Half Your AI Runs on Chinese Models

How a 35x pricing gap quietly redirected nearly half of US enterprise API traffic to Chinese models, and why the government's crackdown on American AI may have accelerated the very migration it should have prevented

Author

Justin Wright
July 13, 2026 • Est. Reading Time: 7 minutes

Happy Monday!

Chinese-origin models now account for 46.4% of routed tokens on OpenRouter, compared to 35.7% for US-origin models. (source: CNBC)

On Monday, CNBC published an investigation that should alarm everyone who spent the past month debating how to regulate American AI. Chinese-origin models now account for 46.4% of routed tokens on OpenRouter, the developer platform that processes roughly 25 trillion tokens per week across 400+ models. US-origin models hold 35.7%. DeepSeek alone processes 5.13 trillion tokens weekly, making it the single largest vendor on the platform. One year ago, US models held 74% of the market. Today they hold roughly a third.

The catalyst is not ideology or espionage, it is the price tag. DeepSeek V4 Flash costs $0.14 per million input tokens, OpenAI's GPT-5.5 costs $5.00; that is a 35x gap. When Coinbase, a publicly traded $60 billion company, routes 1,200 AI agents through Chinese models to cut its AI bill in half, the migration has moved well past the experimental phase.

We have spent the last two issues watching the US government establish a de facto approval process over frontier American models. It pulled Fable 5 and gated GPT-5.6. Altman even proposed giving regulators a 5% equity stake. But while Washington was tightening control over American AI, the market was quietly routing half its traffic to Chinese alternatives that face none of that oversight. The government regulated supply while the market moved demand.

Chinese AI models now capture up to 46.4% of US enterprise API token usage, surpassing US-origin models at 35.7%. A year ago, US models held 74%. The shift is driven by a 35x pricing gap: DeepSeek V4 Flash costs $0.14 per million tokens versus GPT-5.5 at $5.00. Coinbase cut its AI bill in half. Startup Lindy moved 100% of its traffic from Claude to DeepSeek, cutting costs 90%. Congress has opened a joint investigation, but the migration has been underway since February and may already be embedded too deeply in production infrastructure to reverse with policy alone.

TL;DR

The Reversal

The scale of the shift is hard to overstate.

OpenRouter provides the clearest picture of real-world model selection. As of the last week of June, Chinese models processed approximately 18 trillion tokens weekly on the platform while US models processed 5.5 trillion. The reversal was not gradual. The week of February 9-15, 2026, Chinese models processed 4.12 trillion tokens, edging past the 2.94 trillion handled by US models. It was the first time Chinese systems outpaced American ones on the platform.

DeepSeek holds 17.6% of all routed tokens, the single largest vendor by volume. Alibaba's Qwen follows at 13.9% with 2.77 trillion tokens weekly. To put this in context: DeepSeek, a Chinese AI lab that most enterprise procurement teams had never heard of 18 months ago, now processes more tokens through US developer infrastructure than any American model provider.

Token Share Shift: US vs. Chinese Models on OpenRouter

Period

US Model Share

Chinese Model Share

Largest Single Vendor

July 2025

74%

20%

OpenAI

February 2026 (inflection)

~42%

~44%

DeepSeek

Late June 2026 (latest)

~20-35%

~46-48%

DeepSeek (17.6%)

The 35x Gap

The migration has a single primary driver: cost. DeepSeek V4 Flash costs $0.14 per million input tokens while GPT-5.5 costs $5.00. Alibaba's Qwen 3.7 Max matches Claude Opus 4.7 on several agentic benchmarks at roughly half the price. In May, DeepSeek permanently cut V4-Pro pricing by 75%, bringing input tokens to $0.435 per million and output to $0.87. Open-source Chinese models are consistently 60% to 90% cheaper than the leading offerings from Anthropic and OpenAI.

Ramp's chief economist Ara Kharazian identified cost awareness as the primary catalyst. DeepSeek became the number one trending software vendor on the Ramp index in June, signaling that Chinese models have moved from developer sandboxes into formal corporate expense reports. When a model shows up on your company's credit card statement, the experimentation phase is over.

The performance gap between Chinese and American models has narrowed enough that the cost differential becomes the deciding factor for most procurement teams. For high-volume enterprise users processing billions of tokens per month, a 35x pricing gap translates to millions of dollars annually. At that scale, "good enough" at a fraction of the price is a fiduciary obligation.

The Case Studies

The CNBC investigation highlighted two companies that illustrate the migration at different scales. Coinbase, running 1,200 full-time AI agents, cut its internal AI spending by nearly half. CEO Brian Armstrong's team defaulted engineers to two Chinese-origin open-weight models, GLM 5.2 and Kimi K2.7 Code, through an internal LLM gateway. The company did not abandon American models entirely. It built a routing system that directs tasks to the cheapest model capable of handling them. Most tasks, it turns out, do not require a $5-per-million-token model.

