The 62/8.6 Divergence: What Vercel's Token Data Really Says About AI's Value Layer

You think open source won the AI war. The numbers say otherwise. Vercel's latest model usage data shows open-source models now account for 62% of all tokens consumed on the platform. Two months ago, that number was 28.4%. DeepSeek, a Chinese open-weight model, has overtaken Google to become the second-largest model provider by token volume. The narrative writes itself: open source is eating the world. Then you look at the spending side. Open-source models generate 8.6% of total expenditure. Anthropic, with 30% of token volume, captures 65.1% of every dollar spent. That is not a victory lap. That is a margin call dressed in a growth chart. Sentiment is noise; liquidity is the signal. And the liquidity is telling a very different story than the token counters. I have been tracking this divergence since my 2023 Arbitrum arbitrage experiment taught me a simple lesson: volume without value extraction is just noise with extra steps. My MEV bot processed thousands of transactions. It lost $1,200. The mempool was full of activity, but the economic value was concentrated in a handful of sophisticated players who understood latency, slippage, and order flow. The same mechanics are playing out in the AI model market. Token volume is the mempool. Spending is the P&L. They are not the same thing. Let me break down the context before we go deeper. Vercel is a deployment platform used primarily by web developers and front-end engineers. Its AI gateway routes requests to various model providers, making it a useful proxy for real-world developer adoption patterns. The platform's data covers a specific demographic: builders shipping production applications, not researchers running benchmarks. This matters because it reflects actual usage, not hype. The data covers a two-month window, which is short but directionally significant. The platform saw total token volume grow 59% quarter-over-quarter, driven largely by the price elasticity of open-source models. When the cost per token drops by an order of magnitude, developers suddenly find new use cases that were previously uneconomical. That is the demand curve doing what demand curves do. Now let me get into the core analysis. The first thing to understand is the unit economics. Open-source models deliver 62% of tokens at 8.6% of spending. That implies a unit price roughly one-fifteenth of the closed-source average. Anthropic, by contrast, commands a unit price roughly double the market average. This is not a small gap. This is a structural difference in how value is created and captured across the AI stack. The token-to-spending ratio is the single most important metric in this entire dataset. It tells you that open-source models are being used for high-frequency, low-complexity tasks. Code completion. Text classification. Information extraction. The kind of work that needs to happen a million times a day and does not require deep reasoning. These are commodity workloads. They are the AI equivalent of a utility meter reading. Necessary, voluminous, and priced accordingly. Closed-source models, particularly Anthropic's Claude family, are being deployed for the opposite end of the spectrum. Complex reasoning. Multi-step analysis. Creative synthesis. Tasks where a single wrong answer carries real cost. In those scenarios, paying two to fifteen times more per token is rational. The cost of a bad output in a legal document review or a financial model far exceeds the token price differential. This is not a failure of open source. It is a correct market segmentation. DeepSeek's rise deserves specific attention. The model surpassed Google in token consumption on Vercel, which is a significant data point. But I need to be precise about what this means. DeepSeek's token volume is a function of two variables: capability and price. The model's benchmark scores are competitive, but benchmarks do not always translate to production performance. What the Vercel data suggests is that in real-world engineering tasks, DeepSeek's performance-to-price ratio is compelling enough to drive mass adoption. That is a real signal. But it is not the same as saying DeepSeek is a better model than Gemini. It is saying DeepSeek is a better value for a specific class of workloads. Here is where I bring in my own experience. In 2020, I deployed $15,000 into a yield farming protocol that promised 400% APY. I ignored the lack of audits because the returns were too attractive. The smart contract was exploited. I lost $12,000. The lesson was brutal and permanent: when the price is too good, the risk is hiding somewhere you are not looking. The same principle applies to DeepSeek's pricing. The model is being offered at a price point that may be below cost. This is a classic land-grab strategy. It works in the short term. It is not a sustainable business model unless the provider has a path to profitability through scale, optimization, or complementary revenue streams. I am not saying DeepSeek is a scam. I am saying the economics need scrutiny. The MoE architecture and inference optimizations that enable low-cost serving are real engineering achievements. But there is a difference between efficient serving and loss-leading pricing. If DeepSeek is subsidizing usage to build market share, the current token volume is not a reliable predictor of long-term market structure. It is a promotional period. And promotional periods end. Let me talk about the demand elasticity effect, because this is a hidden insight that most commentary misses. The 59% quarter-over-quarter growth in total token volume is not just organic demand. It is induced demand. When the marginal cost of a token drops from $0.01 to $0.0006, developers start using models for tasks they previously would not have automated. They add AI features to products that were not AI-enabled. They build internal tools that were not worth building. This is the Jevons paradox applied to AI: as the cost