GitHub

Companies throughout the industry are beginning to hesitate about the price of artificial intelligence. By April, Uber would have run out of all the A.I. budgets for 2026. MicrosoftInitiating their developersClaude.The code licences were revoked several months later. A Priceline employee told TechCrunch that the extension of the regular Cursor contract was four to five times more expensive.

Although the price of each token has fallen, increased use of artificial intelligence and the promotion of increasingly autonomous agents have contributed to the increasing consumption of the coin. Companies that subscribed to unlimited subscriptions early in 2025 are now busy knowing where their funds are going, reducing their expenditures and whether some return on investment can be saved from budget debris.

At the same time, a market is emerging to meet their needs. Creative companies, established suppliers and new standards agencies are competing to provide companies with tools and language to track their expenditures.

"Six months ago, when I spoke to my client, it was all about what it could do. Is it good enough?"OpenAI Executive Alexander Embiricos told TechCrunch this week at an event in New York City. “This is never our dialogue. Now the conversation is, "Hey, we spent so much money. What's your visibility? What's your auditability? Do you have any token control? What's the efficiency of your model?"

In this context,Linux This week, the Foundation released the Tokenomics Foundation ' s plan, a new standard institution designed to instil the same cost discipline around AI tokens as FinOps for cloud expenditure.

“In April and May, I began to hear the company say, `Oh, my God, our currency budget is three times the total 2026 currency budget, and this is only April,' the executive director of the Linux Foundation's Finops Foundation told TechCrunch. "We began to hear a crisis, and the whole conversation went from currency maximization and `quick' to `We need a fence, how do we control this?"

The CEOs strongly urged their teams to use the best models and move quickly, regardless of cost, to hear such calls in the scientific community. New models released in November, for example Anthropic Claude Opus 4.5, OpenAI GPT-51 and Google Yes. Gemini. 3 Pro has brought significant improvements to proxy tools and has multiplied consumption. It was reported that a company, after forgetting to impose restrictions on its staff, found itself receiving $500 million in Claude ' s bills.

Priceline IT Senior Financial Director Chris Reed said, "It's like a strong cocaine epidemic." He noted that the company had begun to impose symbolic restrictions on certain groups. "They let you try, they make you fascinated, and now you're grateful."

Chief Executive Officer of Faros AI Gordon.Rdon said that he had recently spoken to a chief technical officer, who told him: “One of my engineers spent $40,000 last month on tokens, and I really don't know whether to stop him or tell everyone else like him.”

A March survey by Faros found that out of 20,000 developers, output was increasing, but errors and rewrites were also increasing. The project management platform Jellyfish similarly found that engineers who use the most tokens are working about twice as efficiently as those who use less artificial intelligence, but they spend 10 times more.

The Jellyfish Research Manager, Nicholas Arcolano, by e-mail to TechCrunch, indicated that artificial intelligence expenditure had increased in an explosive manner, largely due to proxy functions, and that the expenditure per developer had increased by about 18.6 times in nine months. In sum, these statistics make the productivity situation more ambiguous than expenditure.

“The return on extreme expenses depends on the final commercial value of the code delivered (e.g. income), which most companies are still unable to measure”, Akolano said.

At least part of the measurement problem is the scale of use of artificial intelligence today.

“Tracking cloud costs is a problem of hundreds of millions of lines per month”, according to Stoment. “Tracing the cost of tokens is a data problem of trillions of lines per month. You can't just put it in any spreadsheet or even basic tool. You have to fundamentally rethink your tools, specifications and accounting systems to do this.”

In Priceline, Reed has seen the difference. He identified problems between the use of vendor reports and Priceline internal data.

“My career begins with telecommunications cost management, from telecommunications to clouds to artificial intelligence, and I see all the same similarities”, he said. “When you introduce new things, mistakes in costing and opportunities for audit and optimization are ripe.”

A market has begun to emerge around this issue. Some pure companies, such as Pay-i, track, measure and optimize the costs and performance of GenAI investments. At the same time, payments allow developers to track costs, measure usage and bill users on the basis of actual value rather than subscription fees.

Then there are companies like Jellyfish, Waydev and Faros AI that provide AI proxy surveillance to prove the return on investment in the developers' tools. Storment indicated that most of the 180 suppliers within the FinOps Foundation preferred this area.

Existing distribution companies are also adding new functions to take advantage of this new market. Ramp has recently entered the area of artificial intelligence expenditure management; Datadog and New Relic have increased cloud cost management, token-level observability and GPU Surveillance services, etc. Next week at the Finops X conference,AWS It is expected that a new financial management functionality will be introduced for the enterprise's artificial intelligence expenditure.

The partner of NEA, Tiffany Luck, believes that token efficiency and observability may be added to the “use layer or application layer”. She noted that Factory, a start-up company that produces artificial intelligence agents for an enterprise, had launched a model router this week that could automatically select the right model for each task.

Gordon expects that forward laboratories and other model providers will use the OpenRouter-type optimization to facilitate inquiries into the cheapest models — a trend that has appeared on the corporate Claude bill.

“The financial report shows how much you spent on Anthropic, even if you call it the Opus model, part of your expenditure will be spent on Sonnet or Haiku because they're smart enough to do it,” Gordon says. "I think it will become more and more important."

However, all of these tools are constructed without common language or shared definitions, which include the cost of tokens, the content generated and how to compare the expenditures of different suppliers. This is where the Foundation for Currency Economics wants to play its part.

The Foundation is developing a normative definition and framework for “currency economics”; AI Open Standards, Norms and Indicators for the Use and Fee of Currency; and new indicators for artificial intelligence economics, such as per smart cost or per wad. It also plans to define indicators of the effectiveness and consumer efficiency of currency factories. The organization is scheduled to be officially launched in July and will announce more members at the FinOps X meeting next week.

In a statement, the Chief Available Officer of Salesforce, Nishant Gupta, stated: “The economics of tokens is fundamentally more abstract and non-transparent than any product of this size that we have managed before.” “It requires operational capabilities different from those of the industry's ability to operate for clouds.”

Nevertheless, Goldman Sachs expects a 24-fold increase in global token use by 2030. Companies that have exceeded their budgets now need a solution, and the Foundation ' s first delivery will take months.

"Maybe we created a steam engine, but we haven't figured out the assembly line," Gordon said.

Akolano believes that it is wise to use it widely and proportionately.

He said: “The best return on investment comes from increasing the use of a wide range of intermediate users from low to medium, rather than pushing up large numbers of users.”

Russell Brandom and Tim Fernholz contributed to this report.