The numbers do not reconcile. That is where I always start.
Vercel's latest model usage data dropped last week, and the signal buried inside it is not what the headlines claim. Open-source models now consume 62% of all tokens processed on the platform, up from 28.4% just two months prior. DeepSeek has surpassed Google to become the second-largest model provider by token volume. The open-source camp has crossed a threshold that many in the industry assumed would take years.
But here is the anomaly that nobody is talking about: those 62% of tokens account for only 8.6% of total customer spend on the platform. Anthropic, with a mere 30% token share, captures 65.1% of the money.
That is not a market. That is a structural bifurcation disguised as a migration trend.
I have spent fourteen years auditing smart contract systems, and the pattern here is not new. It is the same mispricing of risk and value that we saw in the DeFi summer of 2020, when total value locked became the vanity metric and actual yield generation was ignored until the music stopped. Token volume is the total value locked of the AI industry — a surface-level metric that conceals where the real economic gravity lives.
Liquidity is just trust with a price tag. And in this market, trust is being priced in two entirely different currencies.
The Numbers That Do Not Add Up
The Vercel dataset is small, developer-centric, and possibly biased. But for what it captures — real production traffic from working applications — it is far more honest than benchmark leaderboards. When you deploy a model in production, you are betting real uptime, real latency budgets, and real money on the result.
Here is what the data says:
- Open-source models: 62% of token consumption, 8.6% of spending
- Anthropic: 30% of token consumption, 65.1% of spending
- Total token volume: 59% month-over-month growth
- DeepSeek: now the #2 model provider by token volume, ahead of Google's Gemini
Run the math. If open-source models consume 62% of tokens but only 8.6% of spend, their effective unit price is roughly 1/15th of the closed-source average. That is not a price differential. That is an entirely different economic species.
Anthropic's effective token price is roughly 2.2x the market average. Open-source is at roughly 0.14x. The gap between those two numbers is not a gradient. It is a chasm.
Yield is a function of risk, not just time. The same principle applies to AI models. The perceived risk of open-source models — reliability, security, alignment, support — is being priced into that 15x discount. Whether that risk is real or imagined is a separate question. The market is pricing it, and the market is always making a statement.
What 62% Actually Means
I want to slow down here, because there is a tendency to read "62% open-source token share" as "open source wins." That is a category error. Token volume measures the number of discrete API calls being made. It does not measure the complexity of each call, the duration of each call, or the value that each call generates.
In my smart contract work, we measure not just the number of transactions but the value flow through each function. A contract that processes 10,000 transactions of $10 each is different from a contract that processes 100 transactions of $100,000 each. The first is high-frequency, low-value, and efficient. The second is low-frequency, high-value, and critical.
The same logic applies to model consumption.
The 62% token share is likely concentrated in tasks like:
- Code autocompletion
- Text classification
- Information extraction
- Template-based content generation
- Caching heavy deterministic operations
These are high-frequency, low-complexity, cost-sensitive workloads. They are the kind of tasks where a 15x price differential makes open-source models the rational economic choice. If a task costs $0.01 per call with a closed-source model and $0.0007 with an open-source model, the developer will switch. Every time. The quality delta in these tasks is negligible.
The 30% token share that Anthropic holds — the one generating 65.1% of spend — is a different category entirely. These are likely:
- Complex multi-step reasoning
- Long-context understanding
- Code generation for mission-critical infrastructure
- Creative work that requires consistency and judgment
- Financial or legal analysis where errors have asymmetric consequences
In these tasks, the 2.2x price premium is insurance. It is the cost of reducing the probability of a catastrophic failure. And in production environments, catastrophic failures are not theoretical. They are lived experiences.
The DeepSeek Signal
Let me speak about DeepSeek specifically, because it is the most technically interesting data point in this entire story.
DeepSeek has surpassed Google in token consumption on Vercel. That means more production traffic is hitting DeepSeek's endpoints than Google's Gemini. This is not a benchmark score. This is real developer traffic, routed through a real platform, processed by real applications.
Based on my audit experience, when I see an open-source model gaining this kind of production adoption, I do not attribute it to quality alone. I attribute it to three factors working in combination: quality, price, and integration friction. The model must be good enough to pass the bar of production reliability. The price must be low enough to justify switching. And the integration friction must be small enough to not eat the savings.
