The numbers arrived quietly, buried in a developer platform's quarterly transparency report. Over the past two months, open-source models on Vercel's infrastructure surged from 28.4% to 62% of all token consumption. DeepSeek, a Chinese open-weight model, overtook Google to become the second-largest model provider on the platform. Yet here is the dissonance that should trouble every investor, every builder, every observer of the AI economy: those 62% of tokens generated only 8.6% of total spending. Anthropic, with 30% of token volume, captured 65.1% of the dollars.
My eye is on the horizon, not the hourly candle. And from that vantage point, this divergence is not a statistical curiosity. It is a structural signal about where value actually accumulates in the AI stack — and by extension, where it will accumulate in the adjacent crypto and digital asset markets that increasingly settle AI's economic flows.
The Context: What Vercel Actually Measures
Vercel is not a neutral observer. It is a deployment platform favored by web developers, front-end engineers, and application builders — the people who ship AI features into production quickly. Its data reflects real usage patterns, not benchmark leaderboards. When developers integrate a model into their application, every API call, every completion, every inference registers as a token. This is the closest thing we have to a transparent ledger of actual AI consumption.
The platform's total token volume grew 59% month-over-month. That alone tells us something important: the marginal cost of AI inference is dropping fast enough to create entirely new demand. Tasks that were previously too expensive to automate — content classification, code completion, data extraction — are now economically viable. This is the price elasticity effect in action, and it mirrors what we saw in the early days of cloud computing when storage costs collapsed and suddenly every startup became a data company.
But the composition of that growth reveals a more complex story. Open-source models are winning the volume game. DeepSeek's ascent past Google is particularly striking — not because DeepSeek is necessarily a better model, but because developers on Vercel are voting with their API calls, and they are choosing the cheaper, open option at scale.
The Core: Value Density vs. Volume
Let me be precise about what the data shows, because the implications are not obvious at first glance.
Open-source models now handle 62% of token traffic. But they capture only 8.6% of spending. That means the unit economics of open-source inference are roughly fifteen times weaker than the closed-source average. Anthropic, by contrast, commands a token price more than double the market average — and developers are paying it willingly.
This is not a failure of open source. It is a division of labor. The market is sorting itself into two distinct tiers: high-frequency, low-complexity tasks that demand scale and low cost, and high-value, complex reasoning tasks that demand quality and reliability. The first tier belongs to open models. The second belongs to Anthropic and, to a lesser extent, OpenAI.
The 59% growth in total token volume is the real story here. Open-source models are not merely cannibalizing closed-source market share — they are expanding the entire market. Developers are using more tokens because they can afford to. Tasks that were previously deferred are now being automated. This is the classic Jevons paradox: as the cost of a resource falls, consumption rises — sometimes enough to increase total resource expenditure.
But here is the uncomfortable question: what is the economic value of a token? Not all tokens are created equal. A token spent on a complex legal document analysis is worth more than a token spent on autocompleting a function name. The Vercel data suggests that open-source tokens are disproportionately flowing into low-value, high-volume tasks. The 62% figure may be impressive, but it is a measure of activity, not of value creation.
The Contrarian Angle: The Open-Source Victory Narrative Is Premature
There is a seductive story being told right now: open source is winning, the closed models are doomed, and the future belongs to democratized AI. The Vercel data appears to support this — until you look at the spending side.
Here is the contrarian truth: the economic center of gravity in AI remains firmly with closed models. Anthropic's 65.1% spending share on just 30% of tokens is not an anomaly. It is a reflection of what enterprises actually pay for: reliability, safety, accountability, and the ability to handle complex, high-stakes reasoning. When a hospital deploys an AI system for clinical decision support, it is not choosing the cheapest model. It is choosing the model that can be held accountable.
This is where I see a parallel to the crypto markets I navigate daily. In 2021, we saw a similar divergence — the narrative that decentralized protocols would displace traditional finance. The token volumes surged, the usage metrics looked impressive, and yet the economic value remained concentrated in a handful of platforms with real revenue. The bust was not an end, but a necessary pruning. The same pruning is now underway in AI.
DeepSeek's rise is real, but it is a volume story, not a value story. The company may be processing more tokens than Google, but its revenue per token is a fraction of what Google's Gemini commands. This is the classic infrastructure trap: you can win the usage war and lose the profit war simultaneously. I have seen this play out in Layer 2 networks, where dozens of chains boast impressive transaction counts but struggle to generate sustainable fees.
There is also a sustainability question that nobody is asking. DeepSeek's pricing strategy — reportedly below cost — is a market-share play, not a business model. At some point, the subsidies end. When they do, the token distribution will shift again, and the developers who built their applications on ultra-cheap inference will face a painful migration.
The Takeaway: Positioning for the Value Migration
So where does this leave us? The Vercel data tells me three things with reasonable confidence.
First, the AI market is bifurcating into a two-tier structure: open models for scale, closed models for value. This is not a temporary state — it is the equilibrium the market is moving toward. Second, the total addressable market for AI inference is expanding faster than anyone predicted, driven by falling costs. Third, the companies that will capture outsized economic value are those that own the high-value tier — or that build infrastructure enabling the low-value tier to scale profitably.
For those of us watching from the digital asset side, the implications are clear. The AI token economy — whether settled on-chain or off — will increasingly reflect this value density divergence. Projects that position themselves as infrastructure for high-volume, low-cost inference will need massive scale to justify their valuations. Projects that enable verifiable, high-value AI services will command premium multiples.
My eye is on the horizon, not the hourly candle. The next twelve months will tell us whether the open-source surge is a durable structural shift or a subsidy-driven mirage. Either way, the data is telling us something important: in AI, as in crypto, usage is not the same as value. The market is learning this lesson again — and the investors who understand the difference will be the ones who survive the pruning.