Open-Source Models Now Drive 62% of Vercel's Token Traffic — But Only 8.6% of the Spend. Here's the Structural Break
Hook: The Token/Spend Divergence
Let's start with a number that doesn't reconcile. On Vercel's platform, open-source models now account for 62% of all token consumption. They account for just 8.6% of the associated spend. That is not a rounding error. It is not a temporary market anomaly. It is a structural fracture in how the AI industry assigns value.
It's tempting to read this as a victory lap for the open-source movement. It's not. It's a signal that the market has bifurcated into two distinct economies: the high-volume, low-value token flows handled by open-weight models, and the high-stakes, high-complexity workloads that still command premium pricing from closed-source giants. Anyone who conflates these two things is going to misprice the entire sector.
## Context: Vercel as a Data Witness Vercel's CEO published these numbers on August 22, 2024. The data is drawn from the platform's AI Gateway, which sits between developers and a marketplace of model providers. This is a unique vantage point. It's not a survey. It's not a whitepaper. It's raw traffic data from a tool used primarily by web developers, front-end engineers, and AI application builders. It captures real, routed API calls.
The headline figures are stark. In two months, open-source models went from 28.4% of token share to 62%. The second major shock: DeepSeek overtook Google to become the second-largest model provider on the platform, trailing only OpenAI. Meanwhile, Anthropic sits on a peculiar throne: it processes only 30% of tokens but commands 65.1% of the total dollar spend.
For context, the AI market is in its growth phase. The total token volume on Vercel is accelerating. The pie is growing. But the distribution of that pie is shifting underneath us.
Core: The Anatomy of a 14x Price Gap
The first thing I did with these numbers was calculate the effective price per token. The math is straightforward: If open-source models handle 62% of tokens for 8.6% of spend, and closed models handle 38% of tokens for 91.4% of spend, the ratio of unit economics is stark. The closed model token is, on average, about 14 to 17 times more expensive than an open-source token.
That gap is not an engineering reality. It's a strategic choice. Open-source models, particularly DeepSeek's V2/V3 series with their MoE and MLA attention architectures, have genuinely driven down marginal inference costs. But the price point isn't just cost-plus; it's penetration pricing. Open-source vendors are selling tokens at near cost to capture developer mindshare and ecosystem position. They're buying the flow.
However, the most dangerous conclusion to draw is that this is purely a "price war" story. If developers were only switching to save money, they'd switch back when quality slipped. The fact that the migration to open-source has been so rapid suggests something else: for the majority of tasks, the quality gap has closed.
Here's my specific inference from the data. The 62% token share is likely dominated by high-frequency, lower-complexity workloads: code completion, refactoring, unit test generation, documentation, batch summarization, embeddings, and basic classification. These are the everyday tasks of a developer. In this domain, models like DeepSeek-V2/V3 are 'good enough'. They cross the usability threshold.
The high-spend tail for Anthropic tells a different story. The 65.1% spend is concentrated in complex, high-stakes operations: Agentic workflows, long-horizon planning, complex codebase analysis, and sensitive enterprise integrations. This is where reliability and reasoning depth are non-negotiable. This is the "high-value density" tier.
This divergence validates my long-standing position on this industry: we are no longer competing on parameter count or intelligence. We are competing on value density. The question is not, "Can your model generate a JSON object?" It is, "Can your model generate a production-grade, enterprise-compliant, security-hardened codebase without hallucinating?"
The DeepSeek Milestone: More Than Price
Let's focus on DeepSeek specifically. Passing Google on token volume is a marker of a particular kind of victory. Google's research capabilities are not in question. But this data suggests that in the developer ecosystem, the API's pricing, tooling, and iteration cadence have left developers cold. DeepSeek has won on the developer experience and total cost of ownership.
I have to be cautious here. The Vercel sample is a specific cohort: web developers. This is not enterprise IT. This is the long tail of tech. But it's precisely that long tail that is setting the foundation for the next decade of AI applications.
Contrarian Angle: The Hidden Costs of the Open-Source "Win"
The contrarian take, and the one the bulls are missing, is that the 8.6% spend figure is a trap. It's an illusion of cost savings that ignores the full stack of ownership.
That number only covers API calls. It does not include the GPU cluster costs, the DevOps time, the engineering overhead, the security auditing, and the latency management required to self-host these models. If a company is running a local DeepSeek model, the "free" token is subsidized by a massive internal infrastructure bill. That's not accounted for here.
Furthermore, this data is likely distorted by a "long tail" effect. Many of those open-source tokens are probably being used in dev/test environments or batch processing, where the cost of a mistake is low. In production environments, where an error costs revenue, I suspect the ratio shifts significantly. The data may be masking the fact that high-value production traffic is still flowing to the closed-source premium providers.
I think we're also seeing the beginning of a "race to the bottom" in quality expectations. When developers get used to open models for their daily drivers, they lower their expectations. They accept a 90% success rate on a code refactor because it's free. This forms a habit. This is a significant threat to the incumbent closed-source players even in the middle market.
Takeaway: The Divergence Is The Future
The data points to a future where open-source models become the default API for routine workloads, and closed-source models become the specialized, premium tool for critical tasks. This isn't a zero-sum game. It's a stratification of the market.
As I look at this, I see a market that's getting more efficient. The "value capture" is shifting to the applications layer, and the "value creation" is shifting to the infrastructure layer. The unit economics of the model itself are becoming a commodity. This is a critical juncture for investors and developers. The question isn't whether open-source will kill closed-source; it's whether the closed-source vendors can evolve into high-value service providers before the open-source models cross the "good enough" threshold on the complex tasks.
Right now, they have a temporary moat. It's not a security moat. It's a complexity moat. And those are the easiest to erode. The architecture of trust, engineered for failure. The trust in high-cost APIs, engineered for a market that might not be willing to pay a 14x premium for much longer.