The Great Divergence: Vercel Data Reveals the Token-Value Split Rewiring AI's Economic Order

The numbers landed on my desk with the weight of a structural rupture. Vercel's latest telemetry, which tracks AI model usage across its serverless deployment network, shows open-weight models now command 62% of all processed tokens. Two months ago, that figure sat at 28.4%. The growth is not an anomaly; it's a signal. But the signal gets interesting when you cross-reference it with the expenditure data. Those 62% of tokens account for a mere 8.6% of spending. This divergence between raw usage and economic capture isn't just a market quirk. It's the blueprint for a two-tier industry. The architecture of this shift requires context. Vercel is a deployment platform that intermediates API traffic from AI applications to model providers. Its data captures real production workloads, not benchmark sandboxes. The platform's user base skews toward web developers building AI features into front-end products: code generation, content enrichment, and classification workflows. This is a particular slice of the broader AI market, but it's a deeply consequential one—it represents the mass-market, high-frequency layer of AI adoption. The ecosystem breakdown is where the numbers get interesting. Anthropic's Claude models manage to hold a 30% share of token volume. But critically, that 30% generates 65.1% of total platform spend. Conversely, the 62% open-weight share delivers 8.6% of the dollars. The math is stark. The unit economics of Claude tokens on Vercel are roughly 15 times higher than the average open-weight alternative. There is a gap in perceived value and utility here. This is not a story of simple substitution; it's a story of bifurcated roles. From a macro perspective, I see the blueprint for the AI industry's future value chain. The open-weight ecosystem is the commodity layer. It's the crude oil of the AI era—immense volume, pipeline-driven, and thin margins. The closed-weight frontier models are the specialty chemicals. They are the low-volume, high-margin products synthesized for specific, complex reactions. The total token volume on Vercel grew 59% quarter-over-quarter. That's the elasticity of the market being unlocked. The low price point of open models isn't just a substitute for their closed-source rivals; it's a catalyst for new demand. Developers are now building features that were previously uneconomical. They're applying AI to every mundane, repetitive data task, and they're doing it with cheap, open-weight models. This expansion of the total addressable market is why I believe the absolute volume of closed model tokens will also rise, even as their relative share contracts. The most significant data point in the report, however, is not the aggregate. It is the second-place finish. DeepSeek has surpassed Google's Gemini to become the second-largest provider by token volume on the Vercel platform. This is a rug pull on the prevailing narrative of Silicon Valley supremacy. The incident is a forensic clue. It tells us that a relative underdog, with a cost structure optimized for MoE architecture and a willingness to price near marginal cost, can out-execute an incumbent with a superior brand and decades of engineering talent. For developers, the choice is increasingly algorithmic. If the output quality is 95% as good and the price is 15% the cost, the decision becomes a no-brainer. It forces us to ask a harder question: what is the true alpha of the closed model? That alpha is the domain of the Contrarian thesis. The market is reading the data as a blanket win for the open ecosystem. I see a different structure. The market's "value capture" is still firmly in the hands of the closed providers. The token share is a lagging indicator of value, not a leading one. The true leading indicator is the unit of value extraction. Anthropic is not a loser in this transition; they are the winner of the premium segment. Their model is the tool of choice for complex reasoning, where a hallucination is not a nuisance but a liability. A financial audit or a complex contract analysis cannot afford a low-cost model with a 3% error rate. That zero-defect premium is what Claude is selling. The open-weight ecosystem, in turn, is a graveyard of zero-differentiation for the players that don't scale. This is the blind spot. The proliferation of open-weight models, and the ability of anyone to spin up a serverless deployment, commoditizes the entire stack. The unit economic value of a token, on average, is collapsing. The only defense is either scale with extreme cost efficiency, which is DeepSeek's game, or differentiation in quality, which is Anthropic's game. The middle ground—where I see OpenAI and Google stuck—is the most dangerous place to be. They are too expensive to compete on price against DeepSeek and too generic to justify the premium of Claude. Their models are being squeezed out of the middle. This transition has a direct implication for the public markets and venture capital. The valuation of model providers is now a function of their position on this new curve. A "Token-Volume" champion is not a revenue champion. DeepSeek's dominance in usage tells you nothing about their gross margin. If they are selling tokens below cost to buy market share, that's a subsidy strategy that requires infinite capital. It's a P/D (Price-to-Downloads) ratio that's unsustainable. The investment thesis must be built on the "value per token" metric, not the volume. I am building a framework now that looks at the "Net Token Value" (NTV) for a model provider. It's the total dollar spent by users, divided by the total tokens consumed. In the current environment, the only long-term, profitable position is the one with a high NTV. The forward-looking conclusion is unavoidable: this is a market of two speeds. The "token majority" is a low-margin, high-volume utility layer. The "spend majority" is a high-margin, low-volume intelligence layer. The future is not a war for the most tokens, but a war for the most critical tasks. The winner will be the one that owns the workflows that cannot fail. The question I am left with is a positioning one: Are you building infrastructure for the world to think, or are you building the one product that can think better than the world? The former will be huge; the latter will be indispensable. My capital allocation is clear. I am not interested in the infrastructure. I am interested in the indispensable.