We didn't see this coming. Vercel’s latest data drop on AI model usage reveals a split so stark it redefines the entire market. Open-source models now consume 62% of all tokens on the platform. Yet they generate a mere 8.6% of total spending. The numbers are a wake-up call for anyone betting on a simple “open-source wins” narrative.
Context: Why Vercel matters Vercel is the backbone of modern web applications. Thousands of AI-powered apps—from code assistants to content generators—run on its infrastructure. Its token flow mirrors real developer behavior, not just hype. The data covers a two-month window ending mid-2025, and the trend is undeniable: open-source token share doubled from 28.4% to 62%. DeepSeek, a Chinese open-source model, blew past Google’s Gemini to become the second-largest provider by token volume. Meanwhile, Anthropic—the closed-source champion—commands only 30% of tokens but captures a staggering 65.1% of spending.
This is not a story of one side winning. It’s a story of two markets emerging within the same pipeline.
Core: The token-to-revenue ratio is the real signal Let’s break down the core signal. The token-to-revenue ratio is the key metric. For open-source, each token earns roughly 1/15 of what Anthropic’s token earns. That’s not a bug—it’s a feature. Open-source models are optimized for low-cost, high-frequency tasks: text classification, simple code completion, information extraction. They are the workhorses of the AI economy. Closed-source models like Claude and GPT-4 are the specialists—handling complex reasoning, creative writing, and high-stakes decisions.
I’ve seen this pattern before. During the 2022 DeFi summer, I audited a staking contract that had a reentrancy vulnerability—missed by two major audit firms. The protocol had massive TVL but thin fee margins. The same dynamic is playing out in AI. Volume is not value. Token count is not revenue.
The data also reveals a 59% quarter-over-quarter surge in total tokens across the platform. That’s the price elasticity effect: lower costs unlock new use cases. Developers are using more tokens because they’re cheap. Open-source is the engine of that expansion. But the economic value remains concentrated in the premium tier.
DeepSeek’s rise is the most explosive signal. Surpassing Google in token consumption is a milestone. But the question is: is it sustainable? Based on my experience reverse-engineering early ZK-rollup whitepapers in 2021, I know that Chinese AI labs often price aggressively, sometimes below cost, to capture market share. DeepSeek’s unit economics are likely subsidized. The long-term viability of that strategy is uncertain. Meanwhile, Google’s stagnation suggests their model isn’t resonating with developers—a dangerous position for a company that needs to monetize AI.
Contrarian: The open-source 'threat' is overblown The contrarian angle is uncomfortable but necessary. The narrative that open-source is “democratizing AI” is only half true. Yes, it lowers barriers. But it also creates a value hierarchy. The high-value tasks remain locked inside closed-source APIs. Anthropic’s 65.1% spending share proves that. Regulation didn’t anticipate this bifurcation. Current AI governance frameworks treat all models equally, but the economic reality is that open-source and closed-source serve fundamentally different roles. This will have implications for antitrust, data privacy, and export controls.
We didn’t consider that the “open-source revolution” might actually reinforce the dominance of premium closed-source models. As more developers use cheap open-source models for routine tasks, they become more dependent on Anthropic and OpenAI for the hard problems. It’s a symbiotic relationship, not a disruptive one.
We didn’t expect the data to be this lopsided. But the real blind spot is the assumption that token share equals market power. It doesn’t. The 62% open-source token share is a volume game with razor-thin margins. The 8.6% spending share is a cold reality check. The economic value is in the 30% of tokens that generate 65% of revenue.
Takeaway: The next watch The takeaway for investors and builders is clear: focus on value capture, not token volume. The next watch is whether open-source providers can climb the value chain. Can DeepSeek or Llama 3 match Claude on complex reasoning? If not, the 62% token share will remain a volume game with thin margins. And the real money—the 65.1%—will stay with the closed-source incumbents.
In this sideways market, positioning matters. The Vercel data is a compass. Point it toward the value layer, not the noise. The next six months will tell us if open-source can break the ceiling or if it’s destined to be the cheap labor of the AI economy.