DeepSeek's MoE architecture gives it a structural cost advantage that makes aggressive pricing sustainable — at least at current scale. The 59% month-over-month growth in total token volume on Vercel suggests that this low price point is not merely shifting existing demand from one provider to another. It is creating new demand. Developers are building applications that would not have been economically viable at higher price points.
But here is the critical question that I have not seen anyone ask: at what capacity utilization rate does DeepSeek's unit economics break even? MoE architectures have a peculiar cost profile. They reduce the parameter count for each inference, but they require substantial memory to hold all the expert weights. At low utilization, the cost per token is high because you are amortizing memory costs over a small inference volume. At high utilization, the cost per token drops sharply.
The token growth on Vercel suggests DeepSeek is approaching the utilization threshold where their economics become genuinely competitive. This is not a question of model quality. This is a question of operating leverage.
The Hidden Cost Structure
Here is where I want to get technical, because the market analysis ignores this entirely.
The apparent 15x price differential between open-source and closed-source models is not a pure model quality differential. It is also a product of the serving infrastructure and the safety overhead. Anthropic and OpenAI invest heavily in alignment, guardrails, and compliance. These are not free. They are costs that are passed on to users.
Audit reports are promises, not guarantees.
The same logic applies to closed-source model providers. Anthropic's safety stack is a real, operational cost. It requires engineering time, compute resources, and ongoing maintenance. The result is a model that is safer by design — but also more expensive.
The open-source model community, by contrast, has no centralized safety budget. The security posture of an open-source model depends on the community that surrounds it. Some communities are rigorous. Others are not. The variance in safety and reliability between different open-source models is enormous, and developers who deploy them are, in effect, accepting a portion of that risk. If the model is deployed on an edge, the application developer — not the model provider — owns the liability.
This is the hidden transfer of risk. When a developer switches from Claude to a DeepSeek deployment, they are not only paying less. They are also buying less insurance. Whether that is a rational trade depends on the application. For a code snippet generator? Rational. For a system that executes financial transactions? I would think twice.
The Blind Spot in the Market Structure
Now I want to look at the blind spot that nobody in this conversation is addressing: the assumption that open-source models and closed-source models are on a linear competition trajectory. That is the market's belief. I think it is wrong.
The Vercel data suggests something different. It suggests the market is forming a two-layer structure that resembles the cellular architecture of a distributed system:
Layer 1 — The Commodity Layer: Open-source models dominate high-volume, low-complexity tasks. The profit margin per token is thin. The competitive advantage is price and latency. The market structure resembles AWS — a volume game where cost efficiency and reliability determine the winner.
Layer 2 — The Value Layer: Closed-source models dominate high-complexity, high-consequence tasks. The profit margin per token is high. The competitive advantage is quality, safety, and brand trust. The market structure resembles enterprise software — a trust game where brand reputation and reliability determine the winner.
These two layers are not in competition with each other. They are complementary. The open-source layer absorbs the volume that would otherwise be economically irrational for closed-source models to serve. The closed-source layer captures the value that open-source models cannot yet deliver.
The Vercel data supports this hypothesis: open-source models have 62% of the token share but only 8.6% of the spend. Anthropic has 30% of the token share and 65.1% of the spend. These are two markets, not one.
The question that matters is whether this separation is permanent. And that is a question I cannot answer with confidence.
The Price Anchor Effect
The open-source model price points have created a new anchor in the market. The economics of any closed-source model is now judged against the alternative cost of deploying an open-source model. This is a structural shift in pricing power.
In the smart contract world, we call this a "reference price." Once a reference price is established, every higher price must be justified by measurable differences in quality, security, or reliability. This is why Anthropic can sustain a 2.2x price premium — because the quality differential in complex tasks is genuinely observable. This is also why Google is struggling: its models are expensive, but their quality differential does not justify the price premium.
The Google case is instructive. Google's Gemini has been a leading closed-source model, yet DeepSeek's open-source model has surpassed it in token consumption. This is not a reflection of technical quality alone. It is a reflection of value-per-dollar. When developers compare two models for a specific task, they do not compare benchmark scores. They compare the cost of achieving their desired outcome.
If a task can be completed with a DeepSeek model at a fraction of the cost, the benchmark scores become irrelevant. The developer switches.
Google's position is not just a pricing problem. It is a positioning problem. It is not clear what unique value the Gemini ecosystem provides that cannot be replicated by an open-source model. Without that differentiation, the price premium cannot be justified.
