Corporate America was enthusiastic about artificial intelligence. Until it got the bill.
Whiplash around spending on “tokens,” the units of computing power in which A.I. is sold, is hitting engineering teams and board rooms. First there was “tokenmaxxing,” as executives encouraged as much A.I. use as possible. Then there was “tokenminning,” after they rapidly burned through millions of dollars in company money.
The simultaneous urgency and uncertainty is raising complex questions. What is A.I. even good for? How do you know what you’re buying, and measure the value of what it enables? How is it priced, and how will those prices change in the future? What does the spending on it displace?
“It’s a currency where you have no instinct to know what you’re using, and the accounting practices aren’t even there for it,” said Howard Rubin, an economist who advises companies on technology spending. “The A.I. stuff is being treated as an investment right now, but it’s a risky investment in case it has no return.”
The high stakes and scarcity of knowledge have given rise to a new field within economics devoted to figuring out how companies buy A.I. and what value they get from it. Call it tokenomics: the study of how this limited resource is created, traded and converted into things people want.
Businesses are no strangers to adopting new technologies. Electricity allowed power to be generated far from where it was used. The internet radically changed how services are delivered. Cloud computing turned data processing from a clunky on-site task to something that could be rented remotely.
The implications of A.I. could be as profound, but the costs and benefits are more difficult to weigh.
Perhaps the closest analogy to how A.I. works as a business expense is petroleum products. A common feedstock, crude oil, is refined into commodities such as gasoline, diesel and propane. Those fuels have different purposes, but they can be broken down into a standard measure of energy — British thermal units — and they are generally sold wholesale through exchanges where prices fluctuate with supply and demand.
The feedstock for tokens is electricity and semiconductors. The resulting “compute” is fed into models and metered out by tokens, which are generally understood to represent a few characters of English text.
But the petrochemical parallel ends there. There are now thousands of A.I. models, and little transparency around who is paying for what. There is no consensus on how much energy a token requires, which model is best to perform a given task or how many tokens it should take.
“Intelligence as a utility has so many different ways in which it can be deployed, so the value identification problem has become a lot harder,” said Ram Bala, a professor of A.I. and analytics at Santa Clara University. Some models, for example, will keep trying to accomplish a goal even when the mission is hopeless.
“On the one hand, you want to solve a difficult problem. That’s the positive side,” Mr. Bala said. “The negative side is, this could go on in a loop forever, and not really come to a reasonable solution at all.”
The Linux Foundation, a nonprofit that hosts open-source technology projects, previously developed standards for cloud computing that made it easier to compare services across providers. In June, it established the Tokenomics Foundation with similar goals: to create common parameters that A.I. providers agree to disclose.
“There’s a set of decisions that every organization in the world is making right now that we want to standardize the frameworks they’re thinking on, so they can have better starting places for them,” said J.R. Storment, the foundation’s executive director.
Companies are using a new crop of tools and service providers to get some benefit from A.I. without wiping out their profit margins. Revenium, for example, helps track whether token spending improves productivity enough to justify its cost.
Jason Cumberland, Revenium’s co-founder and chief operating officer, said he recommends clients set limits for how much of the expensive models engineers can use. After maxing out their allocated token use, engineers turn to cheaper open-source models, forcing them to think hard about what they really need. Revenium helps clients map their token spending to specific outcomes, like the number of software features shipped.
“The people who come to us are experiencing spending explosions,” Mr. Cumberland said. “They want to understand whether it’s worth it, but also how to curtail it in a way that doesn’t just stop work, which is the challenge.”
One big question they all face: Where does the money come from to pay for all this extra token spending? At first, Mr. Cumberland said, companies scrounged under their proverbial couch cushions to pay for subscriptions to Claude or ChatGPT. As they came to rely on those services, they started to cut software licenses and outside services, like web development and copy writing.
Now, companies are weighing whether token spending should be considered part of a company’s labor budget. That’s the main question economists have: whether A.I. will displace or augment human work. In a thought experiment, analysts at Bain & Company, the consultancy, recently posited a medium-term scenario in which tokens make up about a quarter of corporate operating expenses.
Elisity, a cybersecurity company, says it’s nowhere near there, but it is laying the groundwork with a metric it calls “bionic head count.”
The measure totals up A.I. spending and divides it by the average cost of a worker’s salary and benefits. It then divides annual revenues by the resulting number of human and virtual “employees” to measure their collective output.
To make it all feel a little more realistic, human staff members even write job descriptions for each A.I. agent in order to justify “hiring” it. Each time an engineer runs a model on a new task, it must get an evaluation.
“Essentially, you’re adding virtual head count. Is that resulting in incremental revenue which is all that really matters, or are you just eating at your margins?” said Charlie Treadwell, Elisity’s chief marketing officer. “In a growth stage company, it shifts our mind-set of where are we going to spend the capital to grow faster.”
That equation could change quickly, however, if token prices rise substantially. Elisity is getting by on a flat-rate team subscription for 150 users. If the company were charged by the token — the norm for organizations with more than a few hundred users — spending would quadruple, Mr. Treadwell said, and it would have to reconsider its use. (Elisity has no plans to shed staff, however, and is hiring.)
To make matters more complicated, tokens can be priced differently across cloud providers, and it’s not clear what drives prices up or down. With A.I. laboratories in an arms race to win market share before going public, what they charge may not completely correspond to what tokens cost to produce.
And A.I. can help minimize its own costs. Part of that is because of the increasing use of cheaper so-called cache tokens, which the model has already processed once and can reuse at a fraction of the price. As more businesses use A.I. through software tools that can execute tasks on their own, called agents, those agents are able to delegate more of the work to cache tokens.
In a working paper published this month, a team of economists found that dynamic has driven spending down relative to what it might otherwise be, even as token consumption and the posted prices for frontier A.I. models have risen.
The ability of token costs to stay competitive with human salaries will bear heavily on which kind of intelligence companies lean on in the future. But there is still a shortage of people who know how to deploy A.I. in productive ways. That is keeping consultants like Jue Wang busy.
“I actually think that is the constraint that people don’t talk enough about in the industry,” said Ms. Wang, a partner at Bain. “We often think about ‘Oh, they don’t have capacity.’ But they also don’t have enough people to help deliver the value.”
One of the fundamental difficulties in studying how spending on A.I. is playing out within organizations is the lack of comprehensive data that track it. There is no centralized exchange, no futures market, no government survey that reports prices and spending.
It’s why Aleh Tsyvinski, an economics professor at Yale who co-wrote the recent paper on token use, was excited to analyze a data set from the platform OpenRouter representing just 2 percent of total A.I. spending. It allowed him to study how financial markets react to those expenditures, but many questions remain.
“It’s one of the biggest challenges of our generation,” Mr. Tsyvinski said. “It’s good to understand how A.I. is probably going to probably change everything. Or maybe not change anything, but it’s good to have the measurement.”