of a resource decreases, total consumption increases. The open-source price war is not just stealing share from closed models. It is expanding the entire market. This is good for the ecosystem. It is also a trap for anyone who mistakes token growth for value creation. Now let me address the competitive dynamics. The Vercel data reveals a three-tier market structure forming in real time. At the top, you have Anthropic, capturing outsized economic value through premium pricing on complex workloads. In the middle, you have OpenAI, growing token volume but facing pressure from both sides. At the bottom, you have the open-source cluster, generating massive volume at thin margins. Google's position is the most interesting. Being overtaken by DeepSeek in token volume is a warning sign. It suggests that Google's models are not compelling enough on either capability or price to win developer mindshare. Google has the infrastructure, the talent, and the distribution. What it lacks is a clear value proposition in the developer market. That is a strategic problem, not a technical one. OpenAI's position is more nuanced. The company is growing, but it is caught in a pincer movement. Open-source models are undercutting it on price for commodity workloads. Anthropic is outcompeting it on quality for high-value workloads. OpenAI's differentiation is brand recognition and ecosystem lock-in, which are real but eroding. The company needs to either push further upmarket into Anthropic's territory or find a way to compete on cost with open-source providers. Doing both is difficult. Doing neither is fatal. I want to address the platform bias question, because it matters for how we interpret this data. Vercel's user base skews toward web developers and front-end engineers. This population is more price-sensitive and more likely to experiment with open-source models than, say, enterprise data teams or financial institutions. The token distribution on Vercel likely overstates open-source adoption in the broader market. Enterprise workloads, particularly in regulated industries, still favor closed-source models for compliance, support, and liability reasons. The Vercel data is a directional signal, not a market census. I would not build an investment thesis on it alone. But I would use it to question any thesis that assumes open source is irrelevant. Here is the contrarian angle. The mainstream narrative is that open source is winning and closed source is losing. The data does not support that conclusion. What the data actually shows is a bifurcation of the market into two distinct layers with different economic characteristics. The open-source layer is winning on volume. The closed-source layer is winning on value. These are not contradictory. They are complementary. The market is not moving toward one model. It is moving toward a dual-layer structure where open source handles the long tail of high-frequency, low-complexity tasks, and closed source handles the high-value, high-complexity core. This has profound implications for how we value AI companies. The market has been rewarding token growth as a proxy for adoption. The Vercel data suggests that token growth without economic value capture is a hollow metric. Anthropic's 65.1% spending share on 30% token volume is the kind of metric that justifies a premium valuation. DeepSeek's 62% token share at 8.6% spending is the kind of metric that justifies a utility multiple. Investors who conflate the two are making the same mistake I made in 2017 when I bought ICO tokens based on whitepaper hype. I lost 94% of my savings because I confused narrative with fundamentals. The market does not care about your conviction. It cares about your cash flows. Let me also address the sustainability question. Can open-source providers maintain their pricing? The answer depends on their cost structure. DeepSeek's MoE architecture and inference optimizations are genuine innovations that lower serving costs. But there is a floor. Compute costs, bandwidth costs, and engineering salaries do not go to zero. If DeepSeek is pricing below cost, it is burning capital to acquire market share. This is a viable strategy if the company can eventually raise prices or monetize through other means. It is a death spiral if it cannot. The same logic applies to other open-source providers. The market will eventually sort out which providers have real cost advantages and which are subsidizing adoption. There is also a geopolitical dimension that the data surfaces. DeepSeek's rise on Vercel is not just a commercial story. It is a signal that Chinese AI models are competitive in international developer markets. This has implications for export controls, data governance, and the broader technology competition between the US and China. Western regulators are already scrutinizing Chinese AI models for data security risks. The Vercel data gives them a concrete data point: Chinese models are being used in production environments by Western developers. This will accelerate regulatory attention. It may also accelerate efforts by Western companies to develop competitive open-source alternatives. I want to bring in my 2022 LUNA experience here, because the parallel is instructive. I held $20,000 in UST and Luna, believing in the algorithmic stability model. When the peg broke, I refused to sell because I was emotionally attached to the narrative. I watched the value evaporate to near zero. The lesson was not about stablecoins. It was about the danger of confusing a compelling story with a sound mechanism. The open-source token share story is compelling. But the mechanism that generates 62% token share at 8.6% spending share is not a value-creation mechanism. It is a cost-arbitrage mechanism. Cost arbitrage is real. It is also commoditizing. And commoditized markets do not generate outsized returns for the providers. Let me talk about what this means for application-layer builders, because that is where the real opportunity lies. The collapse in