46% of US corporate AI tokens go to Chinese models — the CNBC survey that changes everything

46% of US enterprise AI tokens go to Chinese models — the CNBC investigation that changes everything

🔎 The figure no one expected

46%. That is the share of API tokens consumed by Chinese models in the US enterprise segment during the weekly peak of June 2026, according to OpenRouter data compiled by Yahoo Finance and the CNBC investigation published on July 7, 2026. A year ago, this same metric fluctuated around 4.5% in the first half of 2025, with an annual average of 11%.

This is not an emerging trend. It is a complete shift. And it raises a simple question that every American CTO still refuses to ask out loud: why pay 10 to 15 times more for a Western model when the Chinese one does the job at least as well on the majority of tasks?

The key points

• The share of Chinese models on OpenRouter has exceeded 30% every week since February 8, 2026, peaking at 46% (Yahoo Finance data, July 2026).
• GLM-5.2 (Z.ai) recorded a ×27 growth in daily token volume and ×80 in customer numbers on Vercel in its very first week of availability.
• Chinese open-weight models cost 60 to 90% less than GPT or Claude equivalents at comparable performance levels, pushing companies to adopt the "advisor model" architecture.
• US lawmakers opened a formal investigation on July 8, 2026, into the supply chain security implications (CNBC).

Recommended tools

The figures: from 4.5% to 46% in 18 months

The trajectory is unequivocal. According to the State of AI 2025 d'OpenRouter report, the share of Chinese open-source models began to accelerate from mid-2025. At the time, no one was paying attention: US labs' margins were too comfortable to worry about peripheral competition.

At the end of 2024, Chinese models represented less than 2% of OpenRouter traffic, according to JPMorgan data cited by Economic Times. In the first half of 2025, the figure was 4.5%. The 12-month rolling average stood at 11%.

Then February 2026 marked a point of no return. The weekly share of Chinese models has never dropped below 30% since February 8. The peak of 46% was reached in June, as documented by Yahoo Finance.

Looking more broadly, Dataconomy reports that Chinese models reach a 61% total market share on OpenRouter across all segments combined (not just US enterprise), with 4 of the 5 most used models being of Chinese origin. Data Gravity confirms this figure of ~61% for May 2026 for open-weight models alone.

Trending Topics specifies that among the top 10 models on OpenRouter in February-March 2026, the token consumption of Chinese models already reached 61%. Analysis by Our World in Data confirms that China has now surpassed the United States in the number of open-weight models in OpenRouter's top 50.

This is a structural shift, not a temporary blip.

GLM-5.2 : the model that panicked Silicon Valley

Z.ai's GLM-5.2 is not a marginal newcomer. It is the catalyst that turned a slow trend into a sudden explosion. The data compiled by ByteIota is unequivocal: in its first week of availability on Vercel, GLM-5.2 saw a 27-fold increase in daily token volume and an 80-fold increase in unique clients.

Two factors explain this explosive adoption. First, raw performance. GLM-5.2 reaches 62.1% on SWE-bench Pro, a benchmark for solving real-world software problems. To contextualize: OpenAI's GPT-5.5, the highest-performing frontier model in agentic, scores 58.6% on this same benchmark. GLM-5.2 doesn't just rival Chinese models: it beats them on their home turf, code.

Next, the price. GLM-5.2 is offered at $0.40 per million input tokens and $0.40 for output. Compared to the pricing of GPT-5.5 or Claude Opus 4.7 Adaptive, which range between $15 and $75 per million tokens depending on the segment, the gap is on the order of 40 to 100x. Even taking into account recent price optimizations by Anthropic and OpenAI on their mid-tier models, the gap remains 60 to 90% in favor of Chinese models.

Z.ai chose an MIT license for GLM-5.2, with an explicit mention: "no regional limits". It is a political message as much as a technical one. No geographical restrictions, no non-use clause for American companies. Z.ai knows that the license is a selling point compared to Western models subject to increasing regulatory constraints.

