The Token Share Mirage: Why 62% Usage and 8.6% Spend Means the AI Model Market Is Splitting in Two

Two months. That's all it took for the open-source model share on Vercel's platform to flip from 28.4% to 62% of all tokens consumed. The same dataset shows DeepSeek — an open-weight model from a Chinese lab — overtaking Google as the second-largest model provider by token volume. On the surface, this reads like the open-source revolution finally arrived. Dig one layer deeper and the numbers tell a different story: those 62% of tokens generated only 8.6% of total spending. Meanwhile, Anthropic's 30% token share captured 65.1% of the dollars. That's not a revolution. That's a market bifurcating into two distinct economies with different rules, different margins, and different survival trajectories. Tracing the noise floor to find the alpha signal — the signal here is not that open source is winning. It's that the AI model market is splitting into a high-volume commodity layer and a high-value intelligence layer, and the two are diverging faster than most investors realize. Let me establish the context before I pull this apart. Vercel is a deployment platform for web applications and front-end infrastructure. Its AI gateway routes developer traffic to various model providers, and the company periodically publishes aggregate usage data. This is not a comprehensive view of the entire AI market — it skews toward web developers, SaaS builders, and application-layer engineers. But that's precisely why the data matters. These are the people actually shipping AI products into production, not running benchmark suites in research labs. When a developer on Vercel routes a request to DeepSeek instead of Google, that's a real production decision with real cost implications. The platform's total token volume grew 59% month-over-month, which tells me the price elasticity of AI inference is far higher than the market consensus assumes. Lower prices are not just shifting share — they're creating entirely new demand. Tasks that were previously uneconomical to run through an LLM are now viable. That's the kind of demand creation that doesn't show up in benchmark leaderboards. The core data point that deserves scrutiny is the divergence between token share and spending share. Open-source models — led by DeepSeek but including Llama variants, Qwen, and others — now move 62% of the tokens on Vercel's platform. Their share of spending: 8.6%. That's a unit economics gap of roughly 15x. Anthropic, by contrast, commands 30% of tokens but 65.1% of spending — a unit price roughly 2.2x the market average. The arithmetic is brutal and clarifying. Open-source models are being used for high-frequency, low-complexity tasks: code completion, text classification, information extraction, summarization. The kind of work where a 90% correct answer at 1/15th the cost beats a 98% correct answer at full price. Closed-source models, particularly Anthropic's Claude family, are being reserved for complex reasoning, multi-step agentic workflows, and high-stakes generation where failure is expensive. This is not a temporary arbitrage. It's a structural division of labor that will harden over time. I've spent enough years auditing protocol economics to recognize this pattern. It's the same trajectory we saw in blockchain infrastructure: base layer commoditization and application layer value capture. In 2020, I was stress-testing Curve Finance's invariant calculations with a $15,000 bot, mapping out slippage mechanisms to find timing attack vectors. The lesson I took from that exercise applies directly here: when a technology layer commoditizes, the value doesn't disappear — it migrates upward. The same thing is happening in AI models. The token itself is becoming a commodity. The intelligence applied to the token is where the margin lives. Code does not lie, but it does hide — and what's hidden in this data is that the open-source token surge is largely a story about cost-driven task delegation, not capability parity. Let me be more specific about what the 62% figure actually represents. Based on my experience deploying and benchmarking open-weight models in production environments, the token distribution is almost certainly concentrated in a few task categories. Code generation and autocomplete — where DeepSeek's coding models have genuinely competitive performance — likely account for a disproportionate share. Data extraction and structured output tasks, where the cost savings dominate any quality delta, are another major bucket. And a significant portion is probably what I'd call 'trial tokens' — developers testing whether open models can handle a task before deciding whether to escalate to a premium provider. The 59% month-over-month growth in total token volume supports this. When you drop the price of inference by an order of magnitude, developers start throwing tokens at problems they previously solved with regex or hardcoded logic. Some of that is productive. Some of it is just cheap compute being wasted. But all of it inflates the open-source token share without adding proportional economic value. The contrarian angle here is uncomfortable for the open-source evangelists. The 62% token share is being celebrated as proof that open models have 'won.' What it actually proves is that open models have been relegated to the low-margin, high-volume end of the market. The spending data — 8.6% — is the more honest metric. It tells you where the economic value actually resides. And it's not with the open models. This is the same trap we saw in the Layer2 narrative: projects touting transaction counts and TVL while the actual revenue and sustainable usage tell a different story. I've been saying for two years that most Layer2 sequencers are centralized nodes wrapped in decentralized marketing. The token share data from Vercel is the AI equivalent — usage metrics that flatter the open-source ecosystem while the revenue concentrates elsewhere. Redundancy is the enemy of scalability, and in this case, the redundancy is the narrative that open-source adoption equals open-source profitability. There's a second contrarian layer worth examining: the DeepSeek overtaking Google data point. On the surface, this looks like a Chinese open-source model beating an American closed-source giant. The reality is more nuanced. DeepSeek's pricing strategy is aggressive — some would say predatory. The