The Token Paradox: When 62% of Usage Generates Only 8.6% of Value

The numbers hit like a cold wave. Two months ago, open-source models commanded 28.4% of token traffic on Vercel's AI gateway. Today, that figure stands at 62%. Yet those same models account for just 8.6% of total spending. Meanwhile, Anthropic—with a mere 30% of token volume—captures 65.1% of every dollar flowing through the platform. Bulls would call this a revolution. I call it a reckoning. The gap between usage and value isn't a statistical quirk. It's a mirror reflecting something deeper about how we measure worth in the age of machine intelligence. Tech changes. Values remain. And the values embedded in this data tell a story that most market commentary has completely missed. Let me step back and explain what we're actually looking at. Vercel operates one of the most widely used deployment platforms for web applications. In late 2024, they launched an AI gateway that routes developer requests to various model providers—OpenAI, Anthropic, Google, DeepSeek, and a growing roster of open-source alternatives. The platform's telemetry offers a rare, real-time window into how thousands of production applications actually consume AI models. This isn't a benchmark leaderboard or a research paper. It's ground truth from the trenches of software development. The data reveals three seismic shifts happening simultaneously. First, open-source models have crossed a credibility threshold that took most industry observers by surprise. Second, DeepSeek—a Chinese open-source model family—has overtaken Google in raw token consumption, becoming the second-largest model provider on the platform. Third, the total token volume grew 59% month-over-month, suggesting that cheap open-source models aren't just substituting for closed alternatives—they're creating entirely new demand. Based on my audit experience across dozens of AI infrastructure projects, I can tell you that this last point deserves far more attention than it's receiving. When the marginal cost of a token drops by an order of magnitude, developers change their behavior. Tasks that were previously too expensive to automate—text classification at scale, content summarization pipelines, code completion for every keystroke—suddenly become viable. The 59% growth isn't just existing demand shifting providers. It's latent demand being unlocked by price elasticity. This is the same dynamic we witnessed in crypto when Layer-2 solutions finally made transaction costs negligible. Usage exploded not because users became more sophisticated, but because the friction disappeared. Now let's dig into the core tension. The 62% versus 8.6% split is the most important number in this entire dataset, and it's being widely misinterpreted. Open-source models are not winning because they're better. They're winning because they're cheap enough to be disposable. The average token from an open-source model costs roughly one-fifteenth of what Anthropic charges. That price differential fundamentally changes the calculus of what tasks get routed where. Here's what the raw numbers conceal: the 62% of tokens flowing through open-source models are likely concentrated in low-to-medium complexity tasks. Code completion. Information extraction. Text classification. Sentiment analysis. These are high-frequency, repetitive workloads where a 95% accuracy rate is perfectly acceptable and the cost savings are massive. Meanwhile, the high-value, complex reasoning tasks—multi-step analysis, creative writing, complex code generation, strategic planning—continue to flow to closed models like Anthropic's Claude family. This creates a two-tier market structure that mirrors what we've seen in every technology revolution. The commodity layer gets commoditized. The premium layer commands premium pricing. Open-source models are becoming the AWS of AI—massive scale, thin margins, infrastructure-grade reliability. Closed models are becoming the specialized consulting firms—smaller volume, but each engagement carries outsized economic weight. The DeepSeek story deserves particular scrutiny. A Chinese open-source model overtaking Google's Gemini in token consumption on a Western developer platform is not a small event. It signals that the performance-to-price ratio of open-source models has reached a point where developers are making rational economic choices that override brand loyalty or geopolitical comfort. Verify the code, trust the community. That's the principle at work here. Developers aren't choosing DeepSeek because they love China. They're choosing it because the math works. But here's where I need to inject a note of caution that most commentary is missing. The open-source victory lap is premature. Token share is not market share. Usage is not revenue. And the 8.6% spending figure reveals a uncomfortable truth: open-source models are generating enormous value for developers while capturing almost none of it for themselves. The providers—DeepSeek, Meta with Llama, Alibaba with Qwen—are running