AI Spending Frenzy Faces Reality Check: Anthropic Leads

AI Spending Frenzy Faces Reality Check: Anthropic Leads

The artificial intelligence boom, marked by massive spending on chips and data centers, is facing a critical question: is the demand real? While many tech giants are betting big on future AI usage, one of the leading AI companies, Anthropic, is taking a more cautious approach.

Anthropic’s CEO, Dario Amodei, suggests that the current spending spree might be built on an overestimation of actual AI demand. He believes many companies haven’t fully grasped the risks involved in this rapid expansion.

At the heart of the issue is how AI is used and paid for. Every interaction with AI, from asking a question to generating a line of code, consumes ‘tokens.’ These tokens are the basic unit for measuring AI usage.

A simple chat might use a few hundred tokens, but newer, more advanced AI agents can run for extended periods, using millions of tokens without users actively monitoring them. This is like using a credit card for endless errands without setting a spending limit.

AI Budgets Already Strained

The impact of this high token consumption is already being felt. Uber’s Chief Technology Officer recently stated that their AI coding tools exceeded the company’s entire yearly AI budget by April.

This highlights a significant challenge: businesses are spending far more on AI than they initially planned. Research from Goldman Sachs indicates that companies are exceeding their AI budgets by large amounts, with AI costs potentially matching the spending on engineering staff this year.

Some companies are even encouraging employees to use more AI, not necessarily for productivity, but to boost usage metrics. Meta and Shopify have reportedly used employee leaderboards to track AI usage.

Nvidia’s CEO, Jensen Huang, has also indicated that engineers should be using a substantial amount of AI resources. However, focusing on usage rather than the actual results can lead to people finding ways to use more AI simply to meet targets.

The ‘Tokenmaxxing’ Problem

This practice, sometimes called ‘tokenmaxxing,’ is where the system can be easily gamed. When the goal becomes simply to consume tokens, there are many simple ways to do that, according to Databricks CEO Ali Ghodsi. His company helps manage AI workloads for many businesses and sees this trend firsthand.

Eric Glyman, CEO of Ramp, a company that tracks spending for thousands of businesses, points out that efficiency is key. Using the most powerful AI model to edit an email might be overkill and unnecessarily expensive.

Another major issue is the ‘unlimited usage’ model, which has been common for consumers. Services like ChatGPT Pro offer unlimited use for a set monthly fee.

However, Anthropic is finding this model unsustainable, especially as AI evolves from simple chatbots to more complex agents. These agents consume tokens much faster than basic chatbots.

Anthropic Shifts to Per-Token Billing

Anthropic has stopped offering unlimited subscriptions for popular third-party tools like OpenClaw. One estimate suggested that a $200 monthly plan could actually cost the provider $2,000 to $5,000 in computing power.

The company is also moving its enterprise clients away from flat-rate plans to billing based on the exact number of tokens used. When a direct cost is attached to every token, consumers and companies that expected a fixed monthly bill might reduce their AI usage.

This shift is crucial because the entire AI investment cycle relies on token demand. Companies like Nvidia sell billions of dollars in chips based on this expected usage.

Data center companies build massive new facilities, and major tech firms invest billions in infrastructure, all assuming that AI usage will continue to grow rapidly. If a significant portion of this usage is driven by employees gaming metrics, AI agents stuck in loops, or companies overspending budgets they cannot maintain, then the vast infrastructure being built might be sized for a demand that isn’t real.

The ‘Cone of Uncertainty’

Amodei refers to this situation as a ‘cone of uncertainty.’ Building data centers takes one to two years. Right now, AI companies are making huge financial commitments based on demand that has not yet fully appeared. Investing too little risks losing customers, while investing too much could lead to revenues that never materialize, causing the financial models to fail.

Anthropic is the first major AI lab to adjust its strategy. By eliminating flat-rate plans, billing for each token, and building capacity based on verifiable demand, they are taking a different path.

Misjudging the timing by even a couple of years could be disastrous for companies making such large investments. Both Anthropic and another major AI firm are expected to go public soon.

Market Impact and Investor Outlook

Anthropic’s approach of providing clear data on per-token usage and understanding what customers are truly paying for could set it apart. The market is already starting to reward companies that set realistic goals and focus on delivering real value. Underpromising and overdelivering is often a safer and more sustainable strategy for long-term success.

The company that prices its services based on actual, verifiable demand will likely present a much stronger financial picture than those that continue to bet on unproven, inflated usage figures. As the AI race continues, the focus on sustainable business models and realistic demand forecasting will become increasingly important for investors.

The next major AI earnings reports will likely show how these different strategies are playing out in the market.

Source: AI Demand Is Overstated — Only Anthropic Is Being Realistic (YouTube)