Companies rushing to replace human workers with AI agents are running into an uncomfortable truth...the bills are coming in higher than the salaries they hoped to eliminate.

Must read: Meta reached out for interview, then fired 8,000 techies: 'AI layoff culture has gotten unbelievably toxic'

Technology companies are finding out that the “token” costs tied to running AI agents are spiralling out of control, especially as engineers increasingly use multiple AI coding agents simultaneously.

What exactly is driving these massive AI bills?

Most AI systems charge based on “tokens,” the chunks of text processed by large language models during prompts, reasoning and outputs. The more tasks an AI performs, whether writing code, debugging software or reviewing documents, the more tokens it consumes.

The issue has become more visible as AI coding assistants gain deeper adoption across the tech industry.

“For my team, the cost of compute is far beyond the costs of the employees,” Bryan Catanzaro, vice president of applied deep learning at Nvidia, told Axios.

The economics become even more challenging when engineers run several AI agents at the same time. Instead of a single assistant helping with one task, companies are now seeing developers deploy fleets of autonomous coding agents in the background, all generating continuous API requests and consuming compute power.

Must read: AI layoffs may hurt companies too, not just workers: Study warns of ‘automation trap’

AI-generated code is becoming mainstream

The surge in costs comes as AI-generated code becomes increasingly common inside major technology firms.

In a post on social media platform X, Boris Cherny, head of Claude Code at Anthropic, claimed earlier this year that “pretty much 100%” of Anthropic’s code is now AI-generated.

Executives at Google and Microsoft have also said that AI now contributes to roughly a quarter of their companies’ code output.

Meanwhile, Meta has reportedly begun incorporating AI usage into employee performance reviews, reflecting how aggressively companies are pushing internal AI adoption.

What is ‘tokenmaxxing’?

As AI adoption intensifies, a new workplace trend has emerged among developers: “tokenmaxxing.”

The term refers to engineers consuming large amounts of AI compute power, often using millions of tokens per day, to maximise productivity or experiment with autonomous coding workflows.

According to The New York Times, some heavy users are generating monthly AI bills exceeding $150,000.

Must read: ‘We need to take down the size of the company’: Meta CEO Mark Zuckerberg signals more layoffs

“I probably spend more than my salary on Claude,” Stockholm-based software engineer Max Linder told the publication.

The problem has moved beyond individual excess. According to a report by The Information, Uber engineers using Anthropic's Claude Code have already exhausted the company's entire AI budget for 2026.  

Why AI companies may actually benefit from this

While enterprises struggle with soaring compute bills, AI model providers stand to gain significantly.

One OpenAI investor told Axios that concerns around token consumption could ultimately benefit OpenAI if its Codex tools prove more token-efficient than rivals such as Anthropic’s Claude Code.

Anthropic, meanwhile, has already raised prices for some of its AI services as demand for coding agents grows. Microsoft recently announced that it was moving Copilot on GitHub from request-based billing to usage-based billing.

The economics of AI infrastructure are also creating new competitive advantages. Companies are now evaluating models not just on performance, but also on how efficiently they use tokens and compute resources.

The rising costs have become significant enough for executives to rethink compensation structures around AI access.

In March, Jensen Huang proposed allocating AI tokens to software engineers equal to roughly half their base salary. Huang framed the idea as a potential recruiting incentive.

Must read: AI isn’t the real reason for tech layoffs? Salesforce CEO Marc Benioff calls it a ‘lazy way out’

Are AI agents actually improving productivity?

Despite the massive spending, questions remain over whether AI agents genuinely improve productivity at scale.

Several studies have suggested that forcing employees to use AI tools may sometimes increase complexity rather than reduce it, particularly when workers spend additional time validating AI-generated outputs or correcting errors.

Critics argue that while AI agents can accelerate repetitive tasks like coding, summarisation and debugging, they may also introduce new forms of inefficiency, including hallucinations, security vulnerabilities and quality-control overhead.

For now, companies seem to be caught between two competing pressures, the fear of falling behind in the AI race and the growing realisation that scaling AI agents may not be as economically straightforward as earlier expected.