The numbers do not align.
Vercel's CEO recently released platform data that should disturb every model vendor. Open-source models now command 62 percent of all token traffic on the platform. They account for only 8.6 percent of customer spending. The gap is not a rounding error. It is a structural revelation.
The protocol does not lie. The interface does. In this case, the interface is the pricing page. And the data confirms a truth that marketing departments have worked hard to obscure: we are in the midst of an industrial-scale migration of workloads to open-source models, but the revenue remains locked behind closed APIs. To understand the next phase of the AI and crypto infrastructure wars, we must first understand why this paradox exists.
Context
Vercel is not a random sample. It is the deployment layer for a massive segment of the modern web. The developers building on Vercel are shipping production applications. They are not playing with toy prompts. When the company's CEO shared that open-source token share had surged from 28 percent to 62 percent in a matter of months, the signal was not about hobbyists. This is a statement about where real engineering work happens.
The major providers in this data set are OpenAI, Anthropic, Google, and DeepSeek. The story is not uniform across them. OpenAI and Anthropic are still growing in absolute terms. The market is expanding. But the competitive structure has snapped into a new alignment. DeepSeek, a Chinese open-source model vendor, has overtaken Google to become the second-largest model provider on the platform. That is a seismic event. It is also one that most Western analysts have handled poorly, preferring to dismiss it as price-driven arbitrage. My audit experience tells me the truth is more nuanced.
Core: The Token-Value Disconnect
Let me state the obvious for the record: token volume is not a proxy for economic value. The current data proves this. The ratio is stark. Open-source models deliver the majority of tokens consumed but receive less than a tenth of the revenue. By my calculation, the unit price of open-source tokens is roughly one-fourteenth that of closed-source tokens. That is not a cost-of-goods differential. That is a pricing strategy. It is what we call in the industry a penetration play.
The DeepSeek Signal
DeepSeek's rise is not simply about being cheap. If price were the only factor, developers would migrate for a test and then return when the quality gap hurt them. The fact that token share has held and expanded tells me the actual capability of the models has crossed a usability threshold for routine tasks. Code completion, boilerplate generation, testing, and documentation. These are the tasks that make up the bulk of daily traffic. DeepSeek's architecture, with its mixture of experts and multi-head attention mechanisms, has cracked the cost-efficiency curve in a way that forces a reckoning.
I have been in this industry long enough to know that when a new player hits a cost-performance ratio this aggressively, it is not just an incremental step. It is a breakthrough in the engineering of the model itself. The MoE architecture allows them to activate only the necessary parameters for a given token. The attention optimization reduces the key-value cache overhead. These are not marketing claims. They are structural changes. They are the reason DeepSeek can undercut the market without bleeding capital on every inference.
The Value Enclave
Anthropic is the other side of this coin. Thirty percent of token consumption generates 65.1 percent of the spending. This is the value-density signal. Claude is being used for the hard, high-stakes work: complex code generation, sophisticated agentic workflows, deep analysis where a single error costs more than the entire token bill. The market is not irrational. It is segmenting.
This segmentation is the natural equilibrium of a mature technology market. The cost-sensitive, high-volume work moves to open-source. The complexity-critical, low-volume work stays with the premium vendors. The problem for the premium vendors is that the cost-sensitive work is a massive volume. And volume creates an ecosystem. As more applications are built on open-source foundations, the feedback loop accelerates. The open-source models improve faster, and the flywheel spins.
There is a parallel here to my own research in the crypto space. For years, the centralized players held the 'security' high ground. But as the code matured, the capability gap narrowed. The cost of the blockchain layer dropped. Suddenly the 'secure' option was no longer the only option. The same is happening with AI models. The quality gap for 80 percent of tasks has collapsed.
The Value Migration
The forecast is now clear. I have been hearing for months that the end state is a world where closed-source models hold 15-25 percent of tokens but 60-90 percent of economic value. The Vercel data is the first concrete evidence that this trajectory is real. But there is a condition. This future requires that the closed-source models maintain their lead in frontier capability. If open-source models close the reasoning gap within the next two years, the value distribution inverts. The high-value tasks will not stay still. They will follow the cost curve.
Contrarian: The Blind Spots in the Narrative
I am, by nature, a skeptic of the hype machine. Let me point out what the bullish open-source story misses.
First, the Vercel sample is not the enterprise. Vercel is the home of the modern web developer. It is the frontier of the frontend. But it is not the Fortune 500 IT department. The enterprise is where the real high-value contracts live. The data may be overstating the penetration of open-source into the mainstream core of the economy.
Second, the cost comparison is incomplete. The 8.6 percent spending on open-source is the API spend. It does not include the self-hosted GPU costs, the engineering labor to deploy and maintain the model, and the operational overhead. For many companies, the total cost of ownership of a self-hosted open-source model is not significantly less than a managed closed-source API. The 'cheap' label is an illusion if your organization does not have the DevOps maturity to manage the inference cluster.
Third, there is a market distortion. The 'token' is a proxy for workload, but it is not a proxy for profit. A developer who is happy to use open-source for a summary feature will still use a closed-source vendor for the core proprietary algorithm. The 62 percent figure is a testament to the long tail of low-complexity tasks. It does not measure the sophistication of the workload.
This is why I look at this data with a certain amount of cold pragmatism. The shift is real. But the narrative that 'OpenAI and Anthropic are doomed' is the same narrative we heard about Bitcoin during every bear market. The dominant players have a moat in the form of the highest-value use cases and the trust of the institutions. The moat is not in the code. It is in the memory and the compliance. And we in this industry know that the hardest thing to code is trust.
Takeaway
We are in the 'pipeline phase' of the AI economy. The open-source movement is winning the battle for volume and is losing the battle for value. The interesting question is how long that can persist. The value of a token is not stable. It is a function of the underlying model's capability and the context of the task. As open-source models advance, the unit cost of their intelligence will drop even further. The 'value war' will then be a 'context war.'
To own the chain is to own the history. In this case, the chain is the code, and the history is the data. The developer choosing the open-source model is choosing a different kind of history. They are choosing a history where the intelligence is cheap, and the value is derived from the application layer. In that world, the model becomes a commodity. And the winner will be the one who can own the interface, not just the underlying tech.
I have to wonder if the next disruptive project will be the one that treats the model as a shared utility. We build in the dark to light the public square. The lights are on now. The only question is who owns the power grid.
