AI demand is inflated — only Anthropic is being realistic

The main demand signal for AI looks explosive on paper but may be significantly exaggerated. By pricing its tools based on this reality, Anthropic may be the best-positioned AI company if a correction were to come.

Tokens are the basic unit of AI usage: words and the characters that make up both the queries users submit and the output models generated.

Chatting with an AI consumes several hundred coins per paragraph. Agentic AI, where models write code, browse the web, and execute multi-step workflows, burns through thousands more per session.

Using the rates of Anthropic’s latest model, one million token inputs (requests) cost $5, while one million token outputs (the model’s responses) cost $25.

AI companies cite the boom in token consumption as a way to justify hundreds of billions of dollars being spent on the infrastructure to serve them.

But token consumption becomes a skewed measurement.

Meta And Shopify They say they create internal leaderboards that track how many tokens employees use. Nvidia CEO Jensen Huang said an engineer making $500,000 a year would be “deeply alarmed” if he didn’t use at least $250,000 worth of computing. Measuring what an engineer spends on AI instead of what he produces with AI.

As companies begin measuring AI adoption by volume, employees are optimizing based on measurement rather than outcome.

“If your goal is to burn a lot of money, there are easy ways to do that,” said Ali Ghodsi, CEO of Databricks, which processes AI workloads for thousands of businesses. “Resubmit the query to ten places. Create a loop that does this over and over again. This will cost a lot of money and yield no results.”

Jen Stave, managing director of the Harvard Business School Artificial Intelligence Institute, hears the same thing from corporate leaders.

“I’ve talked to a dozen CTOs or CIOs and they all say: ‘I’m actually having a really hard time coming up with an ROI framework for this,’” he said.

Anthropic is planning in case demand forecasts are wrong.

CEO Dario Amodei described what he calls the “cone of uncertainty”: data centers take one to two years to build, so companies are committing billions of dollars to demand they can’t yet verify. When you don’t have enough capacity, buy too little and lose customers. If you buy too much, the revenue won’t come in as planned, the math won’t work.

“If you take a few years off, it can be devastating,” Amodei said. Dwarkesh Patel podcast In February. “I get the impression that some of the other companies don’t write the spreadsheet. They just do it because it sounds good.”

Anthropic’s response was to move away from flat-rate enterprise pricing and move to per-token billing, so the revenue it collects reflects actual usage. While it also eliminates some third-party tools that consume large amounts of tokens, OpenAI makes AI cheaper and consumable at scale.

Flat-rate pricing dominated the early years of AI adoption, with fixed monthly fees for generous or unlimited AI access. This model worked when people were chatting with AI. But brokered usage turned the cost of thousands of tokens per session into millions and disrupted the economy.

Anthropic’s most generous consumer offer, the $200-per-month Maximum plan, has become a case study.

Developers were routing this subscription through third-party agent tools like OpenClaw, running their AI agents around the clock on a schedule designed for chatting. Based on Anthropic published odds For its latest model, a heavy Claude Code Max user could be paying as little as $200 per month for use that would cost the user up to $5,000 without a subscription.

On April 4th, Anthropic discontinued these tools. Boris ChernyClaude Code’s president wrote that subscriptions at X “are not designed around the usage patterns of these third-party tools.”

The same recalibration is happening in the corporate arena.

Older Anthropic contracts included standard and premium seats; Fixed monthly fees with built-in usage allowance. These are now labeled as “legacy seat types that are no longer available for new Enterprise contracts,” according to the company’s statement. support page. New enterprise plans charge per license and token consumption is billed at API rates on top.

Antropik was the first to act, but pressure is mounting across the industry.

OpenAI’s ChatGPT head Nick Turley confirmed at BG2 podcast “It’s possible that in the current era, having an unlimited plan would be like having an unlimited electricity plan. That just doesn’t make sense.”

If every token now has a price, companies and consumers who have budgeted for fixed-rate AI will begin to ask what they are actually getting in return.

Ramp CEO Eric Glyman, who recently launched a token tracker, sees the dynamic from the financial side.

AI spending across Ramp’s customer base increased 13x last year, and no one knows how to budget for it. He noted that Anthropic’s approach is a more cautious long-term strategy, and raises a question that will interest OpenAI investors: If your business model depends on achieving maximum token spend, do you have an incentive to help customers use AI more efficiently?

Salesforce is making a similar bet, introducing a new metric it calls “delegated work units” that tracks the work AI completes rather than the tokens it burns.

Both Anthropic and OpenAI are expected to IPO this year. When they do this, the first thing public market investors will try to answer is the question of demand.

By switching to per-token billing, Anthropic will have cleaner data on what its customers actually value. OpenAI’s numbers will be bigger, but it will be harder to prove how much of them are real.

Even if even a meaningful portion of today’s AI demand is inflated, the company that prices based on reality will be the company left standing when the correction comes.