The numbers do not reconcile. Sixty-two percent of all tokens consumed on Vercel's platform now flow through open-source models. Those same models generate 8.6 percent of the spending. Two months ago, that token share stood at 28.4 percent. The gap is not an anomaly. It is a structural statement about where value actually accumulates in the AI stack.
I have spent the better part of a decade tracing value through ledgers. Token flows are just another ledger. And this particular ledger tells a story that most market commentary refuses to confront: usage volume and economic value have decoupled. Permanently.
Let me establish the methodology before the evidence. Vercel's platform data captures developer traffic across AI-powered applications, primarily web and front-end deployments. This is not a perfect sample of the entire AI market. It skews toward code generation, content workflows, and lightweight inference tasks. But it is a real-time, production-environment dataset. Not a benchmark. Not a whitepaper. Actual developer behavior, measured at the point of consumption.
The dataset covers token consumption and associated expenditure across major model providers. The key metrics: total token volume grew 59 percent quarter-over-quarter. Open-source models now command 62 percent of that volume. DeepSeek, a Chinese open-source model family, surpassed Google to become the second-largest provider on the platform. Anthropic holds 30 percent of token volume but captures 65.1 percent of all spending.
The divergence is stark. Open-source models deliver the majority of raw inference work. They capture a fraction of the economic value. This is not a temporary pricing distortion. It is the emerging architecture of the AI industry.
The evidence chain runs deeper than the headline numbers.
DeepSeek's ascent is the first data point. Surpassing Google in token consumption on a major developer platform is not a trivial event. Google's Gemini models are competent. They are well-marketed. They are integrated into a massive ecosystem. And yet developers are routing more inference traffic through an open-source alternative. The reason is not mysterious. Performance-per-dollar has crossed a threshold where the quality gap no longer justifies the price premium for a significant class of tasks.
This aligns with what I observed during my DeFi audit work. When a cheaper alternative reaches 90 percent of the incumbent's capability, the market does not split evenly. It tips. The cheaper option captures the long tail of use cases almost immediately. The incumbent retains only the highest-value, most complex workloads. The same dynamic is playing out in AI inference.
The second data point is Anthropic's position. Thirty percent of tokens. Sixty-five percent of spending. The unit economics are roughly four times the market average. This is not inefficiency. It is pricing power derived from genuine differentiation. Anthropic's models are being used for tasks where failure is expensive. Complex reasoning. Multi-step analysis. High-stakes generation. In those contexts, the cost per token is irrelevant compared to the cost of a bad output.
I have seen this pattern before. In 2020, when I audited Aave's interest rate models, I found that the market consistently overpaid for perceived safety. Protocols with audited, battle-tested code commanded premium valuations despite offering nearly identical yields to unaudited competitors. The premium was not for the code. It was for the certainty. Anthropic is selling certainty. The market is buying it.
The third data point is the 59 percent growth in total token volume. This is the price elasticity effect in action. Open-source models have lowered the marginal cost of inference to near zero. Developers are now using tokens for tasks they would have previously deferred or handled with deterministic code. Text classification. Information extraction. Format conversion. These are not glamorous workloads. They are the plumbing of modern applications. And they are consuming the majority of inference traffic.
The contrarian angle: token volume is not a proxy for value creation.
This is where the narrative breaks down. The prevailing market story treats token consumption as a proxy for model quality and market share. The Vercel data suggests the opposite. Token volume is increasingly a proxy for commoditization. The more tokens a model family processes, the more likely it is being used for low-value, high-frequency tasks.
Consider the implications for valuation. DeepSeek has surpassed Google in usage volume. But its revenue is almost certainly a fraction of Google's AI business. Investors who conflate usage leadership with revenue leadership are making a category error. The same error that plagued the NFT market in 2021, when wash-traded volume created the illusion of organic demand. I documented 450 interconnected wallets executing circular trades to inflate Bored Ape floor prices. The volume was real. The value was manufactured.
Token volume can be manufactured too. Not through wash trading, but through pricing strategies that sacrifice margin for adoption. If DeepSeek is pricing below cost to capture market share, its token volume is a liability, not an asset. The question is not how many tokens flow through the model. The question is whether the provider can sustain the cost structure.
There is also a platform bias to consider. Vercel's developer base is concentrated in web application development. This overrepresents code generation and content tasks. It underrepresents enterprise workflows, where closed-source models maintain stronger footholds due to compliance requirements, data governance, and integration complexity. The 62 percent open-source share on Vercel may not translate to the broader enterprise market.
The structural prediction is straightforward.
Closed-source models will settle at 15 to 25 percent of token volume. They will capture 60 to 90 percent of economic value. This is not a forecast. It is an extrapolation of current data. The unit economics are already visible. Anthropic's 30/65 split demonstrates the ceiling. Open-source models' 62/8.6 split demonstrates the floor. The market is polarizing into two distinct tiers.
The first tier is the high-value inference layer. Complex reasoning, specialized domain expertise, tasks where accuracy is paramount. This tier will be dominated by a small number of closed-source providers with genuine differentiation. They will command premium pricing. They will capture the majority of AI revenue. Their valuations will be justified by margin, not volume.
The second tier is the commodity inference layer. High-volume, low-complexity tasks where cost is the primary consideration. This tier will be dominated by open-source models running on efficient infrastructure. Margins will be thin. Competition will be brutal. Value will accrue to the infrastructure layer, not the model providers.
This is the same pattern I observed in the DeFi lending market. Compound and Aave captured the high-value borrowing demand. Smaller protocols competed on price for the long tail. The market bifurcated. The same bifurcation is now occurring in AI.
The takeaway is not about which model wins. It is about where value concentrates.
For developers, the implication is clear: open-source models are now viable for production workloads. The cost advantage is too significant to ignore. But the choice of model should be driven by task complexity, not ideology. Use open-source for the long tail. Pay the premium for high-stakes reasoning.
For investors, the implication is more uncomfortable. The AI model market is not a winner-take-all landscape. It is a two-tier structure with divergent economics. Usage leadership does not equal revenue leadership. Token volume does not equal value capture. The metrics that matter are unit economics, retention, and the defensibility of the high-value tier.
For the industry as a whole, the 62/8.6 divergence is a healthy sign. It means AI is becoming accessible. It means the cost of building intelligent applications is collapsing. It means the application layer will continue to expand. But it also means the model layer is commoditizing faster than most observers expected.
I have been tracking on-chain data long enough to know that when usage and value diverge this sharply, the market eventually corrects. The correction will not be a collapse. It will be a repricing. Open-source models will be valued as infrastructure. Closed-source models will be valued as software. The distinction matters.
Logic is the only audit that never expires. The data is clear. The question is whether the market will read it correctly.
s silence.