The numbers hit me like a cold block of data on a Monday morning. Last week, Vercel's CEO dropped a bombshell that should have shaken the AI world to its core. Open-source models now account for 62% of all tokens processed on the platform. Yet they generate just 8.6% of the spending. My first instinct as a data analyst was to check the methodology. My second instinct was to realize this is the most important metric of the current AI cycle.
For those who haven't been tracking this shift, Vercel sits at a unique vantage point in the AI ecosystem. It's the neutral ground where developers build and deploy. Its AI Gateway routes requests across every major model provider, from OpenAI and Anthropic to Google and DeepSeek. This is not a theoretical market analysis. It is the actual routing data from thousands of developers making daily decisions about which model deserves their workloads. When the CEO of Vercel publicly shares these numbers, we are not getting a vendor pitch. We are getting a literal blockchain of developer intent.
The data paints a picture that my 2017 ICO audit instincts find deeply familiar. We are watching a classic market in transition. The headline numbers tell one story, but the underlying transaction flows reveal the actual power dynamics. Over the past several months, I have watched this data shift in real-time. The token share for open-source models has exploded from 28.4% to 62%. This is not a gentle evolution. It is a supply shock, and the market is still trying to price it in.
The core of this story is the divergence between usage and value. When I first ran the numbers, I had to double-check my spreadsheet. Open source models process the majority of the tokens, but they capture a fraction of the revenue. The math is stark. The unit economics for open-source providers are roughly 1/14th of what the closed-source giants command. This is not a sustainable equilibrium, but it is the current reality of the market. We are seeing the birth of a two-tiered AI economy, and the line between them is drawn on the actual value of the task being performed.
The tale deepens when you look at the specific players. Anthropic is the most fascinating outlier in this dataset. They process 30% of the tokens but capture 65.1% of all spending. This is not a sign of inefficiency. This is the signature of a high-end provider. Claude's models are being used for the critical tasks, the complex code generation, the enterprise workflows, the Agentic loops. The developers are not paying a premium for the brand. They are paying for the reliability of the output, the safety of the architecture, and the consistency of the result. In the same period, DeepSeek has surpassed Google to become the second-largest model provider. This is a milestone that deserves its own analysis. The Chinese open-source model is not winning on cost alone. It is winning because it has crossed a capability threshold. Developers are not choosing it to save money. They are choosing it because it works.
Here is where my contrarian angle kicks in. This is correlation, not causation. The token numbers are seductive, but they hide a deeper structural reality. I have spent years auditing supply chains, and I can tell you that a 62% share of token volume is not the same as a 62% share of value. The open-source models are winning the volume war, but they are doing so by dominating the low-complexity, high-frequency tasks. The data pipeline, the embeddings, the simple code generation, the classification tasks, these are the foundation of the token volume. But the value lies in the complex reasoning, the agentic workflows, the multi-step processes where a single mistake costs a company thousands of dollars. The closed-source models are still the only ones trusted for those.
Based on my experience auditing the 2022 LUNA collapse, this is a liquidity map of the AI industry. The whale-like spending on Anthropic tells me where the institutional capital is parked. The migration of tokens to open-source models tells me where the retail development flow is heading. But the real signal is in the stability of the top tier. Anthropic is not losing its high-value market share. It is maintaining its premium by offering a reliability that developers trust for their most critical assets. In a bear market, this is exactly what I look for. The protocols with the highest value density survive, while the ones with just token volume bleed out. The question is whether this AI economy is heading toward a similar consolidation.
The biggest risk I see is the overvaluation of the open-source movement. The narrative is very seductive. Open-source is winning. The people's models are taking over. But the data tells a different story if you read the gas rather than the hype. The 62% token share is being subsidized by a cost structure that is not sustainable. The actual Total Cost of Ownership includes the GPU infrastructure, the engineering hours, the maintenance, and the integration. When you factor those in, the price advantage of the open-source models shrinks dramatically. We are seeing a classic penetration pricing strategy. DeepSeek is buying market share with low token prices. It is building a user base and a data flywheel that will be very difficult to dislodge. But the monetization is not there yet. The revenue per token is a fraction of what the closed-source models generate.
The institutional-grade infrastructure is still the closed-source domain. When I look at the data for the Enterprise API calls, the usage is focused on the highest-value tasks. This is where the loyalty lies. This is the core of the value proposition. The open-source models are improving their capabilities, but they are still trailing in the long-context, complex-reasoning, and tool-calling scenarios. The gap is narrowing, but it is not closed. The cost of switching, the cost of retraining, and the cost of failure are high. The developers who are putting their jobs on the line are staying with the reliable players.
The fundamental takeaway from this data is that we are not watching a replacement. We are watching a specialization. The AI model market is not becoming a winner-take-all scenario. It is becoming a two-tiered system. The open-source models are the base layer, the cheap compute that handles the bulk of the work. The closed-source models are the premium layer, the heavy lifters, and the critical reasoning engine. This is a stable equilibrium, not a temporary disruption. The developers will continue to use both, and the smart builders will route their workloads to the most efficient model for the task.
The next signal to watch is the pricing floor. As the open-source models continue to mature, they will push the price floor lower. This will force the closed-source providers to prove their value density. They will need to demonstrate that their outputs are worth the 10x or 15x premium. If they can not, then the token share will begin to erode the revenue share. The value migration will only happen if the quality gap narrows. Until then, I will be watching the liquidity, the flow, and the value density of each model. It is the data that will tell us where the real value is being created.
Follow the gas, not the hype.
