The numbers landed on my terminal at 09:47 Istanbul time. Two months. That is all it took for open-source models to flip the Vercel platform from a 28.4% token share to 62%. Not a gradual drift. A vertical spike. And buried in the same dataset, the detail that matters more: those open-source tokens generated only 8.6% of total spending on the platform. Meanwhile, Anthropic—with just 30% of token volume—captured 65.1% of every dollar spent.
Let that sink in. The market is not shifting from closed to open. It is bifurcating into two distinct economies that happen to share the same API endpoints.
I have been tracking AI model consumption patterns since the 2017 ICO blitz taught me one lesson: volume without value is just noise with a timestamp. This Vercel dataset is the cleanest real-world signal we have on developer behavior—not benchmark scores, not marketing decks, but actual production traffic. And it is telling us something uncomfortable for anyone betting on a simple open-source victory narrative.
Context: Why Vercel Data Matters
Vercel is not a random sample. It is the deployment layer for a massive slice of the modern web application stack. Developers building on Vercel are shipping real products—frontend applications, API routes, serverless functions, AI-powered features. When those developers integrate AI models, they are making economic decisions with their own money or their employer's money. This is not a research lab running benchmarks. This is production traffic with a budget attached.
The platform's token volume grew 59% month-over-month. That is the demand-side story: AI features are becoming cheaper to integrate, so developers are integrating more of them. The price elasticity effect is real. When the marginal cost of a token drops by an order of magnitude, developers find new tasks to throw at models—tasks that were previously uneconomical. Code completion, text classification, information extraction, content generation at scale. These are the high-frequency, low-complexity workloads that now run on open-source models because the cost structure finally makes sense.
But here is the critical context most analysts are skipping: Vercel's developer base skews toward web application and frontend development. That means the platform over-represents code generation, UI content, and lightweight AI features. It under-represents the enterprise-grade, complex reasoning workloads that run inside large organizations—the ones that justify premium pricing for models like Claude and GPT-4-class systems. The Vercel data is a directional signal, not a market census. I have made this mistake before, back in 2020 when I was auditing DeFi yield farms and thought Curve's early pool data told the whole story. It did not. Platform bias is real, and it distorts every conclusion you draw from a single data source.
Core: The Volume-Value Divergence
Let me walk through the actual numbers, because the divergence is the story.
Open-source models: 62% of token volume, 8.6% of spending. That is a unit economics ratio of roughly 1:15. For every dollar spent on open-source tokens, closed-source models generate fifteen times the revenue per token. Anthropic alone: 30% of tokens, 65.1% of spending. The unit price of an Anthropic token is more than double the market average. And developers are paying it. Willingly. At scale.
This is not a story about open-source models being "good enough." It is a story about two different product categories serving two different job functions.
Open-source models are winning the volume game because they are the right tool for high-frequency, cost-sensitive, quality-tolerant workloads. If you are building a feature that classifies support tickets or generates product descriptions, you do not need Claude's reasoning depth. You need speed, low cost, and adequate quality. DeepSeek and its open-source peers deliver exactly that. The 59% token growth on Vercel is largely open-source-driven incremental demand—tasks that did not exist before because they were too expensive to automate.
Closed-source models, particularly Anthropic, are winning the value game because they handle the workloads where failure is expensive. Complex reasoning, multi-step analysis, code generation for critical systems, anything where a wrong answer costs more than a premium token price. In those scenarios, the 15x price premium is insurance. And enterprises are buying that insurance.
Now, the DeepSeek data point. DeepSeek surpassed Google to become the second-largest model provider on Vercel by token volume. That is a headline. But what does it actually mean? Based on my audit experience—I have spent the last two years dissecting model provider economics across multiple platforms—this is a price-driven adoption story, not a capability-driven one. DeepSeek's aggressive pricing has created a new market segment: developers who would not have used any AI model at previous price points are now building features around DeepSeek's API. The token volume is real. The economic value capture is not.
Google's position is more interesting. Gemini's token volume falling behind DeepSeek on a developer platform suggests a distribution problem, not a capability problem. Google has the research muscle. But its developer-facing API strategy has been fragmented, its pricing has not been aggressive enough for the long-tail market, and its models have not established a clear identity in the developer community. On Vercel, developers vote with their API keys. They are voting for DeepSeek's price and Anthropic's quality. Google is stuck in the middle—not cheap enough to win the volume game, not differentiated enough to win the value game.
