Tokens processed per day are now tracking above 400 trillion, up more than 185% since October 2025 and over 4,250% versus October 2024.
The numbers point to a monumental increase in AI demand, however, demand is not necessarily the same as value creation. As enterprises spend more money on AI, a lingering and important question is whether rising token consumption is translating into equal economic returns.
This question sits at the center of an escalating debate between Palantir and leading AI labs. Palantir has taken direct shots at OpenAI and Anthropic’s business models, describing the ecosystem as a self-serving “token industrial complex.”
The point that Palantir is making is an interesting one for investors to consider, which is that OpenAI and Anthropic largely monetize increased AI usage, while Palantir is attempting to maximize the economic value enterprises generate from that usage.
The result is two very different business models—and, so far, the two business models are dramatically different in profitability.
Palantir vs AI Labs: The Battle Between Token Consumption and Value Creation
The enterprise model for AI labs is simple: more token consumption generally equals more revenue. However, this usage-driven model also could create a misalignment of interests between labs and their customers. While more token consumption directly benefits labs, it also directly increases the cost to customers, potentially damaging the economic value they generate from each token.
A recent survey from EY signals that enterprises have real questions about whether implementing AI through a usage-based model is the best approach. EY notes that 82% of senior leaders whose organizations are investing in AI are concerned about token usage and its related costs. Furthermore, 98% of these leaders say that AI token usage and related costs have caused their organizations to reconsider their approach.
Amid this, Palantir’s model for enterprise AI monetization is inherently different. Rather than focusing on higher token consumption to drive higher revenue, it aims to help customers get the highest value out of each token they consume. It then charges based on enterprise adoption of its products across different workflows.
The difference should make a customer’s decision to expand their relationship with Palantir significantly more deliberate versus doing so with AI labs. With Palantir, the costs for a business are more likely to be stable unless additional contracts are signed to expand into additional products or workflows. In that case, it is reasonable to think that enterprises have a need to see hard evidence of Palantir’s value creation before committing further.
Meanwhile, token consumption could increase passively simply from certain teams submitting more prompts, transitioning to more token-intensive models, or implementing AI agents to tackle more complex problems.
The counterargument for AI labs is that customers would not increase their token consumption unless they see value being generated. However, multiple pieces of evidence from top AI adopters and researchers, including AI labs themselves, suggest that this calculus is not so straightforward.
More AI Tokens Don’t Always Mean More Business Value
Uber provides a high-profile example of a company that consumed tokens uncontrollably with little tangible evidence of value creation. The company reportedly spent it’s entire 2026 AI coding tool budget in just the first four months of the year. Amid this, the company’s COO Andrew Macdonald went on to note, “it’s very hard to draw a line” between AI usage and the number of additional useful features the company is providing to consumers.
Research from the University of Michigan, Stanford, Google, and Microsoft sheds light on the underlying reasons why token usage can be difficult to control. Their study found that total token usage can vary by up to 30X on the same task, and that higher token usage does not translate into higher accuracy. This highlights the difficulty of estimating token costs going into a project, and thus understanding the return that could be generated.
Lastly, research from OpenAI itself shows that higher token consumption does not necessarily translate into higher customer revenue. An OpenAI-led study, which included researchers from Columbia Business School and Wharton, analyzed nearly 400 public U.S. companies in 2024 and 2025. Notably, the researchers found that “revenue per employee is not meaningfully associated with output tokens per employee or messages per active user.”
These pieces of evidence highlight several flaws in the token-consumption model, and lend credence to Palantir’s view that consuming more tokens doesn’t inherently lead to better business outcomes.
Rising token consumption alone does not tell investors which AI stocks are generating long-lasting value – but the I/O Fund’s institutional-grade research does. We publish hundreds of articles per year on the AI trade, hold weekly webinars, and offer real-time trade alerts. Subscribe now to follow the I/O Fund, with six stocks already up more than 100% YTD.
Palantir’s Land-and-Expand Model Is Driving Higher-Quality Growth
We can see how these different business models result in very different financial profiles between Palantir and AI labs.
Recent reports suggest that OpenAI recently crossed $40 billion in ARR, doubling in approximately seven months. Meanwhile, Anthropic’s ARR has reportedly hit $65 billion in July, increasing by more than 7X in seven months. Assuming these figures are calculated by multiplying monthly revenue by 12, they would imply monthly revenues of $3.33 billion for OpenAI and $5.20 billion for Anthropic.
Particularly in the case of Anthropic, it is fully possible that no other tech company in history has grown this quickly off of what was already a very sizeable revenue base to start the year.
Palantir’s revenue is growing at a rapid pace as well, with sales rising 93% YoY last quarter to $1.94 billion. However, OpenAI and Anthropic are still growing dramatically faster off larger revenue bases.
Still, the differences in their business models can help explain some of the gap. When a new enterprise begins using their AI, the consumption-based model means that revenue can begin to accrue almost instantly. This makes the AI lab business model extremely scalable, as token consumption often simply needs to continue to rise for revenue to grow.
Contrast this with Palantir, which has a much slower-moving monetization model. Notably, the company deploys forward deployed engineers (FDEs) into organizations it works with. These FDEs spend a significant portion of their time embedded on-site with customers and build customized solutions.
This should allow FDEs to create solutions that actually generate value, build trust, and then expand the relationship into more workflows. Notably, this “land-and-expand” strategy is driving the vast majority of Palantir’s commercial sales growth.
Palantir’s revenue from commercial customers increased by $495 million, or 110% YoY in Q2. Of the increase, $407 million was from commercial customers existing as of the end of 2025. In turn, 82% of Palantir’s commercial revenue growth came from existing customers; a clear representation of its land-and-expand success, and an indicator of real value creation.
Palantir’s net dollar retention rate of 157% last quarter also demonstrates this, with existing customers increasing their spend by 57% from the prior year. Furthermore, Palantir’s U.S. commercial remaining deal value (RDV), surged 124% YoY to $6.24 billion, showing that the company is rapidly growing its contracted revenue.
Meanwhile, OpenAI and Anthropic’s ARR figures provide little clarity into the quality of their sales growth. They do not show how much growth is coming from new customers experimenting versus existing customers spending more, or what percentage of sales are contractually obligated.
