The 62% Illusion: Why Token Share Is Not Economic Gravity

The numbers arrived without fanfare, buried in a developer platform's quarterly report. Over two months, open-source models on Vercel's infrastructure surged from 28.4% to 62% of all tokens consumed. The same data shows these models generated just 8.6% of total spending. Meanwhile, Anthropic—with 30% of token volume—captured 65.1% of the dollars. Read that again. The market is not rewarding usage. It is rewarding trust under pressure. I have spent twenty-five years watching systemic shifts, from ICO manias to DeFi yield cascades. This pattern feels familiar. It is the same divergence I flagged in the summer of 2020, when triple-digit APYs masked an absence of real revenue. Back then, the narrative died when the ledger bled. Today, the ledger is telling us something subtler: token share is a vanity metric, but economic value is a gravity well. The Vercel data deserves scrutiny. Developers on this platform lean toward web applications and front-end engineering. That skews the sample toward code generation, text classification, and content workflows—the long tail of AI tasks where cost sensitivity dominates. The 59% quarter-over-quarter growth in total token volume suggests price elasticity is real. When the marginal cost of a token drops to near zero, developers find new problems to solve. This is not substitution; it is creation. But the asymmetry is stark. Open-source models command volume; closed-source models command value. The unit economics differ by roughly 15x. This is not a temporary anomaly. It is the structural signature of a two-layer market forming in real time. Let me be precise about what the data does and does not show. DeepSeek overtaking Google as the second-largest model provider on this platform is a signal, not a verdict. It tells us that for a specific class of tasks, the performance-to-price ratio of open-source models has crossed a threshold. Developers are voting with their compute. But this is a vote for adequacy, not excellence. The 62% of tokens flowing through open models are likely concentrated in high-frequency, repeatable operations—autocomplete, extraction, summarization. The 8.6% of spending flowing to those same models reveals their ceiling: when a task carries real consequence, buyers still reach for the expensive option. This is where my skepticism sharpens. The conventional reading celebrates open-source ascendancy. The contrarian reading asks why Anthropic's share of spending is 7.5 times its token share. The answer is not just model quality. It is accountability. Enterprises do not pay premiums for intelligence alone; they pay for predictable behavior, for audit trails, for a vendor who answers when something breaks. Open-source models shift that burden to the application developer. Most teams are not equipped for that responsibility. I recall auditing smart contracts in 2017, when every project claimed decentralization while running on centralized nodes. The pattern repeats. We celebrate the open protocol while ignoring the custodial layer. The math was sound; the trust was the variable. In AI, the trust variable is now priced at a 15x premium. Consider the investment implications. The market is beginning to separate usage leaders from revenue leaders. DeepSeek's volume surge will attract attention, but its revenue capture is a fraction of its footprint. If its pricing strategy operates below cost—a plausible scenario given the economics—its growth is subsidized. That is not a moat; it is a burn rate. The valuation logic for open-source providers will increasingly resemble infrastructure utilities: high scale, thin margins, capital-intensive. The valuation logic for Anthropic and its peers will follow the software playbook: defensible margins, pricing power, customer lock-in. The prediction that closed models will capture 60-90% of economic value despite representing 15-25% of token volume is not a forecast. It is a description of the present, projected forward. The question is whether this equilibrium holds. History does not repeat; it rhymes in code. In 2020, DeFi protocols with astronomical usage but negative cash flows saw their valuations collapse when liquidity evaporated. Liquidity is not a floor; it is a horizon. The same principle applies here. Token volume without economic density is a mirage. There is a second-order effect worth watching. The price anchor set by open-source models is compressing the entire market's pricing envelope. Closed models must now justify premiums through demonstrated superiority on high-stakes tasks. This pushes them toward specialized domains—complex reasoning, multimodal integration, enterprise workflows—where the cost of error is high. Meanwhile, open models capture the elastic demand they created. Efficiency is the enemy of resilience, and the market is optimizing for efficiency at the edge while preserving resilience at the core. One more signal deserves attention. The token mix on Vercel reflects developer sentiment, but it does not reflect enterprise procurement. The gap between these two worlds is where the next disruption will emerge. When enterprise workloads begin migrating toward open models—not for cost, but for data sovereignty and customization—the value distribution will shift again. That migration will be slower, but it will be stickier. For now, the data paints a clear picture. The open-source wave is real, but its economic gravity remains weak. The smart money is not chasing token volume; it is tracking dollar density. Correlation is the smoke; divergence is the fire. Watch the divergence between usage and spending. It will tell you where the next cycle turns. I have been through enough cycles to know that the narrative dies when the ledger bleeds. The ledger here is not bleeding—it is bifurcating. The question for investors is not which model wins the token war. It is which model wins the margin war. And that war has only one metric that matters: the price of trust when the system is under stress.