The proclaimed demand for artificial intelligence, while seemingly robust on paper, may be significantly inflated. Anthropic, by strategically pricing its services for a more grounded reality, appears best positioned to weather any potential market correction.
At the core of AI usage lie “tokens” – the fundamental units representing words and characters that constitute both user queries and model outputs. A casual paragraph-long interaction with an AI typically consumes a few hundred tokens. However, the advent of agentic AI, which empowers models to write code, browse the web, and execute complex multi-step workflows, dramatically escalates consumption to thousands of tokens per session.
To illustrate, Anthropic’s latest model pricing dictates that one million input tokens (prompts) cost $5, while one million output tokens (the model’s responses) are priced at $25. AI companies frequently cite this surge in token consumption as justification for the immense capital being invested in the underlying infrastructure.
However, a growing consensus suggests that token consumption is becoming an increasingly distorted metric. Industry giants like Meta and Shopify are reportedly implementing internal leaderboards to track employee token usage. Nvidia’s CEO, Jensen Huang, famously stated he would be “deeply alarmed” if a highly compensated engineer wasn’t utilizing at least $250,000 worth of compute – a measure focused on expenditure rather than tangible output.
The underlying issue is that when companies begin to quantify AI adoption based purely on volume, employees naturally optimize for that metric, potentially at the expense of genuine productivity. Ali Ghodsi, CEO of Databricks, which manages AI workloads for numerous enterprises, commented, “If your goal is to just burn a lot of money, there are easy ways to do that. 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.”
This sentiment is echoed by enterprise leaders. Jen Stave, executive director of the Harvard Business School AI Institute, notes, “I’ve talked to a dozen CTOs or CIOs who are all saying, ‘Actually, I’m having a really hard time finding an ROI framework for this.'”
Anthropic is proactively preparing for the possibility that current demand projections may be overly optimistic. CEO Dario Amodei has articulated his concern about a “cone of uncertainty,” emphasizing that data center construction takes one to two years, necessitating substantial upfront investment for demand that remains unverified. Over-provisioning risks financial strain if demand fails to materialize, while under-provisioning could lead to lost customers due to capacity limitations.
“If you’re off by a couple years, that can be ruinous,” Amodei remarked on a recent podcast. “I get the impression that some of the other companies have not written down the spreadsheet. They’re just doing stuff because it sounds cool.”
In response, Anthropic has shifted from flat-rate enterprise pricing to a per-token billing model, ensuring that revenue directly correlates with actual usage. Furthermore, the company has curtailed access for certain third-party tools that were significant token consumers, a stark contrast to OpenAI’s strategy of making AI more affordable and accessible at scale.
The early stages of AI adoption were largely characterized by flat-rate pricing, offering generous or unlimited AI access for a fixed monthly fee. While this model proved effective for conversational AI, the rise of agentic AI, where sessions can consume millions of tokens, has disrupted the underlying economics.
Anthropic’s $200-a-month Max plan became a notable case study. Developers were leveraging this subscription through third-party agentic tools like OpenClaw, effectively running AI agents around the clock on a plan originally designed for conversational use. Based on Anthropic’s published rates for its latest model, a power user of Claude Code Max might have been paying as little as $200 per month for usage that would have otherwise cost upwards of $5,000.
On April 4, Anthropic ceased support for these tools. Boris Cherny, head of Claude Code, explained on social media that the subscriptions “weren’t built for the usage patterns of these third-party tools.”
A similar recalibration is underway within enterprise solutions. Older Anthropic enterprise contracts included standard and premium seats with fixed monthly fees and bundled usage allowances. These are now designated as “legacy seat types that are no longer available for new Enterprise contracts.” Current enterprise plans are priced per seat, with token consumption billed separately at API rates.
While Anthropic has taken the lead in this transition, industry-wide pressure is mounting. Nick Turley, head of ChatGPT at OpenAI, acknowledged on a recent podcast that “it’s possible that in the current era, having an unlimited plan is like having an unlimited electricity plan. It just doesn’t make sense.”
As every token now carries a direct cost, companies and consumers who budgeted for flat-rate AI will inevitably scrutinize their return on investment. Eric Glyman, CEO of Ramp, which recently introduced a token-tracking tool, observes this dynamic from the financial perspective. AI spending across Ramp’s customer base has surged thirteenfold over the past year, with many struggling to budget effectively. He views Anthropic’s approach as the more prudent long-term strategy and raises a critical question for OpenAI’s investors: if a business model hinges on maximizing token expenditure, does it inherently disincentivize helping customers optimize AI efficiency?
Salesforce is adopting a similar strategic stance, introducing a new metric called “agentic work units” to track the actual work completed by AI, rather than simply the tokens consumed.
With both Anthropic and OpenAI anticipated to pursue initial public offerings this year, the question of AI demand will be a primary focus for public market investors. Anthropic’s move to per-token billing will provide clearer data on customer value, while OpenAI, despite potentially larger reported numbers, may face greater challenges in demonstrating the authenticity of its demand. In the event of a market correction, the company that has strategically priced for reality is more likely to emerge resilient.
Original article, Author: Tobias. If you wish to reprint this article, please indicate the source:https://aicnbc.com/20775.html