company has reportedly priced its API below marginal cost in certain tiers, a classic subsidize-to-capture-market play. When you're giving away inference at a loss, token share is easy to buy. The question that matters is sustainability. Can DeepSeek maintain this pricing once the subsidy runs out? Or will it be forced to raise prices, ceding the token share back to competitors? I've seen this movie before. In the crypto bear market of 2022, protocols burned through treasury reserves to subsidize yields and inflate TVL. When the subsidies stopped, the TVL evaporated. The same dynamic is playing out in AI inference pricing. The market is rewarding the subsidy, not the underlying capability advantage. That said, I don't want to dismiss DeepSeek entirely. The engineering behind their models is genuinely impressive. Their Mixture-of-Experts architecture achieves competitive performance at a fraction of the compute cost of dense models. The inference optimization work they've done — including their MLA (Multi-head Latent Attention) mechanism — is legitimately innovative. This is not a case of pure subsidy masking inferior technology. The models are good. The question is whether they're good enough to command premium pricing once the subsidy window closes. Based on my benchmarks, the gap between DeepSeek's best models and Anthropic's frontier models on complex reasoning tasks remains significant. On routine tasks, the gap is negligible. That's exactly the profile of a commodity provider: good enough for the mass market, not good enough for the premium segment. The investment implications are substantial. The market is currently valuing AI model providers on a mix of revenue growth and narrative momentum. The Vercel data suggests a more nuanced valuation framework is needed. For closed-source providers like Anthropic, the 65.1% spending share on 30% token share validates a premium valuation — the market is demonstrably paying for quality. For open-source providers, the valuation logic is closer to infrastructure: high volume, low margin, scale-dependent. That's a very different multiple. I've been through enough market cycles to know that investors eventually figure this out. The correction may be brutal for open-source model companies that have been valued on usage metrics rather than revenue quality. Volatility is the price of entry, not the exit — and the volatility here will come when the market recalibrates its expectations for open-source model monetization. There's also a geopolitical dimension that the raw data obscures. DeepSeek's rise on Vercel's platform means Chinese open-weight models are now embedded in Western developer infrastructure. This raises questions about data governance, model auditing, and supply chain security that the crypto industry has already grappled with in the context of cross-border protocol dependencies. Logic gates are the new legal contracts — and the logic gates in DeepSeek's models are not subject to Western regulatory oversight. For most developers, this is a non-issue. The models are open-weight, auditable, and running in controlled environments. But for enterprise deployments with strict compliance requirements, the provenance of the model weights becomes a procurement issue. This is a friction point that could limit open-source model adoption in regulated industries, further entrenching the closed-source premium segment. Let me also address the platform bias question, because it matters for how we interpret this data. Vercel's user base skews toward web developers, front-end engineers, and indie hackers. This is not the enterprise AI procurement market. The token distribution on Vercel likely overrepresents code generation, content creation, and lightweight automation tasks — precisely the categories where open-source models are most competitive. In enterprise settings — legal document analysis, financial modeling, medical reasoning — the closed-source premium is likely even more pronounced than the Vercel data suggests. The 62/8.6 split might actually understate the economic concentration in the broader market. If anything, the real-world divergence between token share and value capture is probably wider than what Vercel reports. The forward-looking question is whether this bifurcation is stable or transitional. My read is that it's structural. The cost of training frontier models continues to escalate, pushing closed-source providers toward higher-value, higher-complexity tasks where they can justify premium pricing. Meanwhile, open-source models benefit from a virtuous cycle of community contributions, quantization improvements, and inference optimization that steadily narrows the capability gap on routine tasks. The two tracks are diverging, not converging. The market is settling into a stable equilibrium where open models handle the long tail of high-frequency, low-complexity inference, and closed models own the high-value, high-stakes intelligence layer. The economic value will continue to concentrate in the closed-source segment — the article's prediction of 60-90% value capture by closed models looks conservative to me. For developers and builders, the strategic implication is clear: architect your systems to use both layers. Route routine tasks to open models to control costs. Reserve premium models for the tasks where quality differential actually matters. The teams that treat this as a portfolio optimization problem — rather than a religious war between open and closed — will have a structural cost advantage. Build first, ask questions later. The data is already telling you where the margin lives. I've been in this industry long enough to watch narratives form, inflate, and collapse. The open-source token share surge is a real phenomenon with real momentum. But the spending data is the ground truth. And the ground truth says the economic center of gravity in AI models remains firmly with the closed-source providers. The open-source ecosystem is building critical infrastructure — the commodity layer that makes AI accessible at scale. That's valuable. But it's not the same as owning the intelligence layer. The market will eventually price this distinction correctly. When it does, the token share mirage will fade, and the value concentration will become impossible to ignore. The question is whether you've positioned yourself on the right side of that divergence before the market figures it out.