Nearly half the AI tokens flowing through US enterprise workloads in peak weeks now run on Chinese models. That's the finding of a CNBC analysis of OpenRouter routing data published this month: Chinese-origin models hit a weekly peak of 46% of US enterprise token usage, and haven't dropped below 30% in any week since early February 2026. A year ago that figure was in single digits. This article breaks down what's driving the shift, the price gap that explains almost everything, the security questions US companies are quietly setting aside, and what it means for your company's AI stack.
From 4.5% to 46% in Eighteen Months
The trajectory is steep. According to the CNBC investigation of OpenRouter data, Chinese models accounted for roughly 4.5% of enterprise token volume in the first half of 2025 and about 11% averaged over the following twelve months. Then came the breakout: in the week of February 9–15, 2026, Chinese models surpassed American ones on the platform for the first time, at 4.12 trillion weekly tokens. Every week since February 8 has seen Chinese models hold at least 30% of enterprise token volume, peaking at 46%.
The leaders are familiar names. DeepSeek alone routes 17.6% of OpenRouter's tokens — the single largest vendor on the platform — with Alibaba's Qwen at 13.9%. These aren't hobbyist experiments; the volume is coming from production enterprise workloads: summarization, extraction, classification, code completion and internal agents.
The 60–90% Price Gap That Explains Everything
Set ideology aside and the adoption curve looks like simple procurement math. Open-weight Chinese models consistently price 60–90% below leading American offerings. As of June 2026, DeepSeek V4 Flash costs $0.14 per million input tokens; OpenAI's GPT-5.5 lists at $5.00 — a 35x difference.
The old enterprise logic: frontier model for everything, because quality justified cost. The new logic: match the model to the task. For routine workloads that don't need frontier-class reasoning, a model that costs 90% less and performs within 5–10 percentage points is the rational choice — and CFOs noticed after 2026 budget blowouts made AI spend a board-level topic. It's the same cost pressure reshaping the whole market, from GitHub Copilot's move to usage-based billing to enterprises building internal model-routing layers that send each request to the cheapest adequate model.
The Questions Enterprises Are (and Aren't) Asking
Because these are open-weight models, the standard security objection — "your data goes to China" — doesn't strictly apply. Most enterprise deployments run DeepSeek or Qwen weights on US cloud infrastructure or on-prem GPUs; no traffic touches Chinese servers. That distinction is doing a lot of work in procurement approvals right now.
But harder questions remain. Model provenance: can you verify training data and behavior of weights trained by a foreign lab? Alignment and backdoor risk: subtle, task-specific failure modes are difficult to audit. Regulatory exposure: US agencies have restricted Chinese AI apps in government contexts, and policy could extend to weights. And dependency risk: teams optimizing pipelines around a Chinese model family inherit its release cadence and licensing choices. Washington has taken notice — Congress and the US-China Economic and Security Review Commission have both examined how China's open-weight strategy reinforces its broader AI industrial position.
What to Watch Next: Routing Layers and Policy Risk
Two forces will decide where this settles. First, model routing is becoming standard enterprise architecture — automatic dispatch of each task to the cheapest capable model. That structurally favors whoever is cheapest per unit of adequate quality, which today is often a Chinese open-weight model. Second, policy: any new federal restriction on foreign model weights in regulated industries could freeze adoption in finance, healthcare and government contracting almost overnight. Enterprises are hedging by keeping American frontier models — the competitive landscape we mapped in our analysis of ChatGPT falling below 50% market share — for high-stakes reasoning while routing commodity work to open weights.
What This Means for You
If you run engineering or AI budgets at a US company, three actions follow. First, benchmark your routine workloads against a cheap open-weight model — if quality holds within your tolerance, the 60–90% savings are real and immediate. Second, deploy open weights on infrastructure you control and document that isolation; it's what makes the compliance conversation winnable. Third, write down your policy-risk exit plan: know which workloads run on which model family and what it costs to swap. Cost savings that can't survive a regulatory change aren't savings — they're deferred migration debt.
Frequently Asked Questions (FAQs)
Q: Are US companies really using Chinese AI models?
A: Yes. Per CNBC's analysis of OpenRouter data, Chinese-origin models have handled at least 30% of US enterprise token volume every week since February 2026, peaking at 46%. DeepSeek (17.6% of routed tokens) and Alibaba's Qwen (13.9%) lead the pack.
Q: Is it safe for a US company to use DeepSeek or Qwen?
A: Open-weight models can run entirely on US cloud or on-prem infrastructure, so data doesn't flow to Chinese servers. But provenance, auditability and regulatory risk remain open questions — most security teams approve them for low-sensitivity workloads while keeping regulated data on US frontier models.
Q: Why are Chinese AI models so much cheaper?
A: Open-weight releases create brutal hosting competition — anyone with GPUs can serve the model, so pricing collapses toward compute cost. DeepSeek V4 Flash at $0.14 per million input tokens versus GPT-5.5 at $5.00 illustrates the gap: 60–90% cheaper is typical across the category.
Q: Could the US government ban Chinese AI models for enterprises?
A: Restrictions already exist in government contexts, and Congress has studied China's open-weight strategy as an industrial-policy concern. A broader restriction on foreign weights in regulated industries is plausible, which is why smart enterprises keep swap-ready architectures rather than deep single-vendor integration.
The AI trade war isn't coming — it's already running inside US production stacks, one cheap token at a time. Whether 46% is the peak or just a waypoint depends on the next pricing move from American labs and the next policy move from Washington. Is your company routing to open weights yet? Share this with whoever owns your AI budget.
