The Token Paradox: Why 62% Usage Generates Only 8.6% of AI Revenue

The numbers landed in my inbox at 6:47 AM Stockholm time, and they did not make sense. Not at first glance, anyway. Vercel's latest platform telemetry showed open-source models now account for 62 percent of all token consumption on their network. Two months ago, that figure stood at 28.4 percent. A doubling in sixty days. The kind of hockey-stick curve that usually precedes a correction, not a confirmation. But here is the part that stopped me cold: those same open-source models generate only 8.6 percent of total spending on the platform. Sixty-two percent of the traffic. Eight-point-six percent of the money. The thesis held firm when the charts turned red, but this divergence is not a market anomaly. It is a structural revelation about how the AI economy actually works. I have spent the better part of a decade auditing blockchain protocols and mapping token flows, and I have learned that the most revealing metric is almost never the one the marketing team leads with. In crypto, we learned this the hard way in 2017, when ICO whitepapers promised decentralized utopias while their tokenomics revealed centralized extraction. The same pattern is now playing out in the AI model market, and the Vercel data is the audit trail. The gap between usage share and revenue share is not a bug. It is the entire story. Let me establish the context before I deconstruct the numbers. Vercel is not a random sample. It is the deployment layer for a significant portion of the modern web application stack, particularly among developers building AI-powered features into their products. The platform's telemetry captures real production traffic, not benchmark scores or marketing claims. When a developer integrates a model into their application and ships it to production, Vercel sees the tokens flow. This is the closest thing we have to ground truth on actual model usage patterns in the developer economy. The platform's user base skews toward web application and frontend development, which means the data overrepresents code generation, content synthesis, and lightweight inference tasks. It underrepresents enterprise-grade complex workflows. I am flagging this bias now because it matters for the interpretation that follows, and because any analyst who ignores platform selection effects is not an analyst. They are a propagandist. The core finding is this: open-source models have crossed a threshold that most industry observers did not expect to see until 2027 or later. The 28.4 to 62 percent jump in token share within two months represents a critical mass shift in developer trust. This is not a gradual adoption curve. This is a phase transition. Developers are not experimenting with open-source models in sandboxes anymore. They are shipping them to production, at scale, and the usage data reflects that reality. DeepSeek, in particular, has surpassed Google to become the second-largest model provider on the platform by token consumption. Let that sink in for a moment. A Chinese open-source model, released with an API price that undercuts Western competitors by an order of magnitude, is now consuming more tokens than Google's Gemini family. Google. The company that invented the transformer architecture that made the entire modern AI revolution possible. The company whose research papers are the foundation upon which every model in this market is built. And DeepSeek is eating their lunch on actual production usage. The spending data, however, tells a different story. Anthropic accounts for 30 percent of token consumption but 65.1 percent of spending. That means Anthropic's effective price per token is roughly two times the market average, and approximately fifteen times the average price of open-source models. The math is brutal and clarifying. Open-source models are being used for high-frequency, low-complexity tasks where cost sensitivity dominates. Anthropic's models are being used for high-value, complex reasoning tasks where quality justifies the premium. The market is not choosing between open and closed source. It is segmenting into two distinct tiers with different economic logics. This is where my audit instincts kick in. When I see a 62-to-8.6 divergence, I do not ask whether open source is winning. I ask what kind of tasks are driving that 62 percent, and what their value density actually is. The available evidence suggests that open-source token consumption is concentrated in code completion, text classification, information extraction, and other mid-to-low complexity tasks. These are the workhorses of the AI application layer. They are not glamorous. They do not require deep reasoning chains or complex tool use. But they are the tasks that developers need to execute millions of times per day, and at open-source prices, they become economically viable at scale. The price elasticity effect here is significant. When the marginal cost of a token drops by 90 percent or more, developers start using models for tasks they previously would have handled with deterministic code or simply skipped. This is not substitution. This is net-new demand creation. The 59 percent month-over-month growth in total token volume on Vercel is evidence of this effect. Cheaper tokens do not just steal market share from expensive tokens. They expand the total addressable market for AI inference. But here is the uncomfortable question that the bull case for open source does not want to confront: what happens when the low-hanging fruit is harvested? The 62 percent token share is impressive, but if those tokens are predominantly low-value, high-volume tasks, then the economic ceiling for open-source model providers is