‘Tokenmaxxing’ Is Fading, Say Experts: What It Means For Nvidia, OpenAI, Anthropic And The AI Boom

• The term “tokenmaxxing” refers to companies encouraging their employees to use AI tools aggressively, with some firms reportedly tracking token consumption as a proxy for innovation and productivity.
• The practice fell victim to Goodhart's Law, where a metric loses its value once it becomes a target.
• AI researcher Gary Marcus argued that the decline of “tokenmaxxing” could expose broader weaknesses in the economics of large language models.

A growing number of industry experts and analysts believe the era of “tokenmaxxing” may be fading, as companies increasingly scrutinize the return on investment of AI tools and large language models.

The shift comes amid reports of companies scaling back the push for AI usage while burning through their annual budgets for the technology within months.

What Is ‘Tokenmaxxing’?

The term “tokenmaxxing” refers to companies encouraging their employees to use AI tools aggressively, with some firms reportedly tracking token consumption as a proxy for innovation and productivity.

The practice fell victim to Goodhart's Law, where a metric loses its value once it becomes a target.

Amazon.com Inc. (AMZN) employees generated unnecessary AI workloads simply to boost usage statistics rather than produce meaningful business outcomes, according to a report by The Financial Times. Amazon eventually shut down its AI leadership board.

Why Experts Are Concerned

AI researcher Gary Marcus argued in a post on X that the decline of “tokenmaxxing” could expose broader weaknesses in the economics of large language models.

Marcus suggested that OpenAI and Anthropic could face increasing pressure as AI models become more commoditized, competition intensifies, and profit margins shrink.

“Most of the companies that invested massively in them will struggle to make back their investments,” he said, while adding that Nvidia Corp. (NVDA) will eventually decline, once the market widely recognizes the other effects of falling “tokenmaxxing.”

Meanwhile, technology strategist Thierry Borgeat cited estimates suggesting that several hyperscalers may incur negative returns on AI-related capital expenditures under current assumptions, prompting comparisons to the dot-com era.

“And remember: that's assuming zero costs. In reality, GPUs depreciate, power bills run, salaries get paid. The real returns are worse,” he said.

Borgeat added that this is why the comparison to the dot-com bubble keeps coming up. “Incredible technology does not automatically mean sustainable economics. The internet survived. Most internet companies didn't,” he said.

Ride-hailing platform Uber Technologies Inc. (UBER) confirmed last month that it had burned through its AI budget for 2026 in just four months, according to a report by The Information.

Why This Matters For Nvidia, OpenAI And Anthropic

For Nvidia, the debate is significant because the chipmaker remains one of the largest beneficiaries of AI infrastructure spending.

If enterprises shift from maximizing AI usage to optimizing AI costs, growth in token consumption could slow, potentially tempering demand growth for inference infrastructure sold by Nvidia and other AI hardware providers.

Nvidia CEO Jensen Huang has been touting the current phase of AI as an agentic one. “I said, (it’s) completely opposite. And you can see it. (AI) agents is the going to create the largest opportunity for my partner companies,” he said at the ongoing Computex conference.

IO Fund’s Beth Kindig stated in a note in March this year that Nvidia’s $20 billion acquisition of Groq is aimed at driving up token usage, which would boost the company’s revenue and profits.

“Nvidia is preparing to position its GPUs to be among the best inference options available, utilizing Groq’s unique SRAM-based architecture to significantly turbocharge token throughput and accelerate inference performance,” she added.

Cost Another Major Concern For Model Providers As Well As Users

For OpenAI and Anthropic, growing pressure to control costs could lead enterprises to rely more heavily on cheaper models, smaller models or routing systems that reserve premium models for only the most complex tasks.

According to pricing data compiled by Artificial Analysis, some of Anthropic's flagship Claude models rank among the most expensive frontier AI offerings, while OpenAI’s ChatGPT, Google’s Gemini, xAI’s Grok, and DeepSeek are among the more affordable models.

This comes amid reports that Microsoft Corp. (MSFT) has begun canceling Claude Code licenses for its employees, instead turning to its in-house GitHub Copilot Command-Line Interface (CLI).

In another hit to its users, Microsoft-owned GitHub announced today, June 1, 2026, that it is transitioning to usage-based billing.

Copilot has changed significantly over the past year, evolving from an in-editor coding assistant into an agentic platform capable of handling extended, multi-step development tasks, the firm stated.

It added that Copilot can leverage the latest AI models, work across entire code repositories, and independently iterate on complex workflows, resulting in substantially higher compute and inference requirements.

The Invesco QQQ Trust (QQQ) is up 42% over the past 12 months, while the iShares U.S. Technology ETF (IYW) is up 59%.

Also See: NVDA, AMD, MU, MSFT, ORCL: Dan Ives Lists Stocks In His Shopping Bag Amid Ongoing Memory Supercycle

For updates and corrections, email newsroom[at]stocktwits[dot]com.