That poses a dilemma for the Federal Reserve, which needs to make decisions about how to manage inflation.

Some of the immediate costs of AI are easier to spot than the potential benefits, said Ronnie Chatterji, chief economist for OpenAI. 

“For it to impact the economy, it has to be adopted by organizations,” Chatterji said. “Those organizations have to realize value.”

While that is happening, Chatterji acknowledged that “it’ll still be a little while before we see it sort of clearly for productivity statistics.”

Capital expenditure on the AI buildout is expected to reach $581 billion this year in the U.S., and as much as $1 trillion globally, Goldman Sachs Research recently estimated. Spending in the U.S. alone amounts to 1.8% of gross domestic product, a share the firm estimates will rise to 2.8% by 2028. 

A survey by the Census Bureau published in May found that between 17% and 20% of U.S. businesses reported using AI, which is far more prevalent at large firms than small ones. 

Peter Boockvar of One Point BFG Wealth Partners compared AI to the last major tech-driven productivity boom: the internet. Even during that period of automation, the U.S. saw only a 1.5% gain in productivity over a 30-year period, Boockvar said. If you zoom out 50 years, productivity averaged 2.5%. 

“To think that generative AI is going to bring that level of enhancement to the economy, relative to the internet, is tough,” Boockvar said. “Technology has always made people more productive. But is generative AI multiple step functions higher? We just don’t know.”

Inside companies, some executives who have put AI into widespread use are cautioning that the industry’s promises need to be taken with a grain of salt. 

“The reality is that the technology is there,” said Julie Averill, Lululemon’s former chief information officer, who oversaw AI adoption at the company. “The hype is around the ease of the technology in a large organization.”

Lululemon used AI to help executives predict where products would sell best. That was a lot more complicated than using a chatbot.

“The things that have always made implementations in large companies difficult still exist, which is people,” Averill said. “Getting people to change their behaviors, taking them along the journey with you, and getting them to trust the model, that’s hard.”

Chatterji said he’d seen similar patterns in OpenAI’s data. He said the power users of AI deploy it at eight times the rate of average companies, measured by tokens per user. The gap has grown from two times since OpenAI published a report on it three months ago.

“It is growing incredibly fast in terms of the gap between the frontier firms and the typical firms,” Chatterji said. “The companies that are reorganizing their workflows around it and changing the way they work around AI, they’re having more success.”

Economists who study AI have a term for what Averill experienced: weak links. That refers to tasks that can’t be easily automated.

AI makes us more productive by automating work like reading a radiological scan — something AI can do very well. But jobs are really bundles of tasks, some more amenable to automation than others, according to Stanford professor Charles Jones, a leading scholar of how AI will affect growth. He’s now on leave at Anthropic.

The Nobel laureate technologist Geoffrey Hinton predicted in 2016 that radiologists wouldn’t be needed within five to 10 years. Instead, their numbers kept growing as AI made radiologists more valuable to the economy.

“It turns out that radiologists do more than just read scans, and AI tools complement those other skills by automating a fraction of the tasks that radiologists perform,” Jones writes. The other things radiologists do — talking to patients and working with colleagues, for instance — fall in the category of weak links.

The pervasiveness of weak links won’t be clear until companies adopt AI at a bigger scale. 

Last month, Fed Chairman Kevin Warsh appointed Jones to a task force that will inform how the Fed thinks about AI and its effect on the economy. Venture capitalist Marc Andreessen, whose firm is aggressively backing AI startups, is also on the team, and is among those predicting an era of “hyper-deflation.”

When Jones, Andreessen and others report back in a few months, they will join a roiling debate at the Fed about AI. 

The Fed needs to raise its growth forecasts to account for AI, Warsh wrote in November, before he was confirmed to the top job.

“AI will be a significant disinflationary force, increasing productivity and bolstering American competitiveness,” he wrote. It’s a position that helped his standing with President Donald Trump, who is lobbying for lower interest rates.

Warsh’s new colleagues aren’t convinced. Fed officials voted in July to leave interest rates unchanged at a range of between 3.5% and 3.75%. The decision wasn’t unanimous, as some Fed officials worried openly that they needed to restrain the economy to hold back AI-driven price increases. 

“The massive investment in data centers has also added a new demand element to the high inflation Americans are experiencing,” Minneapolis Fed President Neel Kashkari said in a statement explaining his decision to dissent in favor of a higher interest rate.

The rush to build power-hungry data centers is contributing to rising utility bills for many Americans. Household electricity prices rose 10.1% in the two years leading up to June. That was faster than the overall 6.3% increase in prices in that period, according to consumer price index data from the Bureau of Labor Statistics. 

Other Fed officials have raised concerns about supply chain constraints for the servers needed to power advanced models, as AI companies buy up all the chips they can from companies like Nvidia. Chipmakers haven’t been able to ramp up production fast enough to meet demand. 

Prices on certain products have shot up. The cost of dynamic random access memory, or DRAM, will have risen by 400% by the end of the year compared to 2024, JPMorgan Chase estimates. 

CPI data shows the cost of computer software and accessories has risen 22.9% since June 2024, including 17.4% in the past year.

That’s prompted Warsh to adopt a newly cautious tone. Companies’ vast AI spending is laying the groundwork for future growth, the Fed chairman said in July. But for now “the precise timing and magnitude of effects on the supply side remain hard to predict.” 

“The cost and inflationary aspect is really complicating Kevin Warsh’s job,” Boockvar said. “He wants to believe in the productivity enhancements down the road — but it’s not something he can react to.”

In other words, AI may eventually live up to the hype. But for now, the costs are very real. 

— CNBC’S Drew Troast contributed reporting.

WATCH: Former Trump adviser says Fed will need to hike rates