The Token Share Paradox: What Vercel's Data Really Says About the Open-Source AI Takeover

Look at the numbers on Vercel's AI gateway dashboard for February 2025. Open-source models now serve 62% of all tokens. They generate 8.6% of the spending. That is not a typo. In two months, the open-source share of token volume nearly tripled from 28.4% to 62%, while the economic value captured by those same models remained a rounding error. The code does not lie, but the auditor must dig. This is not a story about open-source winning. It is a story about what we mean by winning. Vercel is not a random data source. It is the deployment layer for a significant slice of the modern web application stack. When a developer builds a Next.js app and plugs in an AI feature, the token traffic flows through Vercel's observability. The platform's user base skews toward front-end engineers, indie hackers, and mid-market SaaS teams. That means the data captures the long tail of AI application development, not the enterprise heavy lifting. It is precisely the segment where cost sensitivity is highest and switching costs are lowest. In other words, it is the perfect early-warning system for structural shifts in model adoption. The headline numbers deserve a forensic breakdown. Open-source models went from 28.4% of token share in December 2024 to 62% by February 2025. DeepSeek, the Chinese open-weight model family, overtook Google to become the second-largest model provider on the platform by token volume. Anthropic, meanwhile, holds roughly 30% of token share but captures 65.1% of all spending. The total token volume on the platform grew 59% month-over-month. These four data points, taken together, tell a more complex story than any single one of them suggests. Let me walk through the arithmetic, because the implications are not obvious at first glance. If open-source models serve 62% of tokens but only account for 8.6% of spending, their effective price per token is roughly one-fifteenth that of the closed-source average. Anthropic's models, by contrast, command a price per token that is more than double the platform average. This is not a market failure. It is a market segmentation. The open-source models are being used for high-frequency, low-complexity tasks: code completion, text classification, information extraction, boilerplate generation. The closed-source models, particularly Anthropic's Claude family, are being used for complex reasoning, multi-step agentic workflows, and tasks where a single error is more expensive than the entire API call. Tracing the gas trails back to the root cause, the 59% month-over-month growth in total token volume is the most underappreciated number in this dataset. It suggests that the availability of ultra-cheap open-source models is not simply displacing closed-source usage. It is creating entirely new demand. Developers are now using AI for tasks they would have previously done manually or skipped entirely, because the marginal cost of a token has dropped to near zero. This is the price elasticity effect that economists would predict, but it is happening at a speed that few anticipated. The open-source models are not just stealing share from the incumbents. They are expanding the total addressable market for AI inference. DeepSeek's rise deserves particular attention. In my years of auditing smart contracts and analyzing protocol architectures, I have learned to distinguish between benchmark scores and real-world performance. DeepSeek's token volume surpassing Google's on Vercel is a signal that its models are performing well in production environments, not just on leaderboards. But I would caution against over-interpreting this as a pure capability victory. The price differential is so extreme that many developers may be choosing DeepSeek not because it is better, but because it is good enough at one-tenth the cost. This is a rational economic choice, but it is not the same as a technical triumph. The deeper question is what this means for the competitive structure of the AI model market. The conventional narrative has been that open-source models are catching up to closed-source models, and that the gap will eventually close. The Vercel data suggests a more nuanced reality. Open-source models are winning the volume game, but closed-source models are winning the value game. Anthropic's 30% token share generating 65.1% of spending is not a sign of weakness. It is a sign of pricing power. The market is willing to pay a premium for models that can reliably handle complex, high-stakes tasks. This is not a commodity market. It is a differentiated market with two distinct tiers. Shifting the consensus layer, one block at a time, I see this as a structural realignment rather than a zero-sum competition. The open-source models are commoditizing the low end of the market, where tasks are repetitive and error tolerance is high. The closed-source models are consolidating their position at the high end, where quality and reliability justify premium pricing. This is similar to what happened in the database market, where open-source options like PostgreSQL and MySQL captured the bulk of deployments while Oracle continued to extract enormous profits from enterprise workloads. The same pattern is now playing out in AI inference. But there is a contrarian