The data indicates a bifurcation that most market commentary has missed. Vercel's platform telemetry for August 2024 shows open-source models now command 62% of all AI tokens processed. Their share of the spending, however, is a mere 8.6%. This is not a rounding error. It is a structural revelation about the economic reality of the AI stack.
For context, Vercel is not a neutral observer. The company serves as the deployment layer for a significant portion of the modern web. Its AI Gateway routes requests across a spectrum of models. The shift in the platform's token distribution is therefore a direct proxy for developer behavior in the production environment. The previous quarter had this split inverted, with proprietary models consuming the majority of compute. The migration of actual workloads has been swift.
The core analysis centers on the divergence between volume and value. The 62% token share for open-source models, led by DeepSeek's V2/V3 series, indicates a crossing of the "usability threshold" for mid-tier tasks. Code completion, refactoring, and document generation do not require frontier intelligence. In the absence of a quality penalty, the cost advantage becomes a binary switch. Consequently, the developer defaults to the cheaper option. Based on my audit experience, this is the point where cost optimization becomes the primary driver of architecture.
However, the token share is a misleading metric for revenue. The proprietary models retain 38% of tokens but command 91.4% of spending. Anthropic specifically processes 30% of the tokens yet represents 65.1% of the total expenditure. This discrepancy translates to a unit price ratio of roughly 1:14 between open and closed models. This is not a reflection of marginal compute costs. It is a market signal that Claude and GPT-4o are reserved for specific, high-value functions—complex code synthesis, agentic workflows, and enterprise-critical analysis. The cost per token is irrelevant when the cost of failure is high.
The competitive dynamics are often misread. The fact that DeepSeek surpassed Google to become the second-largest provider is significant, but for a specific reason. Google's research capabilities are not in question. Their developer ecosystem and API pricing are. The data implies that in the field, a model's market share is determined by the speed of inference and price, not by the number of academic papers published. DeepSeek's architecture (MoE with MLA attention) has created a true cost advantage. This is not a subsidy; it is a software engineering optimization.
This leads to a potential long-term inefficiency in the market. If open models control 62% of volume but only 8.6% of revenue, the incentive to improve them is driven by ecosystem growth, not direct monetization. This is a "bug" in the current economic cycle. The unit economics are heavily weighted toward the closed model vendors, giving them the capital to train the next generation of models. This could further widen the gap on the front lines of capability.
The contrarian view, however, requires a note of caution. The open-source proponents are right about the inflection point. The 62% share is not just about price. It is about adequacy. For most internal tools, the open-source output is statistically indistinguishable from the closed-source output. This is a fact that the closed-source vendors have not yet reconciled with. They have retreated to the high ground, but they have surrendered the plateau.
The takeaway
If this trend continues, the ecosystem will stabilize into a two-tier system. The commodity tier (open weights) will absorb the long tail of traffic, and the premium tier will process the high-stakes, high-value tasks. The problem arises if the premium tier fails to justify its cost. If the open-weight models continue to improve, the 65.1% spending share of Anthropic could be cannibalized by the open-source models. The real competition is not between the tokens and the costs; it is between the capabilities and the pricing. The foundation will determine the ceiling. In the absence of data, opinion is just noise. The data here says the foundation is shifting. I have to verify the cost per unit of intelligence, not the cost per token. The market has not priced this shift correctly yet. It is still looking at the top of the chart. It should be looking at the bottom line.
