Open-Source Models Just Took 62% of Token Traffic. Here's Why the AI Value War Is Only Beginning.

The numbers hit my terminal like a block reward confirmation. Vercel's CEO drops platform data: open-source models now chew 62% of all tokens processed through their AI Gateway. Just months ago, that figure sat at 28.4%. The migration is not incremental. It is a stampede. But here is the kicker. Those same open-source models account for only 8.6% of total spend. Closed-source giants—OpenAI, Anthropic, Google—still command 91.4% of the dollars flowing through the same pipes. Volume spikes lie; liquidity flows tell the truth. And the truth here is a market bifurcating at breakneck speed. Let's get the context straight. Vercel is not a random sample. It is the deployment layer for a massive chunk of the modern web. Frontend developers, indie hackers, and increasingly serious product teams route their AI calls through its gateway. That makes this data a live biopsy of developer behavior, not a theoretical survey. When DeepSeek—a Chinese open-weight lab—surpasses Google's Gemini in token consumption on this platform, the competitive map has shifted under our feet. Speed is safety when the exploit is already live. And the exploit here is the assumption that "open source" means "inferior." That narrative is dead. The chart doesn't lie, and neither does the token count. Core data breakdown, forensic style. Open models at 62% token share, 8.6% spend. Closed models at 38% token share, 91.4% spend. Do the math. The implied unit price for open tokens is roughly one-fourteenth that of closed. That is not a small discount. That is a pricing abyss. DeepSeek's MoE architecture and MLA attention mechanism crushed inference costs to a fraction of GPT-4o or Claude 3.5. Developers are rational actors. They did not switch because they hate paying for quality. They switched because the quality bar for their use cases was met at a fraction of the cost. What use cases? Code completion, refactoring, documentation generation, test scaffolding. The boring, high-volume, cost-sensitive workload. This is the long tail of AI development. And open models own it now. Anthropic's position is the flip side of the same coin. 30% of tokens, 65.1% of spend. Claude is the premium pick for complex, high-stakes work—agentic workflows, intricate code generation, long-context analysis. The kind of task where a hallucination costs more than the API bill. Enterprise clients pay for that reliability. That is a value-density play, not a volume play. Now, the contrarian angle. The one nobody in the open-source celebration wants to hear. We don't actually know what the 62% token share is worth. Token counts are not value counts. A million tokens of embedding calls for a vector database cost pennies and create marginal utility. A thousand tokens of intricate, multi-step reasoning might close a funding round. The open-source surge could be inflated by low-value, high-frequency tasks that were always price-sensitive. That would mean the "open-source victory" is partially a victory in the minor leagues. Second blind spot: the 8.6% spend figure only captures API costs. It excludes self-hosted GPU clusters, Kubernetes overhead, and the engineering salaries needed to keep a fine-tuned open model running in production. Total cost of ownership for a serious DeepSeek deployment can rival a closed API subscription. The chart doesn't show that line item. Third, and this is the one that keeps me up at night: sample bias. Vercel's user base skews toward web developers and early-stage product builders. That is not the Fortune 500 procurement office. Enterprise AI spend—the real money—still flows through Microsoft, AWS, and bespoke vendor deals. The 62% number measures the frontier of developer experimentation, not the core of enterprise lock-in. My experience in the 2024 ETF flow analysis taught me to distrust headline ratios. The "Silent Buy Wall" was invisible in exchange volume data but obvious in custody flows. Same lesson here. The visible token metric is flashing open-source dominance. The invisible spend metric says closed models still own the economic gravity. Both are true. The question is which one bends first. Takeaway: watch the premium end. If Anthropic and OpenAI start cutting prices on their top-tier models, they are defending against a real threat. If they hold prices and instead ship dramatically better reasoning, the open-source ceiling gets tested. The next six months will tell us whether this is a permanent restructuring or just the market discovering a new equilibrium. I am betting on the latter, but I have been wrong before. Speed is safety. The data is live. Watch the flow, not the headlines.