angle that most analysts are missing. The open-source token share growth may be masking a fragility that will become apparent in the next market downturn. When the current AI funding cycle cools, the economics of ultra-low-cost inference will come under scrutiny. DeepSeek and other open-weight providers are likely operating at or below cost, subsidized by venture capital or strategic investors. This is not a sustainable business model. It is a market-share acquisition strategy. When the subsidies dry up, prices will rise, and the token share distribution will shift again. The question is not whether open-source models will continue to grow. The question is whether their growth is built on a foundation of genuine cost advantages or on temporary financial engineering. There is also a security dimension that deserves more attention than it is getting. The shift toward open-source models means that a growing share of AI inference is running on infrastructure that lacks the safety guardrails and monitoring that closed-source providers like Anthropic and OpenAI have invested heavily in. In the chaos of a crash, the data remains silent, but the vulnerabilities do not. When a developer routes their application traffic through an open-weight model, they are taking on the responsibility for alignment, safety, and compliance that the model provider would otherwise bear. Most developers are not equipped to handle this responsibility. They are choosing the cheapest option without fully understanding the trade-offs. Based on my experience auditing the Parity multisig vulnerability in 2017, I learned that the most dangerous failures are not the ones that are visible in the code. They are the ones that emerge from the interaction between the code and the assumptions of its users. The same principle applies here. The open-source models are not inherently less safe than closed-source models. But the ecosystem around them is less mature. There is no equivalent of OpenAI's safety team or Anthropic's responsible scaling policy for the open-weight ecosystem. The security burden has been shifted to the application developer, who is often the least equipped to bear it. The investment implications of this data are significant. The market has been valuing AI companies based on usage growth, but the Vercel data suggests that usage growth and value creation are decoupling. A company can serve 62% of the tokens and capture only 8.6% of the spending. Investors who are betting on token volume as a proxy for revenue potential are going to be disappointed. The real value is being created by companies that can command premium pricing for high-quality models, not by companies that are giving away inference at cost. This is a fundamental shift in the valuation logic of the AI sector. Let me be precise about what I am not saying. I am not saying that open-source models are a passing fad or that they will fade away. The opposite is true. Open-source models are here to stay, and their share of token volume will likely continue to grow. What I am saying is that the economic value of that volume is far lower than the raw numbers suggest. The open-source ecosystem is building the roads and bridges of the AI economy, but the toll booths are owned by the closed-source providers. This is not a bad outcome. It is a division of labor. But it has profound implications for who captures the economic surplus. The prediction that closed-source models will eventually account for 15-25% of token volume while capturing 60-90% of economic value is not a forecast of doom for open-source. It is a forecast of a mature market structure. The same pattern has played out in every technology stack, from operating systems to databases to cloud infrastructure. The open-source layer provides the foundation, and the proprietary layer captures the premium. The AI industry is not immune to this dynamic. What should developers and investors take away from this data? First, the choice between open-source and closed-source models is not a binary decision. It is a portfolio decision. Different tasks require different models, and the optimal strategy is to use a mix. Second, the cost advantage of open-source models is real, but it is not permanent. The pricing gap will narrow as the market matures and the subsidy dynamics change. Third, the security implications of the shift toward open-source models are underappreciated and will become more visible as the ecosystem grows. The code does not lie, but the auditor must dig. The data from Vercel is telling us something important about the structure of the AI market. The question is whether we are willing to listen. The next twelve months will be telling. If the open-source token share continues to grow while the spending share remains flat, it will confirm the commoditization thesis. If the spending share starts to catch up, it will suggest that open-source models are genuinely closing the quality gap. Either outcome has different implications for the competitive landscape. The data will tell us which one is happening. The only thing that is certain is that the current equilibrium is not stable. The market is still in flux, and the winners and losers have not yet been determined. In the chaos of a crash, the data remains silent, but the patterns are visible to those who know where to look.