Lindy, a 25-person AI agent startup, made a more dramatic move. CEO Flo Crivello moved 100% of the company's managed-agent traffic from Anthropic's Claude to DeepSeek V4 Flash. The migration took six to nine months of evaluation, prompt re-engineering, and gradual rollout. It cut inference costs on migrated routes by approximately 90%. Crivello told CNBC "you could see that cost curve go down, like, crash to the ground." Lindy's AI costs had become "unsustainable" on American models. Lindy still uses Claude internally, because subscription economics make that viable. But for production workloads at scale, the cost differential was existential.

The Government's Blind Spot

On June 12, the US government pulled Fable 5 off the market via export controls. On June 25, Commerce Secretary Lutnick called Sam Altman to delay GPT-5.6's launch. On July 2, Altman proposed giving the government a 5% equity stake. In three weeks, the government built a de facto approval process for frontier American AI. During that same period, enterprise adoption of Chinese models accelerated.

The Fable 5 shutdown created an 18-day gap where one of the most capable American models was unavailable. GPT-5.6 was restricted to roughly 20 government-vetted organizations for 12 days before going fully public on July 9. Both events gave enterprise teams a reason to diversify their model supply chains, and Chinese models were standing right there at $0.14 per million tokens.

The government is tightening controls on models built by American companies that employ American workers, pay American taxes, and submit to American oversight. The alternatives gaining market share are built by companies subject to different data governance standards, different security norms, and different geopolitical interests. No Chinese model has volunteered for the 30-day review framework established by the June 2 executive order, and congress has noticed.

The House Select Committee on China and the House Homeland Security Committee opened a joint investigation in April, probing companies including Airbnb and Anysphere (Cursor) for their use of Chinese-origin models. Chairman Andrew Garbarino warned that "the Chinese Communist Party is no longer just nipping at our heels in artificial intelligence." Moonshot AI and Zhipu AI were named in a separate congressional security probe in May, but the investigation is reactive. The migration has been underway since February, and by the time policy catches up, the routing decisions will already be embedded in production infrastructure.

What This Means for Practitioners

For enterprise buyers, model diversification is no longer optional. The companies that built routing layers early, like Coinbase with its internal LLM gateway, are the ones capturing the cost advantage. If you are running inference at scale on a single American provider, you are overpaying for tasks that do not require frontier capability. Build the routing infrastructure now, even if you choose not to route through Chinese models. Optionality has value regardless of where you direct the traffic.

For AI startups, the Lindy case study is a warning. If your unit economics depend on American frontier model pricing, a competitor routing through DeepSeek at 90% lower cost will undercut you. Build for model portability from the start. The performance gap is narrowing, but the cost gap is not.

For policymakers, the gap in the current approach is becoming visible. Regulating American frontier models while Chinese alternatives operate outside that framework creates a system where compliance is a competitive disadvantage. The question is whether a framework that applies only to American companies produces the outcome anyone actually wants.

The Bottom Line

One year ago, US models processed three-quarters of the tokens flowing through major developer platforms. Today, Chinese models process nearly half and the reversal happened in five months. The driver was not technological superiority; it was a 35x pricing gap.

The US government just spent three weeks building a regulatory framework for American AI. It pulled Fable 5, gated GPT-5.6, and is negotiating equity stakes and approval processes. None of that applies to the models that now handle the largest share of enterprise API traffic. The government is regulating the industry it can see while the market migrates to the one it cannot.

In motion,
Justin Wright

If the US government's emerging framework for frontier AI oversight only applies to American companies, does tighter regulation make American AI safer, and is there a version of this policy that addresses both the safety concern and the competitive one without choosing between them?

Food for Thought

Quick Hits

  • GPT-5.6 went fully public on July 9 after a 12-day government-gated preview. Three tiers: Sol ($5/$30 per 1M tokens), Terra ($2.50/$15), and Luna ($1/$6). Sol runs on Cerebras at up to 750 tokens per second. (OpenAI)

  • SK Hynix listed on Nasdaq as SKHY on July 10, raising $26.5 billion in the largest ADR offering in market history. The company holds roughly 60% of the HBM market powering AI accelerators. Shares opened at $170, up 14% from the $149 offering price. (CNBC)

  • Grok 4.5 shipped July 8-9 via SpaceXAI API and Cursor. 1.5 trillion parameter MoE architecture at $2/$6 per million tokens, undercutting GPT-5.6 Sol by more than half. Not yet available in the EU. (SpaceXAI)

  • Cloudflare will block "mixed-use" AI crawlers from ad-hosting pages by default starting September 15. New Pay Per Use feature pays publishers when their content appears in AI chatbot answers. First partners: Ceramic.ai and You.com. (TechCrunch)

If you haven’t listened to my podcast Mostly Humans: An AI and business podcast for everyone yet, new episodes drop every week!

Episodes can be found below - please like, subscribe, and comment!