The Token Paradox: Why Open-Source AI Models Win Volume but Lose Value

The numbers are stark. According to Vercel’s CEO, open-source models now account for 62% of all tokens consumed on their platform—yet they command only 8.6% of total spending. Closed-source models, clinging to 38% of token volume, swallow 91.4% of the revenue. This is not a market finding; it is a philosophical rupture. In the chaos of the chain, find the signal. The signal here is clear: we are witnessing a decoupling of usage from value, a split that mirrors the great debates in crypto between decentralization and capture. Context: Vercel is the deployment platform for modern web applications, serving millions of developers. Its AI Gateway routes requests to various LLMs, making it a neutral bellwether for developer sentiment. The data—spanning thousands of apps—shows a tectonic shift. In just a few months, open-source models surged from 28.4% to 62% of token consumption. DeepSeek, a Chinese upstart, leapfrogged Google to become the second-largest model provider by volume. Meanwhile, Anthropic, with only 30% of tokens, captured 65.1% of spending. The market is bifurcating. Core: The numbers tell a story of two economies. The open-source economy is a “volume market” where developers use cheaper, “good enough” models for routine tasks: code completion, documentation, simple classification. The closed-source economy is a “value market” where enterprises pay premium for reliability, safety, and complex reasoning. This is not a bug; it is a feature of how AI markets mature. Truth is not mined; it is remembered. And what we are remembering is that value accrues to scarcity, not ubiquity. DeepSeek’s rise is not just about price—it’s about engineering efficiency. Its MoE architecture and MLA attention deliver GPT-4-class performance at a fraction of the cost. But cost does not equal value. The unit economics are inverted: open-source models cost 1/14th per token, but they are used for tasks that produce 1/14th the economic output. The ratio is consistent. Yet this hides a subtle truth. The 62% token share is not homogeneous. It is composed of millions of small, low-value calls—batch processing, embedding, simple classification. These are the “dust” of the AI economy, analogous to micro-transactions on a blockchain. The real value lies in the complex, high-stakes queries that demand the best models. Anthropic’s Claude is the “L2” of AI—offloading trivial work to cheaper models while handling the heavy lifting. But here’s the contrarian angle: this stratification is not stable. Open-source models are improving faster than closed-source. The “capability gap” is closing. If open-source models can handle 80% of complex tasks within two years, the value distribution will invert. The current 91.4% spending share becomes a historical artifact, not a future blueprint. Culture is the new consensus mechanism. The developer community is voting with its tokens. The shift to open-source is not just about cost—it’s about trust, transparency, and the belief that intelligence should be a commons, not a commodity. But this brings a blind spot. The same dynamic that drives open-source adoption also creates a new centralization risk: the infrastructure layer. Vercel, AWS, and cloud providers sit between the models and the users. They are the new “validators” of the AI economy. Their fee structures, policies, and data access shape the flow. We do not build walls; we build bridges for value. But who controls the bridge? Based on my experience auditing smart contracts and building decentralized education platforms, I see a parallel. The open-source token surge is like a layer-2 scaling solution—it increases throughput but introduces new trust assumptions. The real value is not in the model weights but in the coordination layer: the API gateways, the caching layers, the authentication systems. These are the “stakers” of the AI value chain. Anthropic’s high spending share is not just about model quality; it’s about the enterprise-grade service wrapper—security, compliance, SLAs. That is the moat. Takeaway: The future of AI is not a monopoly of one model; it is a federation of models, each optimized for a specific task. The value will flow to the orchestrators, not the generators. The question is not whether open-source will win—it will. The question is whether the infrastructure that supports it will remain open or become a new walled garden. Ideas have no gas fees, only gravity. The gravity of value is pulling toward the platforms that enable frictionless coordination. That is the signal in the chaos. The rest is noise.