History does not repeat, but it often rhymes in the code. For six years, I have watched liquidity flow through digital ledgers—how value moves when trust is scarce, and how volume rarely translates into economic weight. The pattern emerging from Vercel's latest developer data tells a story that feels familiar to anyone who has studied the gap between adoption and settlement.
Over the past two months, something shifted in the AI development ecosystem. Vercel's production telemetry reported open-source model token consumption surging from 28.4 percent to 62 percent of all tokens processed. DeepSeek, an open-source architecture out of China, surpassed Google to become the second-largest model provider on the platform—a milestone that would have been unthinkable a year ago.
But here is where the data gets complicated. Those open-source models, which now carry more than three-fifths of the platform's token traffic, generated only 8.6 percent of the spending. Meanwhile, Anthropic, with roughly 30 percent of token volume, captured 65.1 percent of all expenditure.
The ledger remembers what the algorithm forgets. This is one of those moments where the underlying economics tells a story that the raw usage numbers are actively hiding.
Context: What the Vercel Data Actually Measures
Let me establish the baseline before we go deeper. Vercel is a developer platform used extensively by front-end engineers and AI application builders. Its data pipeline tracks millions of inference calls across thousands of production projects—not benchmark tests, not research labs, but real code running in production environments.
This distinction matters for three reasons. First, production data reflects actual developer behavior rather than experimental preference. When a developer chooses a model, they are making a practical decision about cost, latency, and reliability—not just reading a leaderboard. Second, Vercel's user base skews toward web application development, which overrepresents certain workloads like code completion, content generation, and information extraction, while underrepresenting enterprise-grade complex reasoning. Third, the platform captures the full lifecycle of model selection—the switch from one provider to another happens in real time, and the data reflects that migration.
That said, the biases cut both ways. Vercel may overstate open-source adoption in these long-tail scenarios, but it simultaneously understates the dominance of closed-source models in high-stakes environments where quality is paramount and cost is secondary.
With that context, the numbers deserve a careful reading.
The Open-Source Surge: Volume Without Value
The headline figure is the token shift: from 28.4 percent to 62 percent in roughly sixty days. That is not an incremental migration; it is a phase transition. Developers are not gradually experimenting with open-source models—they are actively switching their default providers.
I saw the same pattern in the 2020 DeFi summer, when total value locked moved from a handful of audited protocols into a wide range of experimental contracts. The volume numbers looked decisive at the time. But the value creation did not follow the volume. The same dynamic is emerging here.
What the volume growth does not show is the nature of the workloads. The 62 percent token share is almost certainly concentrated in repetitive, deterministic tasks: code autocompletion, text classification, sentiment analysis, data extraction. These are tasks where "good enough" is genuinely enough, where the model's output quality does not justify paying a premium for frontier capability.
Consider the unit economics. If open-source models process 62 percent of tokens but generate only 8.6 percent of spending, their per-token revenue is roughly one-fifteenth of the closed-source baseline. That is not necessarily a bad business—the transaction volume can be huge—but it does reflect a fundamental difference in economic density.
There is a subtle second-order effect hidden here. Total token volume grew 59 percent quarter over quarter. That growth is the price elasticity of demand in action. The low cost of open-source inference is creating entirely new demand—developers are running tasks they previously deferred because the cost did not justify the benefit. At a fraction of a cent per call, ten million classification requests become feasible.
This mirrors what we observed in the L2 scaling boom. When transaction costs dropped from fifty dollars to ten, usage exploded, but the average economic value settled per transaction collapsed. The system grew in scale without growing in economic significance.
DeepSeek's Ascent: Sustained Surge or Subsidized Spike?
The most striking data point is DeepSeek's rise over Google. That is a genuine milestone—an open-source model from a Chinese lab has displaced a major American closed-source provider on a prominent developer platform. But we need to be careful about what the milestone represents.
DeepSeek's pricing is a fraction of Google's Gemini. When you price your model at one-tenth of the leading alternative, you do not need to be superior—you only need to be approximately equal on the tasks that matter. For a large class of developer workflows, DeepSeek is likely good enough. The price differential alone is sufficient to drive massive volume.
This creates a price elasticity effect. At one-tenth of the cost, a developer might run fifteen times more inference calls. The result is a volume surge that does not correlate with a quality breakthrough.
