There's a moment every founder dreads. It's not the funding round that falls through or the product launch that fizzles. It's the quiet realization that the metrics everyone's been celebrating are actually telling a different story than the one we tell ourselves. We see a chart trending upward and call it progress. We see a new leaderboard position and call it a victory. But if we're honest with ourselves, these numbers rarely capture the true weight of what we've built.
Consider a recent disclosure from Vercel's CEO that surfaced across the developer ecosystem. At first glance, it looked like another piece of bullish AI adoption data—open-source models now account for over 62% of all tokens processed on the platform. For those of us who've watched the pendulum swing from proprietary software to open systems for decades, it seemed like a vindication. But the second number in that disclosure is the one that keeps me awake at night. Those same open-source models accounted for only 8.6% of Vercel's AI spending. The closed-source models, with their shrinking 38% token share, still commanded a staggering 91.4% of the platform's expenditure.
The data points to a deeper truth about our industry. We are seeing the rise of an "inverted economy"—a world where the masses are moving through the gates, but the vast majority of treasure remains in a guarded fortress. The battle for AI's future isn't about who gets the most foot traffic; it's about who gets the highest-value cargo. For every two tokens generated by open-source models, closed-source models generate only one, yet they capture ten times the revenue. That's not a market inefficiency; that's a power structure.
My own journey into this space began not with a token, but with a whitepaper. In 2017, during the ICO boom, I applied my financial engineering background to audit over fifty blockchain projects, looking for viable economic models rather than inflated promises. I found only 12 that made any real sense. That experience taught me a crucial lesson: technology serves human trust, not the other way around. And it's the same lens I use today to look at Vercel's data. We are not just looking at a usage report; we are seeing the financial skeleton of an entire industry.
The 62% token share is not a victory for open source—it's a surrender of economic value. It's a classic trap. We've built an economy where the dominant contributor is also the most undervalued. The dream of open-source was supposed to be about freedom and accessibility. But if the only people who can profit from it are those who can spend 14 times more per token, then we've simply built a new kind of walled garden.
DeepSeek's rise to become the second-largest provider on Vercel, overtaking Google, is a landmark event. It proves that a Chinese model with a better engineering approach and significantly lower pricing can disrupt the established hierarchy. But let's not mistake this for a triumph of the open-source ethos. DeepSeek's token count may be high, but its value capture is minimal. This is a marathon being run by a sprinter who has no idea where the finish line is.
Let’s be clear: the developer is not stupid. Developers are the most pragmatic engineers in the world. They migrate from closed to open models for two reasons: cost and control. They aren’t doing it to make a philosophical statement. This mass migration of workload is only happening because the open-source models have reached a quality threshold for many common tasks. I’ve seen this pattern before in the open-source software movement. In the early 2000s, Linux started eating the server market, not because it was more user-friendly, but because it was reliable enough and significantly cheaper. The same thing is happening in AI—but the economic structure around it is different.
This brings us to the central paradox of the AI revolution. The majority of usage is moving to open-source, but the vast majority of capital is flowing to closed models. It is a conflict of scale and value.
Let's dive deeper into the numbers. If open source uses 62% of tokens and only spends 8.6% of the budget, it implies that the unit price for these tokens is about 1/14th of closed source. That's not just a cost difference; that's a deliberate strategy of "penetration pricing." Open-source models like DeepSeek are not priced at their true cost; they’re priced to capture the market share and ecosystem position. They are the loss leaders in a larger campaign. In the short term, this is fantastic for a developer building a product. You can iterate endlessly, test every hypothesis, and build features that would be prohibitively expensive with GPT-4o. This lowers the barrier to entry and accelerates the speed of innovation.
But here's the trap: if you build your entire business on the backs of these cheap tokens, you're building your castle on sand. When the open-source model providers decide to raise prices to achieve sustainability—which they must eventually do to survive—your entire cost structure is threatened. The leverage shifts back to the infrastructure layer, not the model layer.
Now, let's look at Anthropic. With 30% of tokens, it commands 65.1% of the spending. This is not just a premium brand; this is the ability to charge for trust. Claude 3.5 Sonnet's pricing is higher than GPT-4o's, yet developers still choose it. Why? Because for complex, high-stakes tasks—like generating complex code for an entire application or analyzing long, nuanced legal documents—a single mistake is more costly than a thousand extra API calls. The developer is not buying tokens; they are buying reliability. They are buying the assurance that the code will work and the analysis will be correct. In an economy where an error costs millions, the price of the token is irrelevant. This is the value-density that closed-source models have mastered.