what appears to be a subsidy strategy. Prices that low, at that scale, suggest either extraordinary engineering efficiency or deliberate below-cost pricing to capture market position. This is where my contrarian angle comes in. The conventional narrative says open-source is winning and closed models are doomed. I think the opposite is closer to the truth. The data suggests that closed models are actually consolidating their grip on the highest-value segments of the market. Anthropic's 65.1% spending share with only 30% token volume isn't a sign of weakness—it's a sign of pricing power. The market is telling us that complex reasoning is worth 15 times more than routine processing. That premium isn't going to evaporate just because open-source models get better at routine tasks. Consider the parallel to the crypto world. Bitcoin handles the vast majority of transaction volume in terms of network security and hash rate. But the economic value of smart contract platforms—Ethereum, Solana, and their Layer-2 ecosystems—far exceeds Bitcoin's in terms of fee generation and application value. Bulls react. Bears reflect. We build. The builders understand that different layers serve different purposes. The same logic applies here. Open-source models are becoming the settlement layer of AI—ubiquitous, reliable, and cheap. Closed models are becoming the application layer—where the real economic value is created and captured. The deeper question is whether this bifurcation is sustainable. Open-source model providers are burning capital to maintain their price advantage. If DeepSeek's costs are truly as low as they appear—through Mixture-of-Experts architecture and aggressive inference optimization—then the model is sustainable. If they're subsidizing usage to buy market share, we're looking at a classic burn-and-churn play that will end in consolidation or price increases. There's also a governance dimension that the data can't capture. Open-source models shift security responsibility from the provider to the developer. When you call Anthropic's API, you're getting their safety layers, their alignment work, their red-teaming. When you self-host an open-source model, that responsibility falls on you. The 62% token share means a massive amount of AI inference is now running without the protective infrastructure that closed providers have built. This isn't an argument against open source—it's a warning that the ecosystem needs new safety tooling to fill the gap. Looking at the competitive landscape, Google's position is the most precarious. Being overtaken by DeepSeek in token consumption is a wake-up call. Google has the best research team in the world, but their developer-facing model strategy has been unfocused. Gemini's performance is competitive, but the pricing and accessibility haven't created the same developer loyalty that OpenAI and Anthropic have cultivated. In a market where switching costs are near zero, developer experience and price-performance ratio are everything. OpenAI sits in an uncomfortable middle position. Their token volume is growing, but they're being squeezed from below by open-source models and from above by Anthropic's premium positioning. The question OpenAI must answer is whether they want to compete on price or on quality. Trying to do both will leave them stranded in the middle—too expensive to win the commodity tier, not differentiated enough to justify the premium. The investment implications are profound. The market is beginning to understand that token volume is a vanity metric. What matters is the economic value captured per token. Anthropic's valuation premium is justified by their 65.1% spending share. DeepSeek's impressive token numbers, by contrast, may not translate into proportional revenue. Investors need to distinguish between usage leaders and value leaders. They are not the same thing. So where does this leave us? The two-layer structure is forming faster than most analysts predicted. Open-source models will continue to dominate token volume, driving down costs and expanding the addressable market for AI applications. Closed models will continue to capture outsized economic value by serving the complex, high-stakes tasks that demand the best available reasoning. The winners will be the developers who understand which layer to use for which task—and the investors who understand that usage and value are diverging, not converging. The next twelve months will tell us whether DeepSeek's rise is sustainable or a flash in the pan. Whether Anthropic can maintain its pricing power as open-source models improve. Whether Google can recover its footing. But the underlying trend is clear: the AI model market is bifurcating into a commodity layer and a premium layer, and the economic gravity is pulling in opposite directions. The builders who navigate this divide will define the next era of software. The rest will be left holding tokens that don't translate into value. Verify the code, trust the community. But never confuse usage with worth.