fundamentally constrained. The spending data already reflects this. Eight-point-six percent of revenue for 62 percent of usage is not a sustainable business model unless the cost structure is equally asymmetric. And that is the question I cannot answer from the Vercel data alone. What is DeepSeek's actual gross margin? What is their cost per token in terms of compute, electricity, and infrastructure? If they are pricing below cost to acquire market share, then this entire narrative is a subsidy-driven illusion that will collapse when the funding runs dry. If they have achieved genuine architectural efficiency through mixture-of-experts design and inference optimization, then the open-source advantage is real and durable. The distinction matters enormously for anyone making investment decisions in this sector. Let me now address the competitive dynamics, because the Vercel data reveals a more nuanced picture than the simple open-versus-closed binary. The first casualty of this data is Google's positioning. Being surpassed by DeepSeek in token consumption is not a minor embarrassment. It is a signal that Google's models are not resonating with the developer community in the way that their research prestige would suggest. Google has the best AI research lab in the world. They have the distribution advantages of their cloud platform and their consumer products. And yet, on a platform that represents the cutting edge of web application development, their models are being out-consumed by a Chinese open-source competitor. This suggests a fundamental disconnect between Google's technical capabilities and their product-market fit in the developer segment. The pricing of Gemini, the API ergonomics, the model performance on real-world tasks, or some combination of these factors is not competitive. Google's s chaos. The company that should be dominating this market is losing share to a model that costs a fraction of what Gemini costs. OpenAI occupies an increasingly uncomfortable middle position. The Vercel data does not break out OpenAI's specific numbers, but the aggregate picture suggests they are growing in absolute terms while losing relative share. OpenAI is caught between Anthropic's premium positioning and the open-source cost advantage. Their models are not cheap enough to compete with DeepSeek on price-sensitive tasks, and they are not differentiated enough to command Anthropic's premium on high-value tasks. This is the classic sandwich problem, and it is not a comfortable place to be in a rapidly commoditizing market. OpenAI's brand recognition and enterprise relationships provide some insulation, but the structural pressure is real and intensifying. Anthropic, by contrast, has achieved something remarkable. They have positioned themselves as the quality leader in a market that is racing to the bottom on price. The 65.1 percent spending share on 30 percent token share is not just a pricing power story. It is evidence that a significant segment of developers and enterprises are willing to pay a substantial premium for models that deliver superior performance on complex tasks. This is the same dynamic we saw in the early days of enterprise software, where companies like Oracle and SAP commanded premium valuations not because their products were cheaper, but because they solved problems that cheaper alternatives could not. Anthropic has found their moat, and it is not in cost efficiency. It is in capability differentiation. The industry structure that is emerging from this data is a two-tier market. The first tier is the high-volume, low-margin segment dominated by open-source models. This is where the token volume lives, where the growth is happening, and where the competitive dynamics are most brutal. The second tier is the high-value, high-margin segment dominated by a small number of closed-source providers, with Anthropic currently leading the pack. The prediction that closed-source models will eventually account for only 15 to 25 percent of token volume while capturing 60 to 90 percent of economic value is consistent with the current trajectory. This is not a prediction of open-source failure. It is a prediction of market segmentation. Open source wins the volume game. Closed source wins the value game. And the two games have different rules, different economics, and different competitive dynamics. Now let me offer the contrarian angle, because every good audit requires a stress test of the prevailing narrative. The dominant interpretation of this data is that open source is winning and closed source is losing. I think that interpretation is incomplete and potentially misleading. The more accurate reading is that the AI model market is bifurcating into two distinct businesses with different economic characteristics, and the metrics that matter for one are almost irrelevant for the other. Token volume is a vanity metric for closed-source providers. Spending share is a vanity metric for open-source providers. The real question is not which approach is winning. The real question is which approach is building a sustainable economic engine. Here is the blind spot that most analysts are missing. The open-source token surge is being driven by a price point that may not be sustainable. If DeepSeek and other open-source providers are operating at negative gross margins, then the 62 percent token share is a temporary phenomenon that will reverse when the subsidy ends. The history of technology markets is full of examples where free or heavily subsidized products achieved dominant usage share, only to see that share evaporate when the economics normalized. The question is not whether open-source models are technically competitive. They clearly