The same question applies to OpenAI, though less sharply. OpenAI has a stronger brand and a more established ecosystem. But in the Vercel data, the trend is clear: the open-source camp is taking volume share, and the value is consolidating into Anthropic.
The market is dividing into two tiers: a premium tier, where quality is the currency, and a commodity tier, where price is the currency. The middle tier is being squeezed.
The Measurement Problem
I want to step back and think about the measurement problem in this data. The Vercel platform is a web infrastructure provider. Its user base is developers building web applications. That is not a representative sample of the AI market. It over-represents code generation, content generation, and front-end automation. It under-represents enterprise workloads, financial services, legal analysis, and research.
This means the 62% open-source token share is likely overstated relative to the overall market. The enterprise AI market — where closed-source models dominate — is not captured in this data. That is a critical caveat.
However, the direction of the trend is still meaningful. Even if the absolute numbers are skewed, the trajectory is clear. Open-source token share is growing rapidly. The question is not whether it is growing. The question is how much of the total market it can eventually capture.
My estimate — based on the economics I am seeing — is that the open-source models will eventually settle in the range of 50-60% of total token volume in production. But they will likely settle in the range of 15-25% of total spending. The remaining 75-85% of AI revenue will remain with a small set of closed-source providers.
This is not a death spiral for closed-source models. It is the opposite. It is a clarification of their role. They are not competing for volume. They are competing for value. And in the value game, the winners are the ones with the best models, not the cheapest ones.
The Investment Implication
For the investment community, the Vercel data has a deeper meaning. The market is starting to realize that not all AI revenue is created equal. Token volume is a vanity metric. Revenue is a reality metric. The market cap of an AI company should be a function of revenue, not of token volume.
This is the same lesson that DeFi learned in 2020. Total value locked (TVL) was the vanity metric. Revenue was the reality metric. When the TVL stopped growing, the projects with no revenue collapsed. The projects with real revenue survived.
In the AI market, the equivalent of TVL is token volume. The equivalent of revenue is spend. The Vercel data shows that the market is already pricing this distinction. Anthropic has 2.1x the revenue per token of the market average. That is why it is one of the highest-valued AI companies in the world.
The open-source model providers, by contrast, have revenue per token of about 0.7x the market average. Their valuation logic must be different. They are infrastructure providers, not software providers. They are valued on volume and scale, not on margin.
The market has not fully priced this difference yet. There is still a lot of confusion about how to value AI model providers. But the direction is clear: the market will eventually reward revenue, not volume.
The Unsustainable Question
The open-source pricing model has a serious sustainability problem that I want to highlight. When a model provider charges 1/15th of the market rate for tokens, one of two things is happening. Either their cost structure is radically better, or they are subsidizing the price.
DeepSeek's MoE architecture gives them a genuine cost advantage. But is that advantage enough to justify the 15x price differential? I am skeptical.
The likely scenario is that DeepSeek is pricing below cost to acquire market share. This is a classic penetration pricing strategy. It works in the short term, but it is not sustainable indefinitely. At some point, the provider needs to raise prices or find other ways to monetize the attention.
If DeepSeek raises prices, the token share will decrease. If they keep prices low, they will continue to capture share but with minimal revenue. The fundamental tension is that the same strategy that creates market share growth creates revenue compression.
This is not a new problem. It is the same problem that every disruptive pricing strategy has faced. The question is whether the volume growth will eventually justify the unit economics. In the AI market, the cost per token will continue to decline due to hardware improvements and algorithm optimization. So the absolute cost of serving a token will drop. That means DeepSeek might be able to maintain its price while improving margins over time.
But there is a limit to that logic. The cost of serving a token can only drop so far. At some point, the price needs to reflect the true cost of service.
I am not predicting a collapse. I am predicting a recalibration. The 15x price differential is not sustainable. It will narrow. The question is how narrow it will get and how quickly.
The Contrarian Angle: The Data Is Not What It Looks Like
Here is where I want to push back against the prevailing narrative.
The narrative is: "Open source is winning. The closed-source model is losing." The Vercel data does not support that conclusion. It supports a different conclusion.
The open-source model is winning the volume game. But it is losing the value game. And the value game is the one that matters for long-term economics.
