Perspective: AI demand is inflated, and only Anthropic is being realistic

The apparent surge in demand for artificial intelligence may be more illusion than reality. Anthropic, a notable player in the AI sector, has adopted a pricing strategy that reflects a more cautious outlook, potentially positioning itself advantageously should a market correction occur. Tokens, the fundamental units of AI interactions, consist of the words and characters that form user queries and the responses generated by models. Engaging with an AI typically consumes a few hundred tokens per paragraph, while more complex tasks—like coding and executing intricate workflows—can expend thousands of tokens in a single session. At the current rates set by Anthropic's latest model, one million tokens for input prompts costs around $5, while output responses come to $25. AI companies often cite the rising consumption of tokens to justify the massive investments flowing into infrastructure. However, this metric may be misleading. Major companies like Meta and Shopify have instituted internal systems to monitor token usage among employees. Nvidia's CEO, Jensen Huang, expressed concern that engineers earning hefty salaries might not be generating proportional value based on their AI consumption metrics. As firms increasingly focus on measuring AI adoption through volume, there is a risk that employees may prioritize spending over meaningful outcomes. Databricks CEO Ali Ghodsi highlighted that it’s easy to generate high costs without achieving substantial results. He noted that employees might simply re-submit queries multiple times, inflating costs without enhancing productivity. Jen Stave, executive director of the Harvard Business School AI Institute, echoed these sentiments, sharing that many CTOs and CIOs struggle to establish a return on investment (ROI) framework for their AI initiatives. Anthropic is preparing for the possibility that optimistic demand forecasts may not materialize. CEO Dario Amodei introduced the concept of a 'cone of uncertainty,' pointing out that the infrastructure for AI takes time to build, and companies are risking billions on unpredictable demand. Amodei warned that miscalculations in demand forecasting could have severe consequences. Anthropic has responded by shifting from a flat-rate pricing model to a per-token billing system, aligning revenue generation more closely with actual usage. This change also involved cutting off certain third-party tools that were high token consumers, while competitors like OpenAI have aimed to make AI more accessible and affordable. Historically, flat-rate pricing dominated the early adoption phase of AI, providing users with generous access for a fixed monthly fee. However, as usage patterns shifted towards more agentic applications, the costs skyrocketed to levels that the original pricing models could not sustain. Anthropic's most comprehensive consumer package, the $200-a-month Max plan, was frequently exploited through third-party tools that were not designed for such extensive use. On April 4, Anthropic intervened by discontinuing access to these tools, emphasizing that their pricing structures were incompatible with the usage patterns observed. Similarly, changes in enterprise contracts have begun to reflect this shift, moving away from flat monthly fees to a model that charges based on actual token consumption. As the industry grapples with these changes, OpenAI's Nick Turley acknowledged that unlimited plans might soon become impractical, drawing a parallel to unlimited electricity plans. With each token now incurring a cost, organizations that previously relied on flat-rate models will scrutinize the value they derive from AI. Eric Glyman, CEO of Ramp, pointed out that AI spending among their clients has surged 13 times over the past year, highlighting the difficulty in budgeting for these expenses. He praised Anthropic's strategic approach as more sustainable in the long term. With both Anthropic and OpenAI eyeing potential IPOs, investors will be keen to assess the validity of current demand. Anthropic's shift to per-token billing may yield clearer insights into customer value, while OpenAI may present impressive figures but face challenges in validating their authenticity. If a significant portion of AI demand is indeed inflated, the company that has priced its services realistically could emerge as a leader once the market recalibrates.