Open Source Wins Traffic, Closed Source Wins Profit: AI Commercialization Hits a Fork in the Road

In July 2026, the AI industry presented a puzzle that has left many observers scratching their heads: open-source model token usage is soaring, yet their share of total enterprise spending continues to decline. Vercel's AI Gateway dashboard shows that DeepSeek has rapidly climbed to the number-one spot in token processing, now accounting for over one-third of platform traffic. At the same time, Anthropic still captures more than half of the platform's total AI expenditure. Traffic to open source, profit to closed source — AI commercialization is heading down a clearly bifurcated path.

The Data: A Widening Gap Between Traffic and Spend

Vercel's data is just one slice. On OpenRouter, a platform covering a broader market, the divergence is even more striking.

DeepSeek V4 Flash processes approximately 5.3 trillion tokens per week, making it the most popular model by far. Anthropic's most capable frontier model, Opus 4.8, handles about 2 trillion tokens per week — less than 40% of V4 Flash's volume.

But shift to the spending dimension, and the picture inverts entirely. Opus 4.8 averages roughly $1.37 per million tokens, while V4 Flash costs just $0.06 — a roughly 23-fold difference. Extrapolating from this pricing gap, Opus 4.8 almost certainly accounts for the lion's share of platform spending despite its lower token volume.

This is not an isolated case. Zhipu, with its GLM-5.2 model, has risen to fourth place in token volume on Vercel, and Nvidia's Nemotron is rapidly entering the market backed by the company's formidable industry relationships. The open-source camp's traffic share keeps expanding, but what it is consuming are mostly mature application scenarios — where enterprises have already shifted from "validate feasibility with the smartest model" to "run production with the fastest, cheapest model."

Decagon's Practice: Why 90% Open-Source Workloads Still Feed Closed-Source Growth?

Jesse Zhang, CEO of enterprise AI agent company Decagon, offers an even more striking data point: roughly 90% of the company's workloads now run on open-source models. Yet, he notes, overall enterprise spending on expensive frontier models has "barely declined."

The key to this paradox lies in the AI application lifecycle.

"When you run a customer service AI agent in production, latency directly determines whether the product is usable. Nobody wants a product where every conversational turn takes eight seconds to respond. So you need smaller, faster models," Zhang explains. "But off-the-shelf small models can't meet customer quality standards. They only work after extensive fine-tuning on specific tasks."

The problem is that frontier model labs largely do not offer this capability. Enterprises cannot fine-tune OpenAI's or Anthropic's most powerful models to fit their needs, nor do they own the small models these labs provide — they cannot shape them at will. As a result, the "small model plus deep fine-tuning" combination inherently forces enterprises toward open-weight models — not because of cost, not because of open-source ideology, but because there is simply no alternative.

Meanwhile, new application scenarios are emerging at an even faster rate. Every enterprise is using top-tier models to explore the next frontier — and while the token consumption during these exploration phases is low, the unit price is sky-high. Every dollar saved by migrating mature scenarios to open-source models is quickly replenished by exploratory spending on new ones.

Zhang summarizes: "Frontier model labs will continue to dominate application discovery, while open-source models will increasingly dominate production deployment."

Dario Amodei's Provocation: "Open Source Is a Misonomer"

Anthropic CEO Dario Amodei's earlier remarks provide the closed-source camp's annotation of this divergence.

"Even when a model is publicly available, you cannot see its internal mechanisms. That's why the industry typically calls these 'open weight' models, not 'open source,'" Amodei said. "Traditional open-source software can be modified by the community, continuously iterated, and strengthened through collaboration — but that advantage does not fully apply to large models."

He closed with a sharp observation: "When I see a new model, I never first ask whether it's open source. Even whether DeepSeek is open source doesn't matter to me. The only question I care about is: is this model good enough, and can it beat us on the tasks that matter?"

But Amodei's argument inadvertently reveals an uncomfortable truth: the closed-source camp's power to define what constitutes a "good model" is itself part of their profit moat. When the open-source camp redefines "good" from a different dimension — not absolute reasoning capability, but cost-effectiveness and controllability in specific scenarios — market divergence becomes inevitable.

The Essence of the Fork: Discovery vs. Deployment, General vs. Specialized

The underlying logic of this divergence can be distilled into a simple two-stage model.

Stage One: Application Discovery. When a brand-new AI application scenario emerges, no one knows exactly what form the problem will ultimately take. Enterprises default to the smartest general-purpose model available to explore the boundaries, willingly paying a premium for intelligence capabilities they may never actually need. At this stage, proving the scenario works matters more than controlling costs.

Stage Two: Production Deployment. Once a scenario is fully mature — once the enterprise understands the input data distribution, the behavior the model needs to exhibit, and the failure modes it must guard against — the trade-off inverts. General intelligence becomes a liability. What the enterprise truly needs is the smallest, fastest model, fine-tuned to do one specific task exceptionally well.

These two stages are not a zero-sum game — they represent a natural division of labor across different maturity slices of the same market. As Zhang puts it: "The declining share of open-source model spending is not because open-source models are failing. It is because the entire enterprise AI market is still in the earliest phase of its maturity curve."

Who Is Winning?

In the short term, the closed-source camp still firmly holds the profit side — Anthropic's spending share on Vercel and OpenRouter is the clearest proof. But in the long term, the traffic battle may prove more decisive. As DeepSeek, Zhipu, Nvidia, and others in the open-source and open-weight camp cover more and more production scenarios, they control not just token counts but the core pipelines of enterprise applications.

This may be the most important trend on AI's commercialization fork: the open-source camp is winning "mindshare," while the closed-source camp is winning "wallet share." The real victor will be whoever crosses into the other's territory first.