The Token Share Paradox: Why Open-Source AI's 62% Usage Maskers a 91.4% Spending Reality

The logs show a sharp divergence. Open-source AI models now process 62% of all tokens on Vercel's platform, up from 28.4% in January 2024. Yet they account for only 8.6% of spending. The code did not lie; the humans misread the data. Context: Vercel is not a typical AI model provider. It is a deployment platform for web applications, but its AI Gateway routes API calls to multiple providers. Last week, CEO Guillermo Rauch released raw usage metrics. The data cuts through narrative noise. It shows where developers actually allocate their compute and their dollars. Core: The headline number is dramatic: open-source token share doubled in seven months. Dig deeper and the story becomes more nuanced. DeepSeek, the Chinese open-source model, surpassed Google to become the second-largest model supplier by token volume. Anthropic, with only 30% of tokens, captured 65.1% of spending. OpenAI and Anthropic together still control 91.4% of total expenditure. This is not a simple shift to open-source. It is a bifurcation of the market. Open-source models dominate routine tasks: code completion, simple refactoring, documentation generation, batch classification. These tasks are volume-heavy but value-light. Closed-source models own the high-value work: complex code generation, multi-turn reasoning, agentic workflows, enterprise-grade compliance. The average token spend per open-source call is about 1/14th that of a closed-source call. Transition is not an event, but a data stream. The data stream shows that developers are not abandoning closed-source; they are using each category for what it does best. DeepSeek's rise is real but confined to cost-sensitive, high-volume workloads. Its MoE architecture and MLA attention deliver inference at commodity prices. But when the task requires precision, reliability, or safety, the wallet opens for Anthropic. Contrarian angle: The 62% token share is a mirage if you only look at provider revenue. The true cost of open-source includes self-hosting, GPU rental, and engineering overhead. Vercel's data only captures API spending. A developer running a local Llama 3 instance on a rented A100 incurs costs invisible to this metric. The total cost of ownership for open-source may be higher than the 8.6% suggests. Moreover, the token share surge is partly driven by free-tier usage and testing workloads. Production-grade, high-stakes inference still flows to closed-source. Based on my experience auditing on-chain data, I see a parallel: TVL can be inflated by small, active wallets, just as token volume can be inflated by cheap, repetitive calls. The real signal is value density—spend per token. Anthropic's value density is 2.17x that of OpenAI and 30x that of DeepSeek. That is the metric that matters for long-term business models. Takeaway: The AI model market is not consolidating around open-source. It is segmenting into two tiers: commodity and premium. Open-source wins the volume game; closed-source wins the value game. The next question is whether DeepSeek and its peers can climb the value stack. If they bridge the gap on complex reasoning, the spending distribution will shift. If not, the 62% token share will remain a hollow victory. The data is clear. The code did not lie; the humans misread the data.