The "advisor model" architecture: why developers are making the switch

The revolution is not technical. It is architectural. The pattern currently emerging massively among US enterprise developers, documented by ByteIota, is called the "advisor model".

The principle is simple. Instead of sending every request to an expensive frontier model (GPT-5.5, Claude Opus 4.7), a cheap Chinese model is placed on the front line — the "advisor". It processes the request and produces a response. If confidence is sufficient (high certainty score, routine task), the response is validated directly. If the model hesitates, the request is escalated to a Western frontier model.

The mathematical result is devastating for American labs. In practice, 70 to 85% of enterprise requests do not require frontier reasoning. They are routine: document summarization, entity extraction, email sorting, template response generation, first-level debugging. The Chinese model handles them perfectly for a fraction of the cost.

Only the remaining 15 to 30% — complex multi-step reasoning, critical compliance tasks, code generation for production systems — are escalated to GPT or Claude. The token bill mechanically drops by 60 to 80%, and the quality perceived by the end user remains identical.

This is exactly the scenario made possible by models like Z.ai's GLM-5 (Reasoning) and DeepSeek V4 Pro. GLM-5, with its agentic score of 82 and GLM-5.1 at 83 in general, offers a more than sufficient entry point for the majority of mid-tier use cases.

Why Western labs priced themselves out of the market

The answer is uncomfortable for Silicon Valley: it's their own fault. OpenAI, Anthropic and Google pursued a monopoly strategy through high prices, betting that premium quality would justify any price tag. For two years, it worked. Companies paid without flinching because there was no credible alternative.

The simultaneous arrival of GLM-5.2, DeepSeek V4 Pro and Kimi K2.6 shattered this assumption. DeepSeek V4 Pro (Max) scores 88 overall, one point less than GPT-5.4 and two points less than Gemini 3.1 Pro. DeepSeek V4 Pro (High) reaches 84, equaling Kimi K2.6 and surpassing Claude Sonnet 4.6 (83) and GLM-5.1 (83). For 60 to 90% less.

The mid-tier enterprise segment — companies with 50 to 5,000 employees and AI budgets of $5,000 to $200,000 per month — is the most affected. These companies neither have the means nor the need to pay frontier prices for tasks that do not require cutting-edge reasoning. They simply had no choice before 2026.

Now, they have a choice. And they are making it.

The Trump administration's crackdown on AI, documented by CNBC on June 30, 2026, paradoxically accelerated the trend. By restricting access to certain models and complicating the US regulatory landscape, the measures made Chinese open-weight models even more attractive due to their ease of adoption. A developer downloads the weights, hosts them, and that's it. No enterprise contract, no price negotiation, no confidentiality clause to renegotiate.

The real risks: what the numbers don't say

Reasoning solely in terms of cost and benchmarks means ignoring the three elephants in the room: data jurisdiction, political content restrictions, and the reliability of tool-calls in production.

Data jurisdiction

This is the most serious point. When an American company sends sensitive business data to a Chinese API, even via an aggregator like OpenRouter, the question of applicable jurisdiction is blurry. Do the data pass through Chinese servers? Are they stored, even temporarily? The open-weight self-hosted model (like GLM-5 in self-host, agentic score 82) eliminates this problem, but the cloud API does not. BERI published a security/cost decision-making framework specifically to evaluate this risk, but most companies still have no formal process in place.

Political content restrictions

All Chinese models apply guardrails on politically sensitive topics for the CCP. In practice, for 95% of enterprise use cases (code, data, logistics, HR, finance), this has no impact. But for media companies, content platforms, or search tools, it is an absolute blocker. The model refuses to answer or produces evasive answers on certain topics.

Reliability of tool-calls

In agentic, the reliability of tool-calls is critical. Frontier models like GPT-5.5 (98.2 in agentic) and Claude Opus 4.7 Adaptive (94.3) have been optimized specifically for this scenario. Mid-tier Chinese models like GLM-5 (82) or Kimi K2.6 (88.1 in self-host) are reliable for simple tool chains, but can degrade on complex agentic workflows with 10+ steps and conditional branching decisions.

