The gold rush surrounding artificial intelligence has led to one of the most aggressive spending cycles in tech history. From multi-billion-dollar data center expansions to skyrocketing demand for AI chips, the industry is making enormous bets on sustained demand growth. However, a closer look reveals cracks in this assumption of limitless demand, and Anthropic, a leading AI lab, is one of the few companies taking a more cautious, grounded approach.

The Token Economy and Its Challenges

At the heart of AI economics lies the concept of tokens. Tokens are the basic unit of work in AI systems — every query, generated output, or background process burns through them. While lightweight tasks like chat exchanges consume a few hundred tokens, AI agents tasked with more complex or autonomous operations can burn through millions within hours.

This has led to surging costs across both consumer and enterprise use cases. OpenAI's ChatGPT Pro, for example, offers users unlimited messaging for a fixed monthly fee, with little regard for how many resources are being consumed. Meanwhile, advanced AI agents capable of browsing the web, writing code, or managing tasks autonomously exacerbate the problem, potentially costing far more to operate than any flat-rate plan can realistically support.

Industry Recklessness and Anthropic's Realism

The industry’s current enthusiasm is palpable, but it can also be shortsighted. During a recent podcast appearance, Anthropic's CEO highlighted this recklessness, pointing out that many companies are effectively spending without clear calculations of return on investment. While this attitude might sustain the hype cycle, it sets up businesses and the industry at large for future financial strain.

One stark example comes from Uber, where AI tools have already maxed out the company’s annual AI budget — in April. Analysis from Goldman Sachs supports this pattern, noting that businesses are overrunning initial AI inference budgets by orders of magnitude. These mounting operational expenses could soon rival traditional engineering headcount costs.

Making matters worse is the incentive to inflate token consumption. Some firms, such as Meta and Shopify, have reportedly introduced internal metrics to measure employee adoption of AI tools. However, these metrics often use token consumption as the benchmark. Nvidia’s CEO, Jensen Huang, once quipped: "If that $500,000 engineer did not consume at least $250,000 worth of tokens, I am going to be deeply alarmed." It’s easy to see how this focus on raw usage — rather than meaningful outputs — could encourage wasteful practices.

Anthropic’s Break From the Pack

Anthropic stands out by adopting a more prudent and realistic strategy. Recognizing that flat-rate pricing for AI tools is unsustainable, the company recently cut off popular third-party applications from piggybacking on its subscription plans for free. It has also moved enterprise clients to a per-token billing model, directly linking usage to cost. This shift could help businesses better budget their AI expenses and curb reckless spending.

One notable example underscores the scale of the problem. Some estimates suggest that Anthropic’s $200 Max subscription tier for individuals was costing the company $2,000 to $5,000 per user to support due to unmonitored usage. By transitioning to per-token billing, Anthropic is directly addressing an inefficiency built into the AI economy, prioritizing sustainable growth over customer retention at all costs.

The Consequences of Overestimation

As Anthropic adjusts to realistic demand, the broader AI industry is still grappling with overinvestment. The assumption that AI usage will grow endlessly underpins massive infrastructure expansion. Nvidia, for example, justifies its booming hardware sales with the idea that token consumption is only heading upwards, while data centers continue investing billions to meet anticipated capacity demands. Should the projected demand falter, this overextension could lead to severe financial consequences.

Anthropic’s CEO describes this as the "cone of uncertainty," where companies must plan today for demand that may not materialize for years. Data centers typically take one to two years to build, meaning current decisions are based on speculative forecasts. It’s a high-stakes gamble: under-investment risks alienating customers in a booming market, while over-investment risks unused capacity and wasted capital.

A Pragmatic Future

What sets Anthropic apart is its deliberate approach to matching investment with verified demand. By prioritizing transparent per-token billing rather than flat-rate plans, the company aims to marry AI infrastructure growth with real-world usage trends. This approach is designed to avoid the pitfalls of unchecked optimism and ensure that resources are not being squandered on speculative consumption.

In an industry so often driven by hype, Anthropic’s calculated strategy offers a glimpse of a more sustainable AI future. Both Anthropic and its competitors are reportedly preparing for IPOs this year. If that’s the case, their financial disclosures may reveal whether careful planning or unchecked spending will pay off in the long term.

The entire AI ecosystem, from startups to hardware giants like Nvidia, has built its current trajectory on the idea of infinite growth in token consumption. But as Anthropic begins to challenge this narrative, the industry may soon have to reckon with the reality: demand signals might be off, and companies that adjust accordingly will be the ones standing strong a few years down the line.