The Token Trap: Why AI's Favorite Metric Doesn't Mean What You Think
How rising token counts became the new "eyeballs"—and why that should worry investors
During the late 1990s dotcom boom, internet companies justified soaring valuations with metrics like “eyeballs,” “page views,” and “unique visitors.” The underlying assumption: engagement metrics would eventually translate to profits. That assumption proved catastrophically wrong for most. In 2025, AI companies are doing something similar with tokens—the snippets of text that large language models process. Google reports 1.3 quadrillion tokens processed monthly, an eight-fold increase since February. Alibaba says its token use doubles every few months. OpenAI lists 30 customers each processing over a trillion tokens. These numbers sound impressive. They’re supposed to signal surging AI adoption and justify the industry’s spending levels. But the relationship between token growth and actual demand is more complicated than headlines suggest. And the connection between tokens and profits is weaker still.
Why Token Counts Are Misleading
Three factors drive token growth, only one of which represents genuine increased usage. First, actual adoption. More people using AI tools for more tasks generates more tokens. This is the growth everyone wants to see—it suggests AI is becoming essential to workflows and creating real value. Second, AI integration into existing products. Social media platforms use models to improve recommendations and image quality. Google deploys them for AI Overviews that summarize web pages instead of showing link lists. According to Barclays, these summaries account for over two-thirds of Google’s total token output. These features may improve user experience, but they don’t create new revenue. They consume tokens processing tasks that previously happened without AI, adding cost without adding income. Third, model verbosity. As LLMs become more sophisticated, they produce longer answers. EpochAI research finds that average output token counts for benchmark questions have doubled annually for standard models. “Reasoning” models that explain their approach step-by-step use eight times more tokens than simpler models—and their usage is rising five-fold yearly. This trend will accelerate. Newer models are optimized for quality and comprehensiveness, not brevity. They’re designed to provide detailed, thorough responses. Each improvement in capability tends to increase token generation per query. The result: token counts surge even when actual query volume stays flat. You’re not necessarily doing more with AI; the AI is just talking more.
The Cost Problem
Token prices have collapsed. The cost per token to answer a PhD-level science question as proficiently as GPT-4 has fallen about 97% annually, according to industry analysis. You might assume this makes AI cheap. It doesn’t. Generating responses remains expensive because models keep improving—and growing. As Wei Zhou of SemiAnalysis notes, even as token prices fall, better and more verbose models mean more tokens must be generated to complete any given task. So the marginal cost of providing AI services doesn’t approach zero. It stays significant because capability improvements offset price reductions. This creates margin pressure. OpenAI charges developers about $5-15 per million tokens depending on the model. DeepSeek, a Chinese competitor, offers comparable capability at a fraction of that price. According to recent comparisons, DeepSeek’s pricing can be 10-20x cheaper than OpenAI’s premium models. When quality differences narrow, price becomes the deciding factor. And many users are increasingly willing to trade slight quality reductions for substantial cost savings. The competitive dynamics look worrying for model providers. They’re in a race where improving quality requires larger, more expensive models that generate more tokens per query. But pricing pressure from low-cost competitors limits how much they can charge per token. Costs rise while prices fall—classic margin compression.
The Profitability Question
Sam Altman has warned OpenAI investors to expect years of heavy losses. The Wall Street Journal reports that by 2028, OpenAI expects operating losses to reach $74 billion—around three-quarters of projected revenue. Those aren’t startup losses. That’s a company spending $4 for every $3 it earns, sustained at scale. The broader industry shows similar patterns. Most AI companies generating substantial token volumes aren’t profitable on those operations. They’re burning capital to capture market share, betting that scale will eventually produce sustainable economics. This might work if token costs fell dramatically or if pricing power increased substantially. Neither seems likely. Competition from open-source models and low-cost providers like DeepSeek prevents pricing increases. And as models get more capable, they consume more compute per token, preventing dramatic cost reductions. The result: tokens may be the new currency of AI, but they’re a currency that doesn’t generate sustainable profits yet.
What About Enterprise Sales?
Some argue that consumer-facing AI services represent just early adoption, and real revenue will come from enterprise deployments where customers pay premium prices for reliability, security, and support. This thesis has merit. Enterprises do pay more for SaaS tools than consumers pay for equivalent functionality. They value integration, uptime guarantees, and vendor support. But enterprise sales also face token economics challenges. Large companies negotiating annual contracts want predictable costs. Token-based pricing creates unpredictability—costs can spike if usage patterns change or if models become more verbose. To manage this, enterprise contracts often include token allotments or caps. The vendor essentially pre-sells tokens at a fixed price, absorbing the risk that actual costs might exceed revenue. This shifts margin pressure back onto AI providers. Enterprise customers also have more leverage to demand custom models or self-hosting options. They can credibly threaten to build internal AI capabilities or switch providers. This limits pricing power even in premium segments.
The Metrics That Actually Matter
If tokens are a misleading indicator, what should investors and operators watch instead? Revenue per customer, not tokens per customer. How much money does each account generate, regardless of how many tokens they consume? This measures willingness to pay. Gross margin, not token volume. After accounting for all compute costs, how much profit remains? This measures economic viability. Retention rates, not token growth. Do customers renew subscriptions? Do they expand usage over time? This measures value creation. Competitive moat, not capability benchmarks. Can the company sustain pricing power, or will commoditization force margins toward zero? This measures long-term viability. None of these metrics look particularly favorable for most AI companies right now. Revenue growth is strong but often comes from unsustainably low pricing. Gross margins are thin or negative. Retention data is sparse since many products launched recently. Competitive moats are unclear when open-source alternatives exist.
The Path Forward
AI companies face a choice. They can compete on price, accepting low margins and hoping scale eventually produces profitability. Or they can compete on differentiation, building specialized capabilities that justify premium pricing. The first path leads toward utility-style businesses with low margins and slow growth. The second path requires finding defensible niches where competitors can’t easily replicate capabilities. Most companies are pursuing the first path because it’s faster. Scaling token volumes is easier than developing unique, hard-to-copy features. But it’s questionable whether that path leads to sustainable businesses. The dotcom parallel is instructive. Many internet companies in the late 1990s reported surging page views and unique visitors. Those metrics proved hollow when revenue models collapsed. The companies that survived built actual business value—sticky products, network effects, or genuine operational advantages. AI companies reporting surging token counts need to explain how those tokens translate to sustainable competitive advantage. Without that explanation, high token volumes are just vanity metrics. For investors, the lesson is clear: be skeptical of token growth as a success indicator. Ask instead about economics, differentiation, and long-term defensibility. Those questions mattered in the dotcom era, and they matter now. The technology may be different, but the fundamentals of sustainable business haven’t changed.