The corporate race to use more AI is slowing down
A workplace obsession with maximizing generative AI usage is beginning to run into a harder reality: cost discipline. A Fast Company report published July 28 described the fading appeal of “tokenmaxxing,” a term used to describe pushing heavy consumption of AI tokens across tools and workflows in the belief that more usage signals better performance, faster output, or stronger innovation.
That mindset gained traction as companies rushed to operationalize systems from providers such as OpenAI and Anthropic. In practice, token usage became a rough proxy for intensity of adoption. The more prompts, outputs, agents, and automated tasks an organization generated, the more advanced it could appear. But the report suggests that corporate buyers are increasingly separating activity from value. Rising bills have not always been matched by equivalent gains in productivity or business outcomes.
Vincent Gusdorf, head of AI analytics at Moody’s Ratings and author of a report cited by Fast Company, argued for a more disciplined approach. His point is straightforward: generative AI makes it easy to create work, summaries, drafts, and process layers that may not actually be necessary. When each interaction consumes tokens and scales across a company, waste can compound quickly.
From status signal to cost center
The article traces how token-heavy usage was celebrated only months ago. Prominent tech leaders had framed high AI consumption as a sign that teams were experimenting aggressively and taking the technology seriously. OpenAI chief executive Sam Altman said in May that he was excited to see what “tokenmaxxing startups” would build and how they would work internally. Nvidia chief executive Jensen Huang argued that if a highly paid engineer was not also using a large amount of tokens, something was wrong. Meta, according to the report, even ran an internal competition that rewarded token usage.
That framing helped create a culture in which spending on AI was treated as a badge of ambition. The archetypal heavy user was not simply automating routine tasks but orchestrating fleets of AI agents continuously. The assumption was that saturation would uncover value somewhere: faster coding, more content generation, better research, or entirely new product features.
For model providers, that surge in activity translated directly into revenue. But for enterprise buyers, the economics look different. They carry the token bills, integration costs, governance burdens, and data risks. Once those expenses became concrete rather than hypothetical, executives had to ask a less fashionable question: which uses are actually worth paying for?
Why skepticism is growing
The backlash described in the report does not amount to an anti-AI turn. It is more specific than that. Companies still want the gains promised by generative models, but they are becoming less willing to assume that maximum usage is the same as smart usage. That distinction matters because token-based billing can encourage behavior that is easy to scale but hard to justify.
Microsoft chief executive Satya Nadella, in comments cited by Fast Company, acknowledged that tokenmaxxing can be addictive while arguing that customers may be paying twice: once for the tokens themselves and again by giving providers access to proprietary data. While Nadella was also promoting Microsoft’s own strategy, the critique points to a broader enterprise concern. Cost is only one side of the equation. Control over sensitive information is another.
Palantir chief executive Alex Karp used even sharper language, saying something had gone “completely wrong” and channeling what he described as frustration among businesses that felt they were spending heavily for little return. The common thread across these reactions is not rejection of AI as a category. It is rejection of undirected AI spend.
The metric problem in enterprise AI
One reason tokenmaxxing spread so quickly is that tokens are measurable. In a period when many companies were under pressure to demonstrate AI momentum, usage counts offered a visible number. Boards and executives could point to adoption figures, volume of generated output, or expanding internal use cases as evidence that the organization was keeping pace.
But measurable does not necessarily mean meaningful. Token totals say little about quality, time saved, error rates, customer outcomes, or revenue impact. They can even obscure waste by making activity look like progress. A team can produce more drafts, more analyses, and more synthetic research with AI while doing little to improve core performance. The result is a familiar enterprise trap: optimizing for the metric that is easiest to report instead of the one that best reflects value.
This is where the shift described by Moody’s becomes strategically important. If organizations move from maximizing consumption to evaluating return, procurement and product decisions could change quickly. The winners would not simply be the model vendors with the biggest raw usage numbers, but the companies that help customers use models selectively, safely, and with clearer economic logic.
What comes next
The likely outcome is not a collapse in enterprise AI demand. It is a move toward tighter controls, more targeted deployments, and greater scrutiny of where tokens are spent. Organizations may prioritize use cases with direct payoff, such as coding assistance, document workflows with measurable cycle-time reductions, or domain-specific copilots integrated into revenue-generating processes. Experiments that generate impressive volumes but vague business benefits may face sharper internal resistance.
That would mark a more mature phase of enterprise adoption. The early period of generative AI in many workplaces rewarded exuberance and visible experimentation. The next period may reward restraint, measurement, and architecture choices that keep costs predictable. In that sense, the decline of tokenmaxxing is less a retreat than a normalization. The technology is moving out of its performative phase and into the procurement logic that governs most enterprise software.
The important signal from the Fast Company report is not that companies are using less AI because interest has evaporated. It is that they are beginning to ask whether each token spent produces durable value. For an industry built on scale, that may become the question that matters most.
This article is based on reporting by Fast Company. Read the original article.
Originally published on fastcompany.com