The sustainability question is critical. Is DeepSeek's pricing sustainable, or is it being subsidized? The token volume growth suggests the latter—you cannot run a business at one-tenth of the market rate without either exceptional efficiency or external subsidy. In the crypto world, we have seen this pattern before: a protocol with high usage and negative unit economics is not a sustainable business, it is a burn rate.
Trust is borrowed; trust is never owned. When a model provider prices below cost to acquire market share, they are purchasing trust at a discount. The moment prices rise to sustainable levels—which they will if the capital cycle tightens—the trust will not be returned. The market will simply move to the next cheap alternative.
Anthropic's Premium: The Price of Reliability
The opposite extreme of the data is Anthropic. Thirty percent of token volume, 65.1 percent of spending. The unit price per token is roughly two to three times the market average. This is not brand loyalty; this is the premium for reliability.
Anthropic has positioned itself as the model for high-stakes tasks—complex reasoning, legal analysis, financial modeling, code generation with strict correctness requirements. In these domains, the cost of a model error exceeds the cost of the token. A developer running a legal document analysis at 10x the cost of an open-source alternative still comes out ahead if the accuracy is meaningfully better.
This mirrors the L1/L2 dynamic in blockchain. Ethereum processes a fraction of the transaction volume of sidechains and L2s, yet it captures the majority of the economic value. Because the base layer is the trust anchor. The market pays for the assurance that the asset will hold.
In the AI model economy, Anthropic is the trust anchor. Developers are not paying for raw capability; they are paying for reliability, consistency, and the reduced probability of catastrophic error.
The Double-Layer Structure of the AI Economy
The emerging picture is a two-layer market structure. The low-cost, high-volume layer is dominated by open-source models handling repetitive, deterministic, cost-sensitive tasks. The high-value, low-volume layer is dominated by frontier closed-source models handling complex reasoning, high-stakes decision-making, and specialized workloads.
This is not a winner-take-all scenario. It is a differentiated market where each layer serves a different purpose and captures a different share of value.
The prediction that closed-source models will eventually capture 60 to 90 percent of the economic value—even while open-source processes 75 to 85 percent of the tokens—is consistent with this structure. The value does not follow the volume; it follows the marginal utility of the task.
If the marginal utility of a complex, high-stakes analysis is fifty times the marginal utility of a code snippet, then the model handling the analysis will capture fifty times more value per token.
What This Means for Investors
From my perspective as a digital asset fund manager, the Vercel data offers a critical lesson: usage metrics are not revenue metrics. The token volume of an open-source model provider says little about its economic potential. A model with massive usage but no pricing power is a commodity provider—it captures volume but not value.
In the crypto world, we have seen this many times. A token with high on-chain activity but low revenue eventually gets priced on revenue, not on activity. The same logic applies to AI models.
The market's pricing of Anthropic and OpenAI reflects this understanding. Their valuations are justified by their ability to capture economic value, not by token volume. The open-source providers may have the user base, but they lack the pricing power.
Safety is the only yield that compounds over time. In the AI market, as in crypto, the model that earns trust and demonstrates reliability is the model that captures the value. The model that offers speed without trust will capture the volume but not the value.
The Contrarian Angle: Open Source Is Not Winning the Value War
The narrative among developers is that open-source is winning. The Vercel data is presented as proof. But this interpretation is deeply flawed.
Open source is winning the volume war. It is not winning the value war. The 62 percent token share reflects real adoption, but the 8.6 percent spend share is the true economic signal. Open-source models are the infrastructure layer of AI—important, scalable, but capturing minimal margins.
For investors, this means the value opportunity in AI remains concentrated in the closed-source frontier models. The open-source model will generate volume but not value. The premium lies in the trust layer, not the infrastructure layer.
The counterargument—that DeepSeek will eventually translate volume into value as its capabilities improve—is a bet on qualitative leap, not a confirmation of current economics.
The ledger remembers what the algorithm forgets. The Vercel data is a ledger, a record of actual production behavior. It shows that the economic value of AI is concentrated in a narrow set of models that have earned trust. The open-source movement has democratized access but has not yet captured the economic premium. As the market matures, the divide between usage and value will become the defining structure of the AI economy. The question is whether the open-source layer can build the trust that the premium layer has already earned.
