The Vercel Data That Exposes the Open-Source Illusion
The numbers don't reconcile. Open-source models now drive 62% of all token consumption on Vercel's platform. They generate 8.6% of the spending. Two months ago, that token share sat at 28.4%. The growth curve is exponential. The revenue curve is flat. Something fundamental is broken in how we measure AI adoption.
This is not a story about open-source triumph. It is a story about value extraction. And the data proves that the market has already figured out which models matter where it counts.
The Context: What Vercel Actually Measures
Vercel is not a neutral observer. It is the deployment layer for hundreds of thousands of web applications. Its AI gateway routes inference requests from developers building real products. This is production traffic, not benchmark scores. Not demo videos. Not marketing announcements. Actual workloads hitting actual models.
The platform processes billions of tokens monthly. The data cuts across every major provider: OpenAI, Anthropic, Google, DeepSeek, Meta's Llama, Qwen, Mistral. The usage patterns reflect developer preferences in real-time. When a developer in Berlin ships a code completion feature, Vercel logs it. When a startup in São Paulo integrates a summarization endpoint, Vercel logs it. This is the closest thing to a global telemetry layer for AI consumption.
The core dataset reveals a structural decoupling between usage volume and economic value.
Token share measures volume. Spending share measures willingness to pay. The gap between these two metrics is the single most important signal in the AI industry right now.
The Core: An Evidence Chain That Destroys the Adoption Narrative
Let me walk through the data with the forensic rigor this deserves.
Finding One: The Open-Source Surge Is Real, But Misleading
Open-source models grew from 28.4% to 62% token share in sixty days. That is a 118% increase. DeepSeek, specifically, surpassed Google to become the second-largest model provider on the platform. This is not a rounding error. This is a tectonic shift in developer behavior.
But here is what the volume narrative obscures: those tokens are cheap. Absurdly cheap. The unit economics work out to roughly one-fifteenth the price per token of closed-source alternatives. Developers are not choosing open-source because it is better. They are choosing it because it is free.
Finding Two: Anthropic Is the Value King
Anthropic holds 30% of token share. It captures 65.1% of all spending. Let that sink in. A model provider with half the token volume of the open-source ecosystem commands seven and a half times the revenue. The implied price per token is more than double the market average.
This is not an anomaly. This is a business model. Anthropic has positioned its models for complex reasoning, agentic workflows, and high-stakes code generation. Tasks where a single error costs more than the entire inference bill. Developers pay for that reliability. They pay for the ceiling.
Finding Three: The Price Elasticity Effect
Total token consumption grew 59% month-over-month. This is not organic demand expansion. This is the price elasticity of AI inference. When DeepSeek and other open-source providers slash prices to near-zero, previously uneconomical tasks become viable. Developers start running sentiment analysis on every customer email. They build classification pipelines for data they would have ignored. They generate draft documentation for every function.
The 59% growth is a demand illusion created by subsidized supply. It is not a sign of healthy market expansion.
Finding Four: DeepSeek's "Victory" Over Google Is a Cost Story
DeepSeek surpassed Google in token consumption. The mainstream interpretation: Chinese open-source models have achieved technical parity with Western closed-source giants. The data-driven interpretation: DeepSeek's price point is so aggressive that developers route low-stakes traffic there without thinking.
I have audited token consumption patterns across multiple platforms. When a model is 95% cheaper, developers will use it for tasks they previously deemed not worth automating. This is not substitution. This is new, low-value demand creation. Google's Gemini models cost real money. DeepSeek costs pennies. Developers will always choose pennies for trivial workloads.
The victory lap is premature. DeepSeek won the volume war. It is losing the value war. And the value war is the only one that matters for sustainability.
The Contrarian Angle: Why the 62% Number Is a Trap
Here is where the mainstream analysis goes wrong. Everyone is celebrating the open-source share surge as evidence of democratization. They are missing the deeper structural reality.
Correlation between token volume and strategic importance is not causation. It is noise.
Let me break down what that 62% actually represents. Based on my analysis of task distributions across production workloads, the bulk of open-source token consumption sits in:
- Code completion snippets
- Text classification and extraction
- Summarization of internal documents
- Template-based generation
- Low-stakes chat interfaces
These are commodity workloads. They require pattern matching, not reasoning. They tolerate errors. They have no regulatory exposure. The cost of failure is measured in cents, not dollars.
The high-value workloads—complex debugging, architectural reasoning, financial analysis, legal document review, multi-step agentic planning—remain firmly in closed-source territory.
Anthropic's 65.1% spending share is not a lagging indicator. It is the leading indicator. It tells you where the actual economic value sits. The 62% token share tells you where developers are spending their spare change.
There is a second layer to this trap. The Vercel platform itself carries selection bias. Vercel's developer base skews toward web application development. Frontend engineers. Indie hackers. Startup MVPs. These users have different model requirements than enterprise teams running regulated workflows.
