The numbers don't lie. They just refuse to tell the story everyone expected.

Open source AI is eating the world. DeepSeek alone now processes just over a third of every token flowing through Vercel's AI gateway. Z.ai's GLM-5.2 has muscled into fourth place in a matter of weeks. On OpenRouter, DeepSeek V4 Flash chews through 5.3 trillion tokens weekly — more than double the volume of Anthropic's flagship Opus 4.8.

Yet Anthropic still commands more than half of all AI spending on Vercel. Opus 4.8 costs roughly 23 times more per token than V4 Flash. Do the math: even at half the volume, the frontier model captures the lion's share of revenue.

Jesse Zhang calls this a life cycle, not a war. The Decagon CEO argues frontier models and open source aren't competitors — they're consecutive phases. Expensive models prove use cases. Cheap models run them at scale. As mature deployments graduate to lighter weights, new use cases emerge that only the frontier can handle. The pie grows. The slice stays.

He's right. But he's also missing the fault line.

Vercel's dashboard shows Anthropic's spend share has dipped only slightly over the past month — and that drop comes largely from Anthropic's own price hikes, not defection. OpenRouter tells the same story from a different angle: usage has democratized, spend hasn't. The market for AI-addressable tasks is expanding fast enough that frontier labs can dominate discovery while open source floods production.

That's the "yet" in the headline.

Two forces could break this equilibrium. First, the gradient between "only frontier can do this" and "open source does this fine" is steepening. Nvidia's Nemotron arrives with the rare combination of extreme adaptability and Nvidia's distribution muscle. If it delivers on early benchmarks, the discovery window for frontier labs shrinks. Second, enterprise buyers are learning. They're building abstraction layers — gateways, routers, eval frameworks — that make model swapping trivial. The switching cost that once protected Anthropic's moat is eroding by the quarter.

Zhang's data is thin. His theory leans on Vercel and OpenRouter snapshots, not longitudinal studies. But the snapshots are consistent: token volume has decoupled from token spend. That decoupling is the whole ballgame.

Frontier labs know this. Anthropic's pricing strategy — raising rates even as open source alternatives flood the zone — signals confidence that their quality delta still justifies the premium. For now, it does. Complex reasoning, long-context synthesis, agentic reliability: these remain frontier territory. But the territory is shrinking.

The deeper question isn't whether open source hurts Anthropic. It's whether the current pricing model survives the moment when open source catches up on the last few capabilities that matter. That moment isn't here. But Nvidia just showed up. And the enterprise buyers holding the purse strings are watching the gap close in real time.

Anthropic isn't hurting yet. The "yet" is doing a lot of work.