Breakingviews - Corporate AI sticker shock will force restraint
HONG KONG, June 3 (Reuters Breakingviews) - If any trend should strike fear into the finance departments of global companies it is "tokenmaxxing". From Amazon.com (AMZN.O)
to Meta Platforms (META.O)
, firms have been aggressively pushing employees to embrace AI, driving up demand for models and apps. The shift has made once-predictable IT budgets wildly volatile and hard to justify. A correction looms.
Businesses are currently locked in a race to encourage widespread use of tokens - the unit of text or data that large language models process and produce. Tools like coding agents, which often run in the background and go back and forth constantly with models, can consume vast amounts of tokens. That has prompted AI firms from ChatGPT inventor OpenAI to Microsoft's (MSFT.O)
GitHub to introduce usage-based charges on top of monthly or annual subscriptions.
The Reuters Inside Track newsletter is your essential guide to the biggest events in global sport. Sign up here.
The extra cost is a fresh headache for chief financial officers. A Gartner survey found
three quarters of the executives it polled expect technology budgets to increase this year, and nearly half estimate double-digit percentage rises. A sizable chunk of that will go toward AI, which will account for over a fifth of total enterprise technology spending by 2035, up from under 4% today, estimates Oxford Economics.
Advertisement · Scroll to continue
That's great for OpenAI and Anthropic, both of which are readying blockbuster initial public offerings. Worldwide spending on models and AI software, including the pair's chatbots, assistants and agents, is on track to top $680 billion in 2027, more than double last year's total, per Gartner. Thanks to the popular Claude coding agent, Anthropic's revenue is on track to more than double to $10.9 billion from the previous quarter, propelling the firm led by Dario Amodei to its first adjusted operating profit in the three months to June.
Advertisement · Scroll to continue
This profligate spending looks hard for customers to sustain, though. Amazon employees have racked up additional costs after using AI tools for unnecessary tasks to inflate token consumption, according
to the Financial Times, prompting a senior executive to plead, "Please don't use AI just for the sake of using AI". Similarly, competition among Meta engineers resulted in the company using
over 60 trillion tokens over a 30-day period, per the Information. That's roughly $900 million worth of Anthropic's tokens, based on the firm's blended prices, according to one estimate
, though Meta is probably paying cheaper enterprise rates.
Advertisement · Scroll to continue
The challenge for executives is twofold. First, unlike stable software subscriptions, usage-based token bills are hard to track and predict. Uber Technologies' (UBER.N)
chief technology officer admitted in April that the company had already blown past its full-year AI budget; even the larger Microsoft has curtailed access to Claude Code for its engineers due to soaring costs, according
to The Verge.
Many firms are in the same boat: 71% of companies in one survey reported
AI cost overruns last year. Data management, infrastructure, networking charges and other fees are also piling up, making it even more complex for finance teams to model AI spending. Spotify Technology (SPOT.N)
reported a higher-than-expected
17% year-on-year rise in first-quarter operating expenses after adjusting for exchange rate movements and social charges, up from 13% the previous quarter. The music-streaming giant partly attributed the increase to AI and cloud spending.
Moreover, the returns from these tech investments remain hazy. Many companies are still experimenting with AI, but Microsoft boss Satya Nadella recently warned "IT budgets are going to have to be reshaped by a combination of business outcomes". Those will include revenue growth, durable productivity and efficiency gains and cost cuts. Early anecdotal data is sobering, though: an IBM survey of 2,000 global CEOs last year found
just a quarter of AI initiatives delivered the expected returns on investment over the last few years; a separate poll from BCG revealed
60% of companies did not generate any material value from AI. With the daily token cost of running agents surpassing comparable human labour for certain tasks, executives will have to tighten up outlays and pivot to delivering durable gains.
All this points to firms imposing financial discipline in the near future. One option is to perform less sophisticated tasks with cheaper models, such as open-source ones from China's Alibaba (9988.HK)
and its peers. That will require monitoring, capping and allocating budgets. The shift could be similar to how companies managed to control cloud computing expenses following the wholesale shift from local servers to remote processing power. As recently as 2022 one estimate put the proportion of wasted spending
on the cloud at 32%; that figure was down to 27% by 2025. Interestingly, cloud waste is rising again - the first increase in five years - partly due to AI.
As technology evolved and costs fell, cloud companies responded by offering more value-added services on top of basic infrastructure. These include Alphabet's Google (GOOGL.O)
Cloud Data Analytics and Amazon's AWS Elastic Beanstalk, a popular tool that companies use to easily run apps. As businesses focus on the cost of AI and the value they get from the technology, token pricing will probably follow a similar path.
The future of token costs remains the big unknown, though. A shortage of computing power, particularly in the U.S. where data centres face lengthy construction delays, electricity constraints and other hurdles, suggests demand will continue to outstrip supply.
At the same time, however, advances in chips, software and hardware mean costs for processing queries on models, known as inference, will continue to fall. Running an open-source model on AI-ready desktops, laptops and mobile phones, for instance, is already much cheaper than relying on a cloud-based proprietary model. AI upstarts have subsidised the cost of tokens, thanks in part to generous funding. That model looks unsustainable over the long run, though. OpenAI, for instance, doesn't expect to break even until 2030, the Wall Street Journal reported
in April, citing financial documents.
So an AI cost crunch looks inevitable. The silver lining for OpenAI and rivals is if token prices come down and consumers start controlling costs, that will not necessarily be at the expense of AI adoption. The cloud computing model suggests a more viable business model can emerge. Even so, the race to tame AI spending is only just beginning.
Follow Robyn Mak on X
.
Comments
Post a Comment