If this same data were collected from an enterprise inference platform—Azure OpenAI endpoints, AWS Bedrock enterprise accounts, Anthropic's direct enterprise API—the closed-source spending share would be even more extreme.
The 62% number is real. It is also unrepresentative of where AI's economic gravity actually sits.
The Structural Analysis: What This Means for the Industry
The Vercel data maps to a two-tier market structure that is becoming more defined by the quarter.
Tier One: The Commodity Layer
Open-source models dominate this tier. Volume is massive. Margins are razor-thin. The business model resembles cloud infrastructure more than software. DeepSeek, Llama, Qwen, Mistral—these are the AWS of AI. They provide raw compute intelligence at scale. They will never command premium pricing because their models are freely available. Any competitor can run the same weights.
The economics here are brutal. Token prices are racing toward zero. Providers are burning capital to acquire usage share. The unit economics do not work without either massive infrastructure optimization or ongoing subsidies.
Tier Two: The Value Layer
Anthropic currently owns this tier. OpenAI is fighting to maintain its position. The moat here is not model weights—it is the deployment infrastructure, the safety tooling, the enterprise integrations, the reliability guarantees. Companies pay for the system, not the weights.
The market has already voted. 65.1% of spending goes to a provider with 30% of tokens. The signal could not be clearer.
My projection: closed-source models will maintain 15-25% of token share while capturing 60-90% of all economic value for the foreseeable future.
This has profound implications for how we value AI companies. Token volume is the wrong metric. Revenue per token is the only metric that matters. Investors who chase the largest token counts are buying into the commodity layer. Investors who chase revenue per token are buying into the value layer.
The DeepSeek Question: Sustainable Model or Subsidy Play?
The most important unresolved question in this dataset is the sustainability of DeepSeek's strategy.
DeepSeek's token consumption is growing at a rate that suggests aggressive below-cost pricing. This is a classic market-entry strategy: capture share, build brand, then raise prices. The problem is that open-source models cannot raise prices without destroying their competitive advantage. If DeepSeek's weights are freely available, developers can simply self-host and eliminate the API cost entirely.
The only viable path for open-source model providers is to become infrastructure companies. The model is the loss leader. The surrounding services—hosting, fine-tuning, deployment tooling, enterprise support—are the revenue.
This is a fundamentally different business model from closed-source providers. It has different margin profiles. Different growth trajectories. Different valuation multiples.
Based on my experience auditing protocol economics in the DeFi space, I have seen this pattern before. Projects that capture massive usage share through subsidies often collapse when the subsidies end. The usage evaporates. The revenue never materializes. The valuation corrects violently.
The floor is a lie. Only the whale matters. And in this market, the whale is the revenue per token, not the token count.
The Google Problem: A Case Study in Misdirected Strategy
Google being surpassed by DeepSeek on this platform is a strategic failure worth examining.
Google has the best research capabilities in the industry. It invented the transformer architecture. It has unlimited compute. It has distribution through its cloud business. Yet its models are losing to a Chinese open-source provider on a developer platform.
The problem is not technical capability. It is pricing and positioning. Gemini models are priced for enterprise value. They are being used for commodity workloads where price sensitivity is absolute. Developers compare Gemini's price against DeepSeek's price and make the rational choice.
Google needs to decide which tier it wants to compete in. It cannot compete in the commodity layer with its current cost structure. It needs to differentiate on the high-value workloads—and it needs to convince developers that its models deliver measurable value for those workloads.
The data suggests Google has not made this transition successfully. Its token share is being squeezed from below by open-source pricing and from above by Anthropic's quality premium.
The Takeaway: What to Watch Next Week
The Vercel data is a snapshot. The trend lines are what matter.
Watch three signals over the coming weeks:
First, OpenAI's positioning. If OpenAI begins matching Anthropic's pricing on complex reasoning tasks, the value layer is consolidating. If OpenAI moves downmarket to compete with open-source pricing, it is abandoning the premium position. Either signal reshapes the competitive landscape.
Second, DeepSeek's next model release. If DeepSeek's next generation maintains or extends its capability gains, the open-source surge will continue. If the capability gains plateau, the token share will stabilize—and the revenue share will remain negligible.
Third, enterprise migration patterns. Watch whether enterprise inference traffic begins shifting to open-source models. That would be the real signal. The Vercel platform captures developer preference. Enterprise adoption is a different beast with different switching costs.
The open-source token share will keep growing. The spending share will not. That gap is not a bug in the market. It is the market telling you where value actually lives.
The 62% number will be quoted in countless articles as evidence of open-source dominance. Read it correctly: 62% of the volume, 8.6% of the value, and a widening chasm between adoption and economics.
Follow the revenue per token. Not the hype. The token count is a vanity metric. The spending share is the truth. And the truth is that Anthropic is building a moat that open-source models cannot cross.
Code doesn't lie. The data doesn't either. The only question is whether you are reading the right numbers.
