Why AI Spending Is Soaring

Major AI vendors have replaced flat fees with token billing, driving corporate tech spending upward.


This article appears in the December issue of Global Finance Magazine.

The pricing of artificial intelligence resources is evolving as quickly as the technology itself, and corporate treasuries need to gird for the change.

Per-seat licensing has become a relic as major AI model vendors have shifted to usage-based billing.

“Two or three months ago, there were per-seat licenses for $20, $100, and so forth for companies,” said Roni Karp, co-founder and CEO of cloud management vendor Wiv. “It was very cheap. Today, many enterprise programs charge by token usage and have become more expensive.”

When comparing AI license costs per employee, he added, it’s becoming the most expensive tool or software license. “I don’t know if there are any other software companies that pay $800 or $1,000 per month to give developers tools to code.”

Karp also sees AI costs growing faster than other IT spending, such as cloud computing services. “If cloud costs can grow 20% per year, AI costs can even grow 20% monthly.”

Since AI providers, on average, have dropped their token prices from a market peak in mid-July, according to the Token Price Index, AI use continues to skyrocket. 

“We definitely see the price wars as the labs try to get market share by reducing their per-million-token cost, but that is not reducing spending,” said Luis Solano, AI growth leader and senior director at Huron Consulting Group. “It’s actually going way up.”

Jevons’ Paradox

According to 19th-century British economist William Stanley Jevons, as technology becomes more efficient, its adoption increases, and AI is no exception.

Elon Musk said during SpaceX Corp’s third-quarter earnings call that AI, or superintelligence as SpaceX calls it, will require 1,000 times more bandwidth than it does today because of the dramatic growth of agentic AI.

“AI, or SI, agents will, on a per-agent basis, use far more bandwidth than a person,” he said. “And, there will be far more agents than people.”

AI workloads have quickly graduated from simple input-output queries to far more complex tasks

“Developers and cloud customers are already using it for things like untangling legacy code bases, accelerating scientific research, robotics, discovering product improvements, and domain-specific tasks requiring in-depth reasoning,” said Alphabet Inc. CEO Sundar Pichai during the company’s third-quarter earnings call.

Queries using Google’s AI mode typically take three times as long as traditional searches, and 20% of global AI Mode searches involve voice, images, or documents, he added. “We have crossed 50 billion visual searches per month.”

Google’s first-party model APIs are processing more than 32 billion tokens per minute, a 40% increase from the second quarter and more than four times what the search engine giant processed a year ago.

Meanwhile, Microsoft Corp.’s Copilot workloads grew 50% quarter over quarter, CEO Satya Nadella told financial analysts during the company’s first-quarter earnings call for fiscal year 2027 in late October.

Limited Resources

Musk said the market should expect a long-term shortage of AI compute capacity due to energy constraints, which will increase the value of compute per watt rather than lower it.

“The compute-per-watt framing is correct because the power is the constraint here,” said Huron’s Solano. “Three years ago, the constraint was chips. But today, there are reports that Amazon has a bunch of warehouses full of [Nvidia] GB300 processors that are waiting to be installed, yet there is no power.”

However, it is still early days for companies trying to manage their AI spending, from developing key performance indicators to implementing model routing and memory management strategies.

It might be too early to determine exactly how to measure AI KPIs, said Wiv’s Karp. “It is like trying to understand the KPIs for each employee.”

Instead, he suggested that firms focus on using the technology more effectively. “If you use the most expensive model for translation tasks, you’re probably not using it smartly, and it will become very, very expensive.”

Rob Daly covers economics and fintech. Contact him at rdaly@gfmag.com.

arrow-chevron-right-redarrow-chevron-rightbutton-arrow-left-greybutton-arrow-left-red-400button-arrow-left-red-500button-arrow-left-red-600button-arrow-left-whitebutton-arrow-right-greybutton-arrow-right-red-400button-arrow-right-red-500button-arrow-right-red-600button-arrow-right-whitecaret-downcaret-rightclosecloseemailfacebook-square-holdfacebookhamburger-newhamburgerinstagramlinkedin-square-1linkedinpauseplaysearch-outlinesearchsubscribe-digitalsubscribe-printtwitter-square-holdtwitteryoutube