OpenAI occupies a similar middle ground, though with a stronger brand and a more established ecosystem. The data suggests OpenAI's token volume is growing, but not at the rate of open-source models, and its unit pricing sits below Anthropic's. That is a squeeze from both directions. The question is whether OpenAI can maintain its position as the default choice for developers who want a balance of quality and accessibility, or whether it gets caught in the same pincer movement that is pressuring Google.
The Unit Economics Nobody Is Modeling
Here is the part that keeps me up at night. The open-source token share is growing because the price is unsustainably low. DeepSeek and other open-source providers are likely operating at or below cost on a per-token basis. This is a classic land-grab strategy: capture market share through subsidies, then figure out monetization later. It worked for Uber. It worked for Amazon. It is not clear it will work for AI model providers, because the cost structure is different.
In ride-sharing, the marginal cost of a ride approaches zero once the fleet is deployed. In AI inference, every token consumes compute. Electricity. GPU cycles. The marginal cost is real and it scales with usage. A provider running at a loss on every token is burning capital with every successful API call. The more popular the model, the faster the cash burn. This is the opposite of a network effect. It is a network drain.
I have seen this movie before. In 2020, I modeled the token emission rates of early Curve Finance pools and predicted the inevitable dump three weeks before it happened. The mechanics were the same: unsustainable incentives attracting volume that would vanish the moment the subsidies stopped. The Vercel data has the same signature. The 62% open-source token share is a subsidized number. It reflects price, not preference. If DeepSeek raises prices to sustainable levels, that share will contract. The question is how much of it survives the transition.
Contrarian: The Open-Source Victory Is a Mirage
Here is the counter-intuitive angle that the mainstream coverage is missing. The open-source token surge is not a sign that open-source models are winning. It is a sign that the market is stratifying into two layers with fundamentally different economics. And the layer that captures the economic value is the one with the smaller token share.
The prediction embedded in this data: closed-source models will eventually account for 15-25% of token volume but capture 60-90% of the economic value in the AI model market. That is not a defeat for closed-source. That is a consolidation of pricing power. The volume leaders will be the commodity providers. The value leaders will be the premium providers. And the gap between them will widen, not narrow.
This has direct implications for how we value AI companies. The market has been rewarding token volume growth as a proxy for adoption. That is the wrong metric. What matters is revenue per token, or more precisely, the defensibility of the premium price. Anthropic's 65.1% spending share on 30% token volume is the valuation anchor. DeepSeek's 62% token share on 8.6% spending is a cautionary tale. Investors who conflate usage with value are going to get burned.
There is also a platform bias problem that nobody is addressing. Vercel's developer base is not representative of the broader AI market. It is skewed toward web application development, which means it over-represents the exact workloads where open-source models are strongest: content generation, code completion, lightweight classification. It under-represents the enterprise workloads where closed-source models dominate: complex reasoning, data analysis, mission-critical automation. The real market split is probably less extreme than the Vercel data suggests. But the direction is clear.
The Infrastructure Angle
From an infrastructure perspective, this bifurcation is actually healthy. The AI stack is maturing into a layered architecture: commodity models for high-volume tasks, premium models for high-value tasks, and a middleware layer that routes requests based on cost and quality requirements. This is exactly how every other technology market has evolved. Compute, storage, and networking all went through the same commoditization cycle. The difference is that AI models are still in the early stages of that cycle, and the pricing dynamics are still being discovered.
The cloud providers are the quiet winners in this transition. Whether developers use open-source or closed-source models, they need compute. The 59% token growth translates directly into GPU consumption. The cloud providers are selling shovels in a gold rush where both sides of the mining operation are buying equipment. That is the infrastructure thesis that matters.
Takeaway: What to Watch Next
Three signals will tell us whether this bifurcation is durable or transitional. First, watch DeepSeek's pricing. If they raise prices and the token share holds, the adoption is real. If the share collapses, it was subsidy-driven. Second, watch Anthropic's token volume. If it starts growing while maintaining the spending share, the premium model is winning the value game outright. Third, watch Google's response. A serious Gemini push on developer experience and pricing would reshape the competitive landscape overnight.
The market is not choosing between open and closed. It is choosing between volume and value. And right now, the value is concentrated in the hands of a few premium providers while the volume flows to the commodity layer. That is not a revolution. That is a market finding its equilibrium. The question is whether the equilibrium holds when the subsidies run out. Static is a choice. The data is already moving.