The OpenRouter LLM Market:
Who Owns Volume vs. Who Owns Revenue

Chinese open-weight labs have captured the majority of token volume on OpenRouter — but a single Western lab still captures the majority of the dollars. The gap between those two facts is the whole investment thesis.

1 · Vendor share — the two leaderboards

Δpp = percentage-point shift in share of the total — the zero-sum view of who's taking ground.

2 · Substitution in the wild — the same apps, both vendors

The clearest evidence that volume migration threatens revenue: the highest-volume apps run both Anthropic and DeepSeek for the same job. 6 of the top 10 apps mix them — and DeepSeek V4 Flash undercuts Claude Sonnet by ~77× on output price. When a workflow already calls both, switching share is a config change, not a migration cost.

Share of the app's tokens by vendor. Even Anthropic's own Claude Code routes ~13% to DeepSeek; roleplay/agent apps lean majority-DeepSeek. The cheaper model is already inside the funnel.

3 · The substitution engine — families, not versions

Individual versions churn constantly, which makes per-model churn misleading. The real unit is the family: usage migrates within a family (e.g. Claude Opus 4.6 → 4.7 → 4.8) while the family's total tells you whether the franchise is winning. "Lead version" share is a stickiness signal — how fast users consolidate onto the newest release.

4 · Lifecycle of the ecosystem

Every tracked model placed on its curve — launch → ramp → peak → decline → death:

A long "declining" tail is normal — it's last-generation versions bleeding into their successors. "Dead" = had real volume, now ~zero with no in-family heir absorbing it.

5 · Winners & losers (family churn, 30d)

6 · The model treadmill — and China churns fastest

Models don't age gracefully; they get cannibalised by their own successors. Measuring each generation's useful life (launch → usage falling below half its peak) exposes how fast the treadmill runs — and it runs much faster in the East. Average useful life: China ~63 days vs West ~92 days.

Shortest-lived first. The fastest Chinese generations turn over in under two weeks, while Western flagships hold for two to four months — the table shows the full spread. On average a Chinese generation's useful life is ~63d vs ~92d in the West. Faster churn = faster iteration, but a brutal amortisation window on training spend.

7 · Implications

• Underwrite the volume curve, not just the dollar. Volume is the leading indicator; revenue lags. Anthropic's ~55% spend share is the clock that moves last — watch the cheap-model volume share as the early-warning signal.
• Low-end disruption is live. DeepSeek V4 Flash at ~77× lower output-token price is taking volume inside the same apps that pay for Claude. Good-enough at a fraction of the cost is exactly how incumbents lose the bottom of the market first.
• Anthropic's moat is the quality gap, and it's a depreciating asset. The premium holds only while Claude is clearly better for the high-value slice. Track the spend-share trend, not the level.
• Revenue is gated by price, not volume. A Chinese token lists ~2.2× cheaper, so even past majority volume China holds a sliver of the dollars — at today's gap it would need ~68% of all volume for half the revenue. The metric that actually moves the dollar line is China's blended $/token, not its volume share.
• The treadmill runs faster in the East. Chinese generations live ~63d vs ~92d in the West — faster iteration and harder to track, but a punishing amortisation window.
• In-family upgrades are sticky; cross-vendor switching is cheap once both are wired in. Brands retain users across their own versions, but multi-model apps make vendor-switching a routing decision.