AI costs spiral as firms lose spending control

A firm spent $500 million in a single month on Anthropic’s Claude AI because it failed to set usage limits in what may rank as one of the costliest bills in the space. Thousands of employees at the firm ran complex tasks on Claude and the meter kept running nonstop, an IT consultant told Axios.

In another instance, Microsoft reportedly cancelled Claude Code licences due to growing token bills. Uber said it exhausted its 2026 AI budget in just four months. Companies are drowning in token bills due to heavy spending on AI tools without clear returns, success metrics, or guardrails.

Seven global firms including Uber, Microsoft, Commonwealth Bank of Australia, GitHub, Cursor, Klarna and Duolingo have reportedly either pulled back on AI spending, changed pricing tiers or even rehired engineers to save on ballooning AI costs.


A recent report by brokerage firm Jefferies noted that “AI for now is costing more money than it is saving.” Jefferies estimated that global AI capital expenditure could touch $4.7 trillion by 2029, of which $2.6 trillion will be spent on tech hardware.

Engineers are racing to top the charts on a new productivity metric: Tokenmaxxing.

According to some experts, staff are said to be using AI for what may be regarded as unnecessary tasks by stuffing prompts with unnecessary context, running endless agent loops, or feeding it with large documents, burning up tokens. Some even ask questions such as “How’s the weather?”

Meta and Amazon had to shut down their internal usage leaderboard to curb this tendency.

Organisations are realising that treating tokens as a success metric is a trap and instead building measurements to track whether they are actually delivering business outcomes.

For instance, Salesforce said its AI usage is going through the roof but the company is measuring outcomes with an Agentic Work Unit (AWU).

“AWU means a token was consumed to execute a task like open a service case, score a marketing lead or close a sales opportunity,” Salesforce chief digital evangelist Vala Afshar told ET recently.

The company consumed 12.3 trillion AI tokens in 2025, a figure expected to be exceeded within just two months this year due to exploding enterprise AI usage. “We may spend $300 million with Anthropic alone with tokens,” Afshar said, describing them as a “unit of currency” in the emerging AI economy.

Goldman Sachs has forecast global token usage could surge 24x to 120 quadrillion tokens per month between 2026 and 2030. cheaper tokens may not necessarily translate into lower AI bills, Gartner said.

More token burnouts

Although the cost per token is expected to fall by a factor of 100, cheaper tokens may not necessarily translate into lower AI bills, Gartner said. That’s because frontier intelligence will demand significantly more tokens than current mainstream applications. For instance, agentic models require between 5-30 times more tokens per task than a standard genAI chatbot.

“Tokemaxxing will continue to happen as agentic AI and the speed of the output makes it addictively useful,” said Neil Shah, founding partner at Counterpoint Research. “There are no potential checks or token budget planning happening at this point to alleviate the token burnouts.”

He added that organisations lack understanding to optimise token usage.

“In many instances, the user is unaware of how many tokens will be consumed for a particular workflow and the lack of transparency or help from the tool itself makes it a blind bet,” Shah said.

Open-source models such as DeepSeek with on-premise hardware can drive down total cost of ownership by nearly 75%, Shah said.