are. The question is whether the current pricing is a reflection of genuine cost advantages or a strategic decision to buy market share. The answer to that question determines whether the current market structure is stable or transitional. There is also a platform bias that deserves scrutiny. Vercel's developer base is not representative of the entire AI model market. It skews toward web application development, which means it overrepresents tasks like code generation, content creation, and lightweight inference. It underrepresents enterprise workflows, complex reasoning tasks, and regulated industries where data sovereignty and compliance requirements favor closed-source providers. The true market share of open-source models across the entire AI economy is likely lower than the Vercel data suggests. The trend is real, but the magnitude may be overstated. I would estimate that the actual open-source share of the total AI inference market is somewhere in the 40 to 50 percent range, with significant variation by task type and industry vertical. The investment implications of this data are substantial. The valuation logic for AI model providers is shifting from usage growth to revenue quality. A company that generates massive token volume but minimal revenue per token is now valued more like an infrastructure provider than a high-margin software company. This is a fundamental re-rating that the market is only beginning to process. Anthropic's high spending share supports a premium valuation because it demonstrates that the market is willing to pay for quality. DeepSeek's high token share but low spending share creates a valuation paradox. Are they a growth story or a commodity business? The answer depends entirely on their cost structure and their ability to move up the value chain into higher-complexity tasks. The regulatory dimension adds another layer of complexity. The international spread of Chinese open-source models raises concerns that go beyond market competition. Data security, model controllability, and geopolitical implications are all part of the picture. Western regulators are beginning to ask questions about the deployment of models that originate from jurisdictions with different data protection standards and different approaches to AI governance. This is not a technical problem. It is a policy problem that could reshape the competitive landscape through regulation rather than market forces. The open-source community has historically resisted regulatory intervention, but the scale of deployment that the Vercel data reveals is making that position increasingly untenable. Let me now address the question that I think matters most for the next phase of this market. The current two-tier structure is not static. The open-source models that are winning on volume today are not standing still on capability. DeepSeek's rapid iteration cycle, the release of increasingly capable open-weight models from multiple providers, and the growing ecosystem of fine-tuning and deployment tools are all pushing open-source capability upward. The question is whether open-source models can close the gap on the complex reasoning tasks that currently drive Anthropic's premium pricing. If they can, the spending distribution will shift, and the economic foundation of the closed-source premium will erode. If they cannot, the two-tier structure will persist, and the market will settle into a stable equilibrium where open source handles the volume and closed source captures the value. My assessment, based on the data and my experience auditing technology markets, is that the gap will narrow but not close. The reasoning and complex task capabilities that justify Anthropic's premium are not just a function of model architecture. They are a function of data quality, alignment investment, safety engineering, and continuous improvement processes that are difficult to replicate in an open-source model. Open source will continue to improve, and the volume segment will continue to grow, but the high-value segment will remain the domain of a small number of well-capitalized closed-source providers. The market will not become a winner-take-all monopoly, but it will also not become a fully commoditized market. The structure will be oligopolistic at the top and fragmented at the bottom. The Vercel data is a snapshot of a market in transition. The 62 percent token share for open-source models is a genuine milestone, but it is not the end of the story. It is the beginning of a new phase in which the competitive dynamics shift from raw capability to economic sustainability. The models that win the next phase will be the ones that can deliver value at a price that reflects their true cost structure, whether that is a premium price for premium capability or a commodity price for commodity performance. The market is sorting itself out, and the sorting process is not always pretty. Companies that cannot find their position in the two-tier structure will be squeezed out. Companies that can will thrive. The thesis held firm when the charts turned red, and it holds firm now. The divergence between token share and spending share is not a market inefficiency. It is a structural feature of a market that is maturing faster than most observers expected. The next twelve months will determine which model providers have built sustainable businesses and which have built subsidized illusions. The data will tell the story, and I will be reading it with the same audit discipline that has guided my analysis through every market cycle I have witnessed. The signal is in the numbers. The noise is in the narratives. And the gap between them is where the real opportunities and the real risks live.