The data is stark. Open-source models account for 62% of all token consumption on Vercel’s AI platform. Yet they generate only 8.6% of the total expenditure. Anthropic, with 30% of the tokens, commands 65.1% of the spend. The market is not paying for what it uses. It is paying for what it values.
This is not a story about democratization. It is a story about value extraction. The numbers reveal a structural fracture in the AI economy: usage volume does not equal economic value. Those who chase token share alone will find themselves in a low-margin desert. Those who capture the high-value tasks will own the profit pool.
I have seen this pattern before. In 2020, during DeFi Summer, I backtested over 500,000 historical block data points on Compound and Aave. The same divergence appeared: high transaction volume on low-yield pools, while the real value flowed to a few high-quality protocols. The market rewarded quality, not activity. The same rule applies here.
*
Context: The Data Set
Vercel’s AI platform tracks token consumption and expenditure across major model providers. The data set covers the period from October 2024 to December 2024. Three key players dominate: DeepSeek (open-source, China-based), Anthropic (closed-source, Claude models), and OpenAI (GPT-4, GPT-4o). Other providers like Google Gemini and Meta’s Llama also appear, but the top three account for over 90% of the volume.
The numbers are not perfectly representative. Vercel’s user base skews toward web developers and front-end applications. Code generation, text classification, and content drafting dominate the use cases. High-stakes enterprise workflows—like legal document analysis or financial modeling—are underrepresented. Still, the direction is clear: open-source is eating the low-end market, while closed-source captures the high-value niches.
*
Core: The On-Chain Evidence Chain
Let me break down the data methodically.
1. Token Share vs. Expenditure Share: A 15x Gap
Open-source models (DeepSeek, Llama 3, Qwen) consume 62% of all tokens. Their expenditure share is 8.6%. The ratio of usage to spend is approximately 7.2:1. For every $1 spent on open-source tokens, the equivalent closed-source tokens would cost $15. This is a price elasticity effect of the highest order.
But price elasticity creates a trap. Low-cost tokens encourage overuse. Developers throw more tokens at simple tasks because the marginal cost is near zero. This inflates the usage share without increasing the economic value. The 62% token share is a mirage—it reflects a behavioral shift toward cheap inputs, not a shift in value creation.
2. Anthropic: The High-Value Anomaly
Anthropic’s 30% token share generates 65.1% of expenditure. The unit token value of Anthropic is approximately 2.2 times the market average. Why? Because Anthropic’s Claude models handle complex reasoning, long-context tasks, and safety-critical applications. Developers pay a premium for reliability and context length. This is not a commodity market; it is a quality market.
3. DeepSeek’s Rise: Price or Performance?
DeepSeek surpassed Google Gemini in token consumption, becoming the second-largest provider by volume. This is a significant milestone. But the driving force is price, not performance. DeepSeek’s API costs are 1/10th of Google’s Gemini 1.5 Pro for equivalent outputs. The model is competent—good enough for code completion and text summarization—but it does not match Claude’s reasoning depth or GPT-4o’s multi-modal capabilities.
The implication: DeepSeek is winning the low-end market by undercutting on price. This is a volume game, not a value game. The company’s revenue per token is likely below $0.0001, while Anthropic’s exceeds $0.001. Scale compensates for low margins, but only if the cost structure is sustainable. Based on my experience auditing DeFi protocols, I have seen similar “volume for market share” strategies collapse when the subsidy runs out.
4. The Total Token Growth: 59% Quarter-over-Quarter
Vercel’s token consumption grew 59% quarter-over-quarter. This is explosive growth. Open-source models drove the majority of this increase. The low price of open-source tokens created new demand—tasks that were previously too expensive to automate are now economical. This is a classic Jevons paradox: as the cost of a resource falls, consumption increases, not decreases.
But the new demand is concentrated in low-value tasks. The incremental token consumption is mostly for simple classification, minor edits, and repetitive queries. The high-value tasks—like generating legal contracts or drafting investment memos—remain with closed-source models. The growth in volume is not matched by growth in economic value per token.
5. The Structural Gap: Usage vs. Value
The data tells a clear story: the AI market is bifurcating. Open-source models will dominate token volume, capturing 60-80% of all usage. But they will capture only 10-20% of the economic value. Closed-source models will capture 80-90% of the value with 20-40% of the volume. This is not a temporary distortion. It is the equilibrium of a market where quality commands a premium.
*
Contrarian: Correlation Is Not Causation
The natural conclusion is that open-source is winning. The data shows open-source has 62% token share. The narrative writes itself: “Open-source democratizes AI, topples the giants.” But this is a leap. The correlation between high token share and market dominance is false. High usage does not equal high revenue. High usage does not equal high margins. High usage does not equal sustainable competitive advantage.
Consider the parallel in DeFi. In 2021, SushiSwap had higher transaction volume than Uniswap for several months. But Uniswap captured more fee revenue because it attracted higher-value trades. Volume alone is vanity. Revenue is sanity. The same holds here: DeepSeek may have more tokens, but Anthropic has more dollars.
Another blind spot: the Vercel data excludes enterprise internal deployments. Many large corporations use closed-source models for internal workflows, running them on private infrastructure. These deployments generate zero token consumption on public APIs, but they represent significant value. The true expenditure share of closed-source models is likely higher than 65.1%.
Finally, the price of open-source models is not sustainable. DeepSeek’s pricing is below its marginal cost of inference. The company is subsidized by Chinese government grants and venture capital. When the subsidy ends, prices will rise. The current 62% token share is a subsidized share, not an organic one.
*
Takeaway: The Next Signal
The next move will come from the closed-source camp. Watch for Anthropic or OpenAI to lower the price of their high-end models, targeting the upper end of the low-value market. This will compress DeepSeek’s growth and force a revaluation of open-source economics.
Alternatively, watch for a consolidation wave among open-source providers. The current fragmentation—DeepSeek, Llama, Qwen, Mistral—is unsustainable. Only the provider with the best cost structure and model quality will survive. The rest will vanish.
Gravity always wins when leverage exceeds logic. The leverage here is the subsidized pricing of open-source models. The logic is that value flows to quality, not quantity. The data demands respect, not reverence. Do not mistake token volume for market power. The real market is hidden in the expenditure column.
Questions I will be asking: When will Anthropic cut its price? How long can DeepSeek sustain its subsidy? Which enterprise clients will switch from closed-source to open-source, and for which tasks? The answers will determine the next phase of the AI economy.