U.S. Companies Collectively Switch Sides: China’s Large Model Market Share Soars to 46%

46% in a Single Week: American Companies’ Tokens Are Quietly Flowing to China

This CNBC report published on July 7 brought into the open something that had already been widely discussed behind the scenes in the industry: American companies are using Chinese large models, and doing so far more aggressively than most people imagined.

The data comes from OpenRouter — an aggregation platform that offers a strong window into real global developer usage patterns because it is not filtered through PR narratives. The numbers look like this: in the first half of 2025, Chinese models accounted for only 4.5% of tokens used by U.S. enterprises; the average over the previous 12 months was only 11%. But starting February 8 this year, that figure consistently stayed above 30% every week, at one point peaking at 46%.

In other words, within just a few months, Chinese models’ penetration into the U.S. enterprise market increased nearly tenfold.

The Trigger: Even Silicon Valley Couldn’t Handle Claude and GPT Bills Anymore

This shift was not driven by sentiment or ideology — it was purely about money.

Forbes previously reported a highly representative case: Uber, due to large-scale use of Anthropic’s Claude Code, burned through its entire AI budget for 2026 in just four months. This was not an isolated incident, but a signal of an industry-wide paradigm shift — AI companies broadly moved from fixed subscriptions to usage-based billing, long-context retrieval became standard, and simple Q&A evolved into multi-step reasoning. If companies continued using flagship models from OpenAI and Anthropic indiscriminately, costs would grow exponentially.

Even worse, OpenAI and Anthropic both raised token pricing for their flagship models almost simultaneously in the first half of this year. For companies that had already deeply embedded AI into their product pipelines, this effectively threw all unit economics out of balance overnight.

Kyle Chan of the Brookings Institution put it bluntly: “In the past, American companies chose models based only on deployment capability, regardless of cost. Now they’re all starting to calculate carefully.”

Silicon Valley Engineers Acknowledge How Much Chinese Models Have Improved This Year

Matching performance was the prerequisite — otherwise nobody would use them no matter how cheap they were.

Several milestones from the first half of this year stand out:

• DeepSeek once again reset the cost-performance curve early this year with its new-generation model, then released another flagship in April
• Zhipu GLM-5.2, released in early July, was described by Reuters as having coding and agent capabilities “comparable to leading American products”
• The Tongyi Qianwen (Qwen) series has consistently topped Hugging Face’s open-source model download rankings

A Citigroup research report showed that by June 2026, open-source models accounted for 65% of tokens processed by OpenRouter, nearly double the share at the start of the year. During the week of June 15–21, Chinese large models reached 18.81 trillion tokens in global usage, surpassing the United States for eight consecutive weeks and firmly ranking first worldwide. Among the global top five, four were Chinese models.

There is also a key structural difference here: China’s leading models largely follow an open-source/open-weight strategy, while flagship models from OpenAI, Anthropic, and Google are all closed-source. For enterprises, open weights mean local deployment, deep customization, and avoidance of vendor lock-in — all of which are nearly mandatory requirements in compliance and security auditing scenarios.

Lindy’s Migration: A Symbolic Event

In June, AI agent startup Lindy switched all of its production traffic from the Claude series to DeepSeek. This was not a small cost-saving experiment — it was a full migration.

Lindy CEO Flo said something particularly interesting in an interview with CNBC: they did not simply find an “acceptable” substitute. Instead, they discovered that “within their agent workflows, DeepSeek delivered the same quality at only a fraction of Claude’s cost.”

What matters even more is who publicly endorsed this trend: Microsoft CEO Satya Nadella and Coinbase CEO Brian Armstrong both stated that smaller, lower-cost models are already sufficient for a substantial portion of enterprise needs. Coming from Nadella, this carries special weight — Microsoft is OpenAI’s largest shareholder.

Why Chinese Models Can Reach These Price Levels

This is not simply a matter of subsidies or operating at a loss to gain market share. It is the result of an entire industrial chain:

1. Compute costs: China’s “Eastern Data, Western Computing” initiative combined with large-scale infrastructure buildout has pushed compute prices far below Silicon Valley levels. China’s integrated national computing network is expected to reduce training costs by another 25% and inference costs by 10%
2. Electricity costs: More than 60% of AI compute center operating costs come from electricity. Western China has cheap and surplus renewable energy, while data centers around Silicon Valley are struggling with grid expansion
3. Engineering talent advantage: A massive pool of young technical talent allows Chinese teams to validate multiple technical approaches in parallel and experiment at lower cost

Wang Xiwen, president of the Beijing Huaxia Industrial Internet Intelligence Technology Research Institute, summarized it well: “This cost advantage was not saved into existence or subsidized into existence — it is the product of a complete industrial ecosystem.”

The Regulatory Paradox: The More Regulation Tightens, the More Companies Are Pushed Elsewhere

There is another intriguing backdrop here. At the end of June, OpenAI, at the request of the U.S. government, delayed the full launch of a batch of new models. In the same month, after negotiations between the Trump administration and Anthropic, export controls on the Mythos and Fable models were finally lifted.

The U.S. government is simultaneously tightening regulation on its strongest domestic models while exploring ways to block the spread of overseas alternatives. But the result is that domestic closed-source models are slowing their release cadence, while the ecosystem window for open-source/open-weight models is becoming even larger — precisely the arena where Chinese companies dominate.

What This Means for Developers

Pan Helin, a member of the Ministry of Industry and Information Technology’s Information and Communications Economy Expert Committee, offered a practical assessment: excessive reliance on a single vendor means exposure to uncertainties such as vendor lock-in, outages, and sudden price hikes. Only multi-model strategies preserve bargaining power against large model providers.

From an engineering perspective, a reasonable architecture today probably looks something like this:

• Complex reasoning and agent planning: use top-tier closed-source models like Claude Opus or GPT-5
• Code generation and everyday dialogue: use high cost-performance models such as DeepSeek, GLM, or Qwen
• Specialized vertical scenarios: use domain-specific fine-tuned models

The era of using one universal hammer for every nail is over.

This is also why model aggregation layers have become so popular recently — a single key capable of accessing all mainstream models fits the current reality of mixed-model usage and scenario-based routing. OpenAI Hub (openai-hub.com) follows this approach: it supports the OpenAI API format with direct domestic connectivity, allowing GPT, Claude, Gemini, DeepSeek, GLM, and Qwen to all be called through a unified interface. Switching providers requires no code changes, which is genuinely useful in today’s cost-sensitive environment of “use whichever model is cheapest.”

One Final Thought

The real meaning of this news is not that “Chinese models have won,” but that the AI industry’s business logic is shifting from “pursuing the strongest” to “pursuing the most suitable.” Once token bills become a central issue in the CFO’s office, cost-performance is no longer a backup option — it becomes the primary one.

Silicon Valley companies quietly turning toward Chinese models is simply one of the clearest manifestations of this shift.

Sources

• ITHome: With Significant Cost Advantages, Chinese AI Large Models Are Gradually Winning Favor Among American Enterprises — Chinese-language first report summarizing the CNBC coverage and OpenRouter data source