The numbers do not reconcile. August 22, 2024. Vercel's platform data shows open-source models consuming 62% of all tokens. Their share of spending: 8.6%. The ledger does not lie, only the narrative does.
That 53.4-point spread is the largest structural signal in AI infrastructure since GPT-4 shipped. It deserves a forensic teardown.
Context: What the Platform Actually Observes
Vercel is not a model vendor. It is a deployment layer. Its AI Gateway routes developer traffic across multiple model providers, capturing real-time usage patterns without the bias of a single vendor's marketing. When Vercel's CEO publishes this data, it is not a sales pitch โ it is a diagnostic report from the middle of the stack.
The numbers: open-source tokens went from 28.4% to 62% in roughly two months. DeepSeek surpassed Google to become the second-largest model provider on the platform. Anthropic holds 30% of tokens and 65.1% of spend. OpenAI's absolute volume grows, but its relative share erodes.
The trend line is unambiguous. But the interpretation is where most analysis breaks.
Core: The Structural Disassembly
Let me break this down the way I broke down Terra's mint/burn mechanism in 2022. The same forensic approach applies โ identify the incentive structure, trace the value flow, expose the design flaw.
The Price Ratio Is Not Engineering โ It Is Strategy
Compute the unit economics. Open models: 62% of tokens for 8.6% of spend. Closed models: 38% of tokens for 91.4% of spend. The ratio of per-token prices: (8.6/62) / (91.4/38) โ 0.057. Open-source models cost approximately 1/14th the price per token of closed models.
That is not a technology gap. That is a pricing decision.
DeepSeek's architecture โ MoE with MLA attention โ achieves genuine inference efficiency gains. I have audited enough tokenomics structures to know when cost differences reflect engineering reality. The 14x gap is not engineering reality. It is penetration pricing. A deliberate strategy to acquire developer mindshare and ecosystem position before monetizing.
Open-source vendors are buying the market. The question is whether they can hold it.
DeepSeek's Ascent Is Not Just Price
Two months. That is the window in which DeepSeek overtook Google on Vercel. Price alone cannot explain that velocity. Developers do not migrate production workloads for a discount unless quality is acceptable. The data says DeepSeek crossed the usability threshold for a substantial class of tasks.
The harder question: which tasks? Vercel's platform skews toward web development, frontend infrastructure, and API orchestration. These are mid-complexity workloads. Code completion. Refactoring. Documentation generation. Test case writing. Integration glue.
On these tasks, open-source models have reached parity with closed models. The 62% token share is concentrated in this long tail of low-cost, high-frequency, individually-low-value calls. This is the hidden structure behind the data.
Anthropic's Premium Is the Counter-Signal
Anthropic's position is the most revealing. 30% of tokens. 65.1% of spending. Developers pay 14x more per token for Claude than for DeepSeek. That is a capability premium, not a brand premium.
Complex code generation. Long-document analysis. Agentic workflows. Enterprise-grade reliability. These are the workloads that still require frontier intelligence. Anthropic has converted capability advantage into commercial advantage โ the 'safety-first' positioning is not marketing, it is a license to charge premium rates.
OpenAI's volume grows, but its share decays. The market is not rejecting OpenAI. It is segmenting. The structure is clear: closed models own the high-value tier. Open models own the volume tier. The question is whether this stratification survives.
The Long-Tail Distortion
Consider what 62% of tokens actually looks like. A single embedding call is 100 tokens. A batch data-processing job is millions. The long tail of open-source usage is probably dominated by cheap, parallelizable, batch workloads. The asymmetry in cost explains the asymmetry in share.
If Vercel published task-type segmentation, my hypothesis is that open-source models would show concentration in classification, extraction, and simple generation. Closed models would show concentration in multi-step reasoning, tool orchestration, and high-stakes code generation. The data is consistent with this hypothesis.
But here is what I found when I audited AI-agent payment protocols in 2025: models that win on cost also win on latency and availability. The volume does not lie about the quality โ it tells you what developers trust for which jobs.
The Unanswered Question
Vercel's sample is biased. Web developers. Frontend engineers. The platform's user base skews toward a specific segment of the AI economy. It does not capture enterprise data pipelines, financial analysis, or regulated industries. The open-source share may be higher on Vercel than in the overall market.
However, the directional signal is real. Open-source models have crossed the capability threshold for mid-complexity tasks. The market has voted with its tokens. This is a structural shift, not a temporary anomaly.
Contrarian: What the Bulls Got Right
The open-source narrative is right in a way that undermines my natural skepticism. The 62% token share does not represent developer charity or ideology. It represents a rational economic choice. Developers are not paying for open-source models because they believe in open source. They are paying because the quality is good enough for the task.
That is the strongest signal in the dataset. When engineers migrate production workloads to lower-cost models, they are not engaging in activism. They are optimizing for unit economics. The fact that they can do this without sacrificing core functionality says more about the open-source models' capabilities than any benchmark.
But the bulls miss the second-order consequence. The 14x price ratio is not sustainable. As open-source vendors capture developer mindshare, they will raise prices. The current ratio reflects market penetration, not steady-state equilibrium. When DeepSeek and other open vendors reach critical mass, their pricing will converge toward cost-plus. The 62% token share will compress.
Emotion is a variable I exclude from the equation. The hype cycle around open-source models is real. The volume is real. But the economics are a one-time reset, not a permanent repricing.
Takeaway: The Value Density Shift
Structure outlives sentiment; code outlives hype. The industry is not transitioning from closed to open. It is transitioning from model-centric to value-centric competition.
The token share wars are a distraction. What matters is value density โ the economic value created per token consumed. Open-source models will dominate the low-value, high-volume tier. Closed models will retain the high-value, low-volume tier. The AI industry will replicate the same tiered structure as every other software market.
On the same basis, the infrastructure layer โ inference optimization, routing, serving โ becomes the new battleground. Vercel's data proves it. The platform that routes tokens efficiently is the platform that captures the value of the shift.
The ledger does not lie, only the narrative does. And the ledger says: value, not volume, is the only metric that survives. Panic is just poor data processing in real-time. Compute the value per token. Then you will know the winner.