A Citi note cited in trade coverage put leading Chinese models as low as 18 cents per million tokens against roughly $4 for top US frontier models, while the capability gap narrows. The USCC's own example: Kimi K2.5 cost about four times less than GPT-5.2 in January 2026 at an equal intelligence score. No visible mechanism closes this soon.

When a task does not need the best model, teams route it to the cheapest one that clears the bar. That behavior is real but young. Glean's CEO estimates roughly 95 percent of enterprise AI usage still runs on frontier models, so the low-cost tier is taking share from a small base.

If steps one and two continue, the market price for routine inference trends toward the open-weight floor. US labs either match on mid-tier work, which burns margin, or concede the volume, which shrinks the revenue base their valuations assume. Microsoft is accelerating this from inside the US ecosystem, shipping low-cost models and routing GitHub Copilot users to the cheapest fit per task.

Both OpenAI and Anthropic filed confidentially for IPOs in early June 2026. One equity analyst's blunt read: current growth rates are the fastest these companies will ever post, which is itself a reason to list now. If commoditization bites during the first public quarters, the repricing happens in the open, against valuations near a trillion dollars each.

This is the link the casual theory misses, and where a model-layer problem becomes a market-wide one. One analysis of disclosed cloud commitments puts about half of the four largest US cloud providers' revenue backlog, on the order of $1 trillion, as owed by two counterparties with deeply negative free cash flow: OpenAI and Anthropic. Hyperscaler capex is running between roughly $400 billion for the four biggest names, per Bloomberg Intelligence figures in the USCC report, and $660 to $725 billion once Oracle and the wider set are included in secondary tallies. For scale on the US-China gap, see the AI Race Facts compute-lead breakdown.

If the labs' revenue assumptions get marked down, the backlog underwriting that build-out is marked down with them. That is how a token price war reaches Nvidia, Oracle, and the broader index.

Even in the full scenario, "US AI crashes" is imprecise about who crashes. The exposed tier is pure-play model sellers with no product moat, plus anyone whose valuation assumes token growth forever. The protected tier sells outcomes rather than tokens: workflows, agents, compliance, distribution. The likely shape is not extinction but the model layer going the way of airlines or memory, brutal economics at the commodity tier, profit concentrated in whoever owns the customer. This step is one plausible outcome among several and cannot be verified in advance.