The artificial intelligence boom is being measured in tokens, billed in tokens, and increasingly justified by tokens. Yet beneath the surge in usage metrics lies a growing concern across the industry: the core signal used to validate hundreds of billions of dollars in infrastructure investment may be overstating real economic demand.
Tokens, the fragments of text that make up prompts and responses, have become the de facto unit of AI consumption. Every interaction with systems built by Anthropic or OpenAI translates into token flow, and at scale, those flows are immense. Simple chat interactions consume modest volumes, but agentic systems—capable of coding, browsing, and executing multi-step workflows, multiply usage dramatically, often running continuously in the background.
At current pricing, that consumption translates directly into revenue potential. Anthropic charges $5 per million input tokens and $25 per million output tokens on its latest models. Multiply that across enterprise deployments and autonomous agents, and the numbers appear to support the industry’s vast capital expenditure on data centers, chips, and energy infrastructure.
Register for Tekedia Mini-MBA edition 20 (June 8 – Sept 5, 2026).
Register for Tekedia AI in Business Masterclass.
Join Tekedia Capital Syndicate and co-invest in great global startups.
Register for Tekedia AI Lab.
But the reliability of that signal is increasingly under scrutiny. Inside large organizations, token usage is becoming a performance metric. Meta and Shopify have introduced internal tracking systems that rank employees by how much AI they consume. Jensen Huang has gone further, suggesting he would be “deeply alarmed” if a highly paid engineer were not generating substantial AI compute spend.
Such benchmarks create a predictable distortion. When consumption is rewarded, optimization follows. Engineers and teams begin to maximize token usage rather than output quality, effectively turning AI into a budget line to be spent rather than a tool to be optimized.
Ali Ghodsi, chief executive of Databricks, has described how easily that system can be gamed. Re-running queries, duplicating workloads, or looping processes can drive up token consumption with little incremental value. The metric inflates, the bill rises, but productivity does not necessarily follow.
This disconnect is becoming visible at the executive level. Harvard Business School AI Institute executive director Jen Stave says many CIOs and CTOs are struggling to construct a credible return-on-investment framework for AI. The challenge is not adoption; tools are being deployed widely, but attribution. Companies can measure what they spend on AI; they cannot yet consistently measure what they gain.
That gap has implications that extend beyond enterprise budgets. It calls into question the demand assumptions underpinning the industry’s infrastructure buildout. Data centers require years to plan and construct, meaning today’s investment decisions are based on forecasts that may not fully account for behavioral distortions in usage.
Anthropic’s chief executive, Dario Amodei, has framed this uncertainty in operational terms, describing a “cone of uncertainty” around demand. Build too little capacity and risk losing customers; build too much and face underutilized assets and delayed revenue.
“If you’re off by a couple years, that can be ruinous,” he said, highlighting the asymmetry of the risk.
Anthropic’s response has been to tighten the link between usage and revenue. The company is moving decisively toward per-token billing, abandoning the flat-rate subscription structures that defined the early phase of AI adoption. That shift is both defensive and diagnostic: it protects margins while generating clearer data on how much customers truly value different types of AI workloads.
The transition has already exposed inefficiencies. Anthropic recently curtailed access to third-party tools that were routing heavy, continuous workloads through consumer subscription plans. In some cases, users paying $200 per month were generating usage that would have cost thousands under a metered model. The arbitrage highlighted a fundamental mismatch between pricing design and actual usage patterns.
Enterprise contracts are undergoing a similar overhaul. Legacy seat-based pricing, with bundled usage allowances, is being replaced by hybrid structures that combine per-user fees with direct billing for token consumption. The result is a model that scales revenue with compute demand but also forces customers to confront the true cost of their AI usage.
Competitors are converging on the same realization. At OpenAI, ChatGPT head Nick Turley has acknowledged that unlimited plans may be economically untenable, likening them to offering unlimited electricity in an environment where consumption can scale without constraint. The analogy is instructive: as AI shifts from occasional interaction to continuous operation, it behaves less like software and more like infrastructure.
From the financial side, the consequences are already visible. Ramp reports that AI spending across its customer base has increased thirteenfold in a year, yet budgeting frameworks remain immature. Companies are spending heavily without a clear sense of optimal allocation, a dynamic that is sustainable only as long as capital remains abundant.
That dynamic introduces a structural tension. Providers benefit from higher token consumption, but long-term adoption depends on efficiency and demonstrable value. If customers begin to optimize for cost rather than usage, revenue growth tied purely to volume could slow.
Some companies are beginning to anticipate that shift. Salesforce is experimenting with “agentic work units,” an attempt to measure AI output rather than input. The concept reframes the value equation: instead of tracking how much compute is consumed, it asks what work is actually completed.
The distinction is likely to become central as leading AI firms approach public markets. Both Anthropic and OpenAI are widely expected to pursue IPOs, where investor scrutiny will focus less on headline growth and more on the quality and sustainability of that growth. Token counts alone will not suffice; markets will demand evidence that usage translates into durable economic value.
In that environment, pricing strategy becomes a signal. Anthropic’s move toward metered billing may produce slower, more disciplined growth figures, but it also yields cleaner data and more predictable unit economics. OpenAI’s broader reach and more aggressive scaling may generate larger top-line numbers, but with greater ambiguity around how much of that demand is structural versus inflated.
The broader risk is that the industry has entered a phase where activity is being mistaken for demand. If a portion of token consumption is driven by internal incentives, experimental overuse, or poorly optimized workflows, then the true baseline for AI demand may be lower than current projections suggest.
Should that correction materialize, its effects would cascade through the system. Infrastructure investments could face underutilization, pricing models would tighten further, and companies reliant on volume growth would be forced to recalibrate.
In that scenario, the advantage shifts to those who priced for reality rather than momentum. The companies that survive will not be those that generated the most tokens, but those that understood which tokens mattered—and were paid accordingly.