The Flaw in Token Consumption Metrics
While artificial intelligence usage is characterized by explosive growth on paper, many industry experts suggest that the reported demand may be significantly exaggerated. The basic unit of AI interaction is the token—the words and characters used in both user queries (prompts) and model responses. While a simple chat consumes only a couple hundred tokens per paragraph, complex tasks utilizing Agentic AI, such as models writing code or executing multi-step web workflows, can consume thousands more per session.
AI companies currently rely on token consumption to justify the hundreds of billions of dollars being poured into necessary infrastructure. However, this metric is increasingly considered distorted because it measures how much an engineer spends on AI rather than what productive output they generate with it. For instance, Nvidia CEO Jensen Huang stated he would be “deeply alarmed” if a highly paid employee earning $500,000 annually was not utilizing at least $250,000 worth of computing power.
“Once companies start measuring AI adoption by volume, employees optimize for the metric instead of the outcome,” noted the industry critique. Ali Ghodsi, CEO of Databricks, pointed out that it is simple to waste money in this environment; he added, “Resubmit the query to ten places. Put up a loop that just does it again and again. It’s going to cost a lot of money and not lead to anything.”
Anthropic’s Realistic Approach to Uncertainty
In contrast, Anthropic is planning for the possibility that current demand projections are incorrect. CEO Dario Amodei described this as a “cone of uncertainty,” noting that since data centers require one to two years to construct, companies are committing billions now for future demand they cannot yet verify. He warned that miscalculating capacity could be financially devastating: “If you’re off by a couple years, that can be ruinous.” Amodei expressed the impression during a February appearance on the Dwarkesh Patel podcast that some competitors were pursuing growth based merely on novelty rather than sound financial planning.
Anthropic’s strategy reflects this cautious stance. The company has shifted away from flat-rate enterprise pricing toward per-token billing, ensuring that collected revenue directly correlates with actual usage. Furthermore, Anthropic removed certain third-party tools that were major token consumers. This move was preceded by a consumer case study where the $200-a-month Max plan was exploited through agentic tools like OpenClaw. Based on Anthropic’s latest rates, heavy users could incur costs up to $5,000 without a subscription; consequently, Anthropic cut off these specific third-party integrations on April 4.
Industry Recalibration and Financial Pressure
The shift toward usage-based pricing is spreading across the sector. Older enterprise contracts at Anthropic included standard and premium seats with fixed monthly fees and built-in allowances, but these are now classified as “legacy seat types that are no longer available for new Enterprise contracts.” Newer plans require payment per seat, plus API rates based on token consumption.
This move is not unique. OpenAI’s Nick Turley, head of ChatGPT, acknowledged the economic shift during a BG2 podcast, stating that in the current climate, having an unlimited subscription plan “is like having an unlimited electricity plan. It just doesn’t make sense.”
From a financial perspective, Ramp CEO Eric Glyman observed that while AI spending among his customers has grown 13 times over the last year, there is no clear path for budgeting. He views Anthropic’s model as the more sensible long-term strategy and questioned whether companies whose business models rely on maximizing token expenditure have any incentive to promote customer efficiency.
Other major firms are adopting similar metrics; Salesforce, for example, is introducing “agentic work units” to measure the actual work completed by AI rather than simply counting tokens consumed. While both Anthropic and OpenAI are anticipated to pursue Initial Public Offerings (IPOs) this year, Anthropic’s commitment to per-token billing provides clearer data regarding customer value. If a significant portion of today’s reported AI demand is inflated, the company that has priced its services for market reality will be best positioned when the inevitable correction arrives.