But here’s where I start to question the dominant narrative. We in the Web3 community have been saying "Code is law" for years, but we all know that code is just a foundation. It’s the human layer—the governance, the community, the trust—that makes it work. The same principle applies to AI. Anthropic’s high spending isn't just about model quality; it's about brand trust, security guarantees, and the implicit promise of a responsible AI provider. They have built a cultural moat, not just a technical one. And culture eats blockchain for breakfast. It always has.
If we look at the infrastructure layer, the picture becomes even more interesting. The rise of open-source models is not just a story about AI models; it's a story about the power shifting to the infrastructure layer. The underlying technology that enables open-source models to be cheap—like DeepSeek’s MLA attention and MoE architecture—is a technology edge, not a business strategy. The real value creation is happening in the optimization of inference, not just in the models themselves. This is a critical insight that many investors are missing. The massive surge in tokens isn't just a demand for intelligence; it's a demand for computation. It signals that the bottleneck isn't the model, but the chips and the power they consume.
As a community founder, I've seen this happen in the crypto space. When Bitcoin's value rose, the value didn't just go to the tokens; it went to the miners and the ASIC manufacturers. The same is happening in AI, where the value is flowing to the GPU providers and the energy grid. The model is the beautiful interface, but the real estate is in the underlying physical world.
This is where the contrarian angle emerges. The open-source rise is often framed as a victory for decentralization, but it's actually a precursor to a new form of centralization. The power isn't moving from OpenAI to the open-source community; it's moving from the model layer to the infrastructure layer. The new kings are the ones who own the hardware and the energy contracts. The open-source models are just the content that draws you into their platform.
Let me explain by using my experience in Estonia. In 2021, I curated a project called "Art for Access" that minted free NFTs for underrepresented artists. The token was free, and the art was the value. But the value of that token was only realized because we had a platform (like Vercel) that could host it, a community that could validate it, and a market that could exchange it. Without the platform, the token was worthless. The same is true in AI. The open-source model is the token, but the platform and the infrastructure are the value creators.
In the next 24 months, we're going to see a massive divergence between the "token rich" and the "value rich." The open-source ecosystem will continue to proliferate, but it will be a graveyard of platforms unless they find a way to monetize beyond the token. The price anchor set by open-source models will force closed-source providers to innovate even faster on the high end, creating a higher-value market. The gap between the top-tier and the commoditized will only widen.
Let’s talk about Google. DeepSeek overtaking Google is a significant signal, but it's not a signal about Chinese AI. It's a signal about Google's failure in developer relations. Google has some of the best AI research in the world. Yet, developers on Vercel are choosing a cheaper, less known Chinese model over Google's models. This is not just about performance; it's about experience. Google has always struggled with developer tools and API stability, a problem that plagues their ecosystem. They have a research culture that's strong in academia but often fails to translate into a user-friendly product.
And then there's the trust issue. We are building the future, together, but we have to be honest about the fact that we’re building on a shaky foundation of governance. When a model is open-sourced, who is responsible when it generates a harmful output? The developer who uses it? The open-source provider? Or the infrastructure that hosts it? In the closed-source world, this is clear: the provider is responsible. In the open-source world, it's a bureaucratic mess. This is the ethical gap that the industry is still trying to solve.
Let’s not forget that these numbers are from a single platform. Vercel is a deployment and hosting platform for web developers. The data is heavily skewed toward the front-end and Jamstack developer community. These developers are early adopters and are more cost-sensitive than, say, a Fortune 500 enterprise using Anthropic. The sample might be biased. But even with this bias, the data tells a powerful story: the economic weight of the model layer is not moving to open source.
So what's the takeaway? We need to stop celebrating the token count and start analyzing the value creation. We need to understand that the open-source AI revolution is real, but its economic reality is still based on a colonial model. The open-source models are the colony that exports raw resources—the tokens—to the closed-source empires who refine them into high-value products. As a community, we should not just be focused on the number of tokens we can generate, but on the value of the tokens we can extract. The goal is not to become the cheap provider, but to build the infrastructure and the platform that becomes the necessary stop for both the open and the closed.
The next time you see a chart showing the massive token usage of open-source models, don't be too quick to pop the champagne. Ask yourself, who is making the real profit? Where is the value stored? Is it in the model, or is it in the system that processes it? The blockchain maxim "Code is law" is flawed because it ignores the people who write the code and the people who use it. The same is true for AI. Code binds, but people break or build. The future isn't about open-source versus closed-source; it's about who controls the platform. And if we don't pay attention to the economics of the platform, we will be building the future for someone else's profit.
