The market does not hate you; it ignores you. And right now, the AI model market is ignoring the single most important signal in its own data: open-source models now process 62% of all tokens on Vercel's platform, yet they capture only 8.6% of the spending. That is not a rounding error. That is a structural revelation about where value actually accumulates in an economy increasingly built on machine-generated output.
I have spent the last nine years watching value migrate across cryptographic and computational substrates. I have audited ICO smart contracts that promised decentralized everything and delivered centralized nothing. I have stress-tested DeFi lending protocols that collapsed under the weight of their own recursive yield assumptions. And I have learned one immutable lesson: when usage and value diverge by an order of magnitude, the market is telling you something about the underlying architecture of the system. The Vercel data is such a signal. It deserves a proper audit.
The Context: A Platform as a Mirror
Vercel is not the AI industry. It is a deployment platform favored by web developers, front-end engineers, and application builders. But that is precisely why its data matters. Vercel's user base represents the pragmatic middle of the AI economy: developers who are not AI researchers, not enterprise procurement officers, but people who need models to work, cheaply and reliably, inside real products. When this demographic shifts its token consumption patterns, it is not a theoretical debate about model quality. It is a revealed preference, expressed in production traffic.
The numbers, as reported, are stark. Open-source model token share grew from 28.4% to 62% in roughly two months. DeepSeek, a Chinese open-source model family, surpassed Google to become the second-largest model provider on the platform by token volume. Anthropic, meanwhile, commands 30% of token volume but 65.1% of spending. The asymmetry is not merely interesting. It is the entire story.
The Core: Token Volume Is Not Value Capture
Let me be precise about what these numbers mean, because the surface narrative is dangerously misleading.
Open-source models now handle the majority of inference requests on Vercel. That is a genuine milestone. It means that for a significant class of tasks, developers have decided that open-weight models are good enough, and the cost differential is decisive. The price gap is not subtle. If open-source models account for 62% of tokens but only 8.6% of spending, the unit economics are roughly 15:1 in favor of closed models. That is not a marginal advantage. That is a different pricing universe.
But here is what the token share does not tell you. It does not tell you what those tokens are doing. My experience auditing production systems tells me that high-volume, low-cost token consumption clusters around specific task categories: code completion, text classification, information extraction, summarization, boilerplate generation. These are the commodity tasks of the AI economy. They are necessary, but they are not where the economic surplus is created. They are the equivalent of payment rails in the crypto world: high transaction volume, low margin per transaction, and a business model that only works at massive scale.
The 59% quarter-over-quarter growth in total token volume is the more interesting data point. It suggests that the low price of open-source models is not just substituting for closed-model usage. It is creating new demand. Developers are now using models for tasks that were previously uneconomical. This is the price elasticity effect, and it is real. But it also means that the token volume growth is partly a function of the models being cheap, not necessarily being good. The algorithm optimizes for survival, not for you. And the survival strategy of open-source models is to flood the market with low-cost inference.
DeepSeek's surpassing of Google is a signal worth unpacking. On the surface, it looks like a quality story: an open-source model beating a closed model in developer adoption. But my training as a cryptographer makes me suspicious of surface narratives. The more likely explanation is a combination of factors. DeepSeek's pricing is aggressive, possibly below cost. Its architecture, reportedly using mixture-of-experts techniques, allows for efficient inference. And its performance on common benchmarks, while not leading, is sufficient for the commodity tasks that dominate Vercel's traffic. The result is a model that is good enough and dramatically cheaper. That is a volume play, not a value play.
This is where the crypto analogy becomes unavoidable. In decentralized finance, we learned that total value locked is a vanity metric. What matters is the economic value extracted from that liquidity. A liquidity pool with $1 billion in TVL that generates $10,000 in fees is a monument to inefficiency. The same logic applies here. Token volume is the TVL of the AI economy. Spending is the fee generation. And the divergence between the two is the single most important metric for understanding where this market is heading.
The liquidity pool is a mirror, not a vault. It reflects the flows, but it does not hold the value. Vercel's token data is the mirror. The spending data is the vault. And the vault is firmly controlled by closed models.
The Contrarian Angle: The Commoditization Trap
Here is where I diverge from the prevailing narrative. The mainstream interpretation of this data is that open-source models are winning, and closed models are losing. I think that is wrong. What the data actually shows is that the AI model market is bifurcating into two distinct economies with different rules, different margins, and different value trajectories.
The first economy is the commodity inference market. This is where open-source models dominate. It is characterized by high volume, low unit price, thin margins, and intense competition. In this market, the winners will be those with the lowest cost structure, the most efficient inference infrastructure, and the ability to scale. This is not a software company business model. This is a utility business model. Think electricity generation, not enterprise software. The valuations in this segment will be based on infrastructure multiples, not growth multiples.
The second economy is the high-value reasoning market. This is where Anthropic currently dominates, and where OpenAI is fighting to maintain relevance. This market is characterized by lower volume, dramatically higher unit prices, and a premium for quality, reliability, and safety. The tasks in this segment are complex reasoning, creative work, agentic decision-making, and enterprise-grade automation. The margins are software-like, and the valuations will reflect that.
