As AI costs squeeze enterprises, the tokenmaxxers may be in for a reckoning.
Tokenmaxxing, or the idea that heavy AI usage directly equates to enterprise value, has swept through Silicon Valley in 2026, with massive impacts. Gartner has predicted that AI spending will reach nearly $2.6 trillion this year, up 47% from the prior year. But those costs may be starting to hurt.
Companies are questioning the spending sprees, and that includes some of the biggest names in tech.
- In an interview with The New York Timesthis week, Microsoft CEO Satya Nadella said that being a tokenmaxxer is "addictive," but he's now pushing employees to use more efficient models, noting that they "don't use frontier models for non-frontier problems." This follows AI efficiency being center-stage inproducts announced at Microsoft Buildlast week. - Meanwhile, Uber CTO Neppalli Naga told The Informationthat the company burned through its entire Claude Code budget for the year by April. - And at the Nebius' Inflection forum this week, the company's CRO, Marc Boroditsky, said the industry needs to shift to "valuemaxxing," or juicing as much value as possible from your tokens, according to Alex Heath's Sources newsletter. - Nikita Shamgunov, VP of Databricks, said that customers are eyeing costs as “the wheels started to come off” of tokenmaxxing, Heath also reported.
- Major AI firms may also be reading the tea leaves: According to The Wall Street Journal, both OpenAI and Anthropic are considering substantial price cuts for tokens as a means of winning customers from one another ahead of their record-breaking IPOs.
So how did this problem emerge? According to Raj Ramanujam, VP of Global Alliances and Cloud at Dynatrace, enterprises jumped headfirst into AI out of excitement and fear of being left behind, often building their pipelines and workflows without thinking about the downstream costs. Now, he told The Deep View, "people are suddenly waking up to that in a very uncomfortable way."
Rob May, CEO of Neurometric.ai, told The Deep View that this trend was born out of a simple need to measure the adoption and performance of AI in work settings. Many decided that measuring AI activity was a sufficient way to do so. However, while all of these tokens are measured in the same way, not all tokens are spent equally. For instance, a token that's used to write a to-do list may be counted equally to ones spent on complex tasks such as scientific research.
Additionally, the popularity of major model providers has led many to believe that other options, such as open-source models and small language models, aren't as viable. It's why May came up with an alternative: "Tokenminning." This is the idea that not every prompt and task requires ultra-powerful models, and that focusing on efficiency doesn't necessarily mean enterprises need to sacrifice on quality.
"The big labs have done a great job of branding, and people believe they can do things that nobody else can do," he said. "I don't think that's true."
Our Deeper View #
The money being poured into AI is nearly unprecedented, with both OpenAI and Anthropic approaching trillion-dollar valuations ahead of their public offerings, and xAI parent company SpaceX about to make its own debut at more than $1.7 trillion. Still, the question of whether the market is in a bubble continues to rumble under the surface. Now that enterprises have started to wake up to the reality of how much it costs to build AI into every part of their business, they may have to ask themselves whether some processes need the latest AI at all, saving the tech only for the most valuable applications. They may choose lower-cost SLMs and open-source models they can run locally for far lower costs and greater control. If that were to happen, the value of frontier models and frontier labs could fall back to earth.