token prices is a gift to developers. It means you can build AI features that were previously uneconomical. It means you can experiment with models without worrying about your API bill. It means the marginal cost of intelligence is approaching zero. This is the same dynamic that drove the internet boom: when the cost of distribution dropped to zero, the value shifted to the application layer. The same thing is happening in AI. The model layer is commoditizing. The application layer is where value is being created. Builders who understand this will capture the upside. Builders who are still trying to pick winning models are fighting the last war. I also want to address the quality question directly. The data suggests that open-source models are being used for lower-complexity tasks. But this is a snapshot, not a permanent state. Open-source models are improving rapidly. The gap between open and closed models on complex reasoning tasks is narrowing. If that gap closes, the economic bifurcation I described will start to erode. Closed-source providers will need to justify their premium pricing with demonstrably superior capabilities. If they cannot, the spending share will eventually follow the token share. This is the scenario that keeps Anthropic and OpenAI executives up at night. It is also the scenario that makes the current data a critical baseline for future comparisons. Let me get into the mechanics of how this plays out. The token-to-spending divergence is not static. It is a dynamic equilibrium that shifts with model capabilities, pricing strategies, and developer preferences. If open-source models continue to improve at the current rate, the quality gap will narrow. At some point, the premium for closed-source models will exceed the value they deliver. That is the tipping point. When it happens, the spending share will start to migrate. The migration will be slow at first, then sudden. This is how market structure changes happen. They look like trends until they look like cliffs. I have seen this pattern before. In 2024, I identified a basis trade between spot ETFs and perpetual futures. The strategy yielded a steady 8% annualized return with minimal volatility. The opportunity existed because institutional flows were creating predictable price dislocations. I executed the hedge manually across two exchanges. It worked because I understood the mechanics. The same principle applies here. The opportunity in the AI model market is not in picking winning models. It is in understanding the structural mechanics that determine where value accumulates. The Vercel data is a window into those mechanics. Here is what I think the next twelve months look like. Open-source token share will continue to grow, possibly reaching 70-75% of total volume. Spending share will remain concentrated in closed-source models, but the gap will narrow as open-source capabilities improve. Anthropic will maintain its premium positioning but face increasing pressure from open-source alternatives. OpenAI will need to make strategic choices about where to compete. Google will either find a way to regain developer mindshare or continue to lose ground. The dual-layer market structure will solidify. And the companies that thrive will be the ones that understand which layer they are operating in. I want to close with a framework for thinking about this. Trust the ledger, not the legend. The legend is that open source is winning. The ledger says something more nuanced. Open source is winning volume. Closed source is winning value. Both are true. The question is which trend you are positioned to benefit from. If you are building applications, the open-source price collapse is your tailwind. If you are investing in model providers, the spending concentration is your signal. If you are doing both, you need to understand the mechanics of each layer. Sunk cost is the anchor that drowns traders alive. I learned this in 2018 when I held my ICO bags to zero because I could not accept the loss. I learned it again in 2022 when I held UST because I believed the narrative. The market does not care about what you paid. It cares about what the asset is worth. The same logic applies to AI models. The market does not care about how many tokens a model processes. It cares about how much value those tokens create. The Vercel data is a reminder that volume and value are not the same thing. They never have been. They never will be. I don't predict the wave; I build the board. That is my approach to markets, and it applies here. The wave is the open-source adoption curve. It is real, it is powerful, and it is reshaping the industry. The board is the framework for capturing value from that wave. It requires understanding the economic structure, not just the usage statistics. It requires knowing where the margin is, not just where the volume is. And it requires the discipline to act on that understanding, even when the narrative says otherwise. The final question is not whether open source will continue to grow. It will. The question is whether the companies and developers riding that growth will capture the value they create. The data suggests that most of them will not. The value will flow to the application layer, to the infrastructure layer, and to the few model providers that can command premium pricing. That is the market structure. That is the signal. The rest is noise. Here is my forward-looking take. The next twelve months will separate the builders from the believers. The builders will use open-source models to create applications that generate real revenue. The believers will keep arguing about which model is better. The market will reward the builders. It always does. The token data is just the scoreboard. The game is value creation. And value creation does not care about your favorite model. It cares about your unit economics, your customer acquisition costs, and your ability to deliver something people will pay for. That is the ledger. Everything else is legend.