The prediction that closed models will eventually account for only 15-25% of token volume but 60-90% of economic value is not a prediction of decline. It is a prediction of structural separation. The closed models are not losing. They are being pushed upmarket, into the segment where their capabilities justify their prices. And the open-source models are not winning. They are being pushed downmarket, into the segment where their prices justify their capabilities.
This is the commoditization trap. It is the same trap that every technology faces when it matures. The hardware industry went through it. The software industry went through it. The cloud industry is going through it. And now the AI model industry is going through it. The question is not whether open-source models will capture more token share. They will. The question is whether that token share will ever translate into meaningful economic value. Based on the current data, the answer is no.
There is a second contrarian observation worth making. The Vercel data is biased. Vercel's user base skews toward web development, front-end engineering, and application prototyping. These are precisely the use cases where open-source models are most competitive. The data does not capture enterprise workflows, regulated industries, or mission-critical deployments where closed models retain a significant advantage. If the data included those segments, the spending share of closed models would likely be even higher, and the token share of open-source models would likely be lower. The 62% figure is a ceiling, not a floor.
The Deeper Structure: What This Means for the AI Economy
Let me step back and map this onto the broader economic landscape, because the implications extend far beyond Vercel's platform.
First, the pricing anchor is shifting. Open-source models are setting a new price floor for AI inference. This is deflationary for the entire market. Closed models will find it increasingly difficult to justify premium pricing for commodity tasks. Their only defense is differentiation: models that can do things open-source models cannot. This is a sustainable strategy, but it requires continuous investment in frontier capabilities. The moat is not the model. The moat is the ability to push the frontier.
Second, the application layer is being democratized. The low cost of open-source models is lowering the barrier to entry for AI application development. This will lead to a wave of new applications, many of which will be undifferentiated and short-lived. But some will find product-market fit and create real value. The application layer is where the value creation will happen, not the model layer. This is the same pattern we saw in crypto: the infrastructure gets commoditized, and the value migrates to the applications built on top.
Third, the infrastructure providers are the hidden winners. Whether models are open or closed, they all run on compute. The growth in token volume, driven by the price elasticity of open-source models, is a direct tailwind for GPU providers, cloud platforms, and data center operators. The AI economy is becoming more compute-intensive, not less. The value capture is shifting from the model layer to the infrastructure layer. This is the equivalent of the crypto market discovering that the real money is in selling picks and shovels, not in mining the gold.
Fourth, the geopolitical dimension cannot be ignored. DeepSeek's rise on Vercel is not just a commercial story. It is a signal that Chinese AI models are competitive in international developer markets. This has implications for data governance, model control, and the broader technology competition between the US and China. Regulation is the lagging indicator of chaos. The chaos here is the rapid diffusion of AI capabilities across borders, and the regulation will follow, likely in the form of export controls, data localization requirements, and model governance frameworks. The open-source model ecosystem will be the hardest to regulate, which is precisely why it will attract the most regulatory attention.
The Investment Implications
For investors, the Vercel data provides a useful framework for evaluating AI companies. The key metric is not token volume or model quality as measured by benchmarks. The key metric is the unit economic value of the tokens a model processes. Anthropic's 30% token share generating 65.1% of spending means its tokens are worth roughly 2.2 times the market average. That is the metric that matters.
OpenAI is in the most precarious position. It is neither the cheapest option nor the highest-quality option. It is caught in the middle, with a brand that commands attention but a product that faces pressure from both sides. The market is beginning to price this in, and the recent valuation adjustments reflect a growing awareness that OpenAI's position is less defensible than its narrative suggests.
DeepSeek's situation is more complex. Its token volume growth is impressive, but its revenue generation is likely minimal. If it is operating below cost, it is burning capital to acquire market share. This is a classic growth-at-all-costs strategy, and it works only if the market share can eventually be monetized. In a commodity market, monetization is difficult. The exit liquidity is just another person's thesis. The thesis here is that DeepSeek will eventually raise prices or find higher-value use cases. Neither is guaranteed.
The Takeaway: Value Migrates, It Does Not Disappear
The Vercel data is a snapshot of a market in transition. The transition is not from closed to open. It is from a unified market to a bifurcated one. The commodity segment will be dominated by open-source models, with thin margins and infrastructure-like valuations. The premium segment will be dominated by a few closed models, with software-like margins and growth valuations. The infrastructure layer will capture value from both segments.
The question for the next cycle is not which model will win. The question is which layer of the stack will capture the most value. My analysis suggests the answer is the infrastructure layer, followed by the application layer, with the model layer being increasingly commoditized. The models are becoming the new TCP/IP: essential, ubiquitous, and nearly impossible to monetize directly. The value is moving up the stack.
I have seen this pattern before. In 2017, I audited ICO contracts that promised to decentralize everything. The infrastructure was built, the tokens were issued, and the value migrated to the applications that actually solved user problems. The same pattern is playing out in AI. The models are the new infrastructure. The applications are the new value layer. And the investors who understand this will be positioned for the next cycle.
The algorithm optimizes for survival, not for you. The open-source models are optimizing for market share. The closed models are optimizing for margin. The infrastructure providers are optimizing for utilization. And the developers are optimizing for cost. Each is rational. Each is pursuing its own survival strategy. The market is simply the aggregate of these strategies, and the Vercel data is the aggregate signal.
Read it carefully. The token share is the noise. The spending share is the signal. And the divergence between them is the opportunity.