This is the same mistake we made in the early days of blockchain. We counted the number of transactions as a measure of success. But a blockchain with 1 million transactions of $1 each is less valuable than a blockchain with 100 transactions of $1 million each. The value is in the size of the flow, not the number of the flow.
The same applies to AI models. The value is in the size of the spend, not the number of tokens.
The Vercel data tells us that the market is already pricing this difference. Anthropic's 65.1% of spending on 30% of tokens is not a sign of overpricing. It is a sign of value. The market is paying for what it gets.
The contrarian angle is this: the open-source model is not a threat to closed-source models. It is a complement. It is absorbing the low-value volume that closed-source models do not want to serve. It is expanding the market. It is creating new use cases that would not exist at high prices.
And by expanding the market, it is creating more demand for the high-value tasks that only closed-source models can handle.
The end result is not a win-lose. It is a win-win. The open-source model grows the market. The closed-source model captures the value.
The only loser is the middle tier: models that are not cheap enough to be commodity and not good enough to be premium.
The Security Implication
As a smart contract auditor, I want to raise a concern that no one in the AI market is paying attention to. The open-source model is a supply chain security problem.
When you use a closed-source model, you are trusting a centralized provider to ensure the model is safe. The provider has an incentive to protect its reputation. The provider has legal liability. The provider can be held accountable.
When you use an open-source model, you are trusting a community. The community may or may not have the resources to ensure safety. The community may or may not be accountable. The community may or may not exist in five years.
This is not a theoretical concern. We have seen open-source models with backdoors. We have seen open-source models with hidden data collection. We have seen open-source models that were trained on data that was not authorized for use.
The risk is not in the model itself. The risk is in the deployment. When you deploy an open-source model, you are responsible for the entire supply chain. If the model has a vulnerability, you are the one who is liable.
This is exactly the same problem that we face in smart contracts. The code is open. The code is audited. But the risk is not in the code. It is in the deployment. It is in the integration. It is in the edge cases.
Audit reports are promises, not guarantees. The same applies to open-source model reviews.
This is not a reason to avoid open-source models. It is a reason to be careful about how you deploy them. The security posture is different. The risk profile is different. The responsibility is different.
The 15-Month Forecast
Let me make a forecast. I do not make forecasts casually. I make them based on the patterns I see in the data.
In the next 15 months, I expect to see the following:
- The token share of open-source models will plateau. The 62% share is not a linear trajectory to 80%. It will stabilize in the range of 55-65%, because there are high-value tasks that the open-source model cannot perform reliably.
- The value gap will persist. The open-source model will continue to generate 10-15% of the economic value of the market, even at 55-65% token share. This is because the value of a token is not uniform. It is a function of the complexity of the task.
- The premium tier will consolidate. The Anthropic-style players — models that are differentiated by quality — will capture an increasing share of the revenue. The middle tier will be squeezed.
- The pricing gap will narrow. The 15x differential is not sustainable. It will narrow to 5-8x, because the cost structure of open-source providers will improve.
- The security issue will surface. At some point, there will be a significant security incident involving an open-source model deployment. This will be a wake-up call. It will not destroy the open-source ecosystem, but it will change the risk calculus.
- The enterprise shift will be slower than expected. Enterprise AI workloads will not move to open-source models at the same rate as developer workloads. Enterprise needs are different: they need support, accountability, and compliance. These are not free.
The Structural Conclusion
The Vercel data is not a story about open-source winning. It is a story about the AI market becoming more efficient. The market is allocating tokens to the lowest-cost provider for each task. That is how markets should work.
The open-source model is the commodity layer. The closed-source model is the value layer. The market is pricing both correctly.
This is not a zero-sum game. It is a positive-sum game. The open-source model is expanding the market. The closed-source model is capturing the value.
The developers who understand this structure will be the ones who win. They will use open-source models for high-volume, low-complexity tasks. They will use closed-source models for high-value, high-complexity tasks. They will not force a false choice between the two.
Yield is a function of risk, not just time. The same is true for AI models. The yield on a token is a function of the risk-adjusted value of the task. The open-source model has a lower risk-adjusted value per token. The closed-source model has a higher risk-adjusted value per token.
The market is pricing this. The Vercel data is the proof.
The next step is to understand the structure, not to fight it. The market is already telling us what the structure is. The question is whether the market participants are listening.
I am listening. I am not celebrating. I am not mourning. I am mapping the structure. That is the only work that matters.