The political response: investigation, but no regulation

On July 8, 2026, the day before the detailed data was published, US lawmakers opened a formal investigation into the massive use of Chinese models by US companies, as reported by CNBC. The investigation specifically targets AI supply chain security.

But there is a gap between an investigation and effective regulation. The United States has no direct equivalent to the EU AI Act whose Commission just published the labelling playbook with a deadline of August 2, 2026. On the American side, the approach relies on voluntary standards, not on binding obligations.

This regulatory vacuum directly benefits Chinese models. As long as there is no clear framework prohibiting or conditioning the use of foreign models for sensitive data, companies follow the economic logic. And the economic logic points to Beijing.

The political pressure is real, but it is also counterproductive. Every hostile political statement reinforces the perception that US restrictions are motivated by the protection of US labs, not by national security. And this pushes even more developers toward Chinese models as a reflex of resistance to regulatory capture.

The broader context: the GPT-5.6 family does not reverse the trend

While Chinese models are chipping away at market share, OpenAI announced the GPT-5.6 family with Sol, Terra, and Luna. This is an expected response: segmenting the offering to cover the mid-tier with models cheaper than GPT-5.5 while maintaining the frontier premium.

The problem is that the mid-tier segment is precisely where Chinese models are the most competitive. A model like Gemini 3.1 Pro (92 overall, 87.3 in agentic) or GPT-5.4 (89 overall, 87.6 in agentic) is now directly challenged by DeepSeek V4 Pro Max (88 overall) and Kimi K2.6 (84 overall, 88.1 in agentic self-host) at a fraction of the price.

OpenAI's segmentation strategy arrives a quarter too late. Developers have already integrated Chinese models into their pipelines. Switching architectures costs time and money. The switching barrier is now on the side of the Chinese models, not the American ones.

What businesses must do now

The decision framework is simpler than it appears. It comes down to three questions.

First question: is the data sensitive? If yes (patient data, state secrets, regulated financial data), the model must be self-hosted or operated by a provider under US/European jurisdiction. Open-weight models like GLM-5 (self-host, 82 in agentic) or Kimi K2.6 (self-host, 88.1) are technically possible but require internal legal analysis. If the company does not want to take jurisdictional risks, stick with GPT-5.5 or Claude Opus 4.7 Adaptive.

Second question: is the use case complexly agentic? If the workflow involves more than 5 sequential steps with conditional tool-calls, frontier models remain superior. GPT-5.5 (98.2) and Gemini 3 Pro Deep Think (95.4) have no Chinese equivalent at this level of complexity. Claude Opus 4.7 Adaptive (94.3) and GPT-5.4 Pro (91.8) complete the podium.

Third question: everything else. For the remaining 80% of use cases (generation, summarization, extraction, sorting, first-level code), a Chinese model via the advisor model architecture is the rational choice. GLM-5.2 for code, DeepSeek V4 Pro for general, Kimi K2.6 for volume. The savings are on the order of 60 to 90% with no perceptible loss of quality.

❌ Common mistakes

Mistake 1: Confusing open-weight and open-source

GLM-5.2 is distributed as open-weight under the MIT license. This does not mean that the training code, data, or RLHF processes are public. An open-weight model gives access to the weights, not the recipe. Calling it "open-source" is a categorization error that many decision-makers still make, and it skews their risk assessment.

Mistake 2: Ignoring the total cost of adoption

$0.40 per million tokens is an introductory price. The real cost includes integration into the existing pipeline, setting up the advisor model pattern, regression testing, quality monitoring, and potentially rewriting prompts optimized for GPT/Claude. The TCO is not linear with the price per token.

Mistake 3: Thinking that the political inquiry will stop the trend

The inquiry opened on July 8, 2026, by US lawmakers is a political signal, not a regulatory tool. Without binding legislation—which the United States has not yet produced at this time—companies are free to use whatever models they want. Relying on a ban to justify not evaluating Chinese models means falling behind by a decision cycle.

Mistake 4: Using a Chinese model for public content without a label

With the EU AI Act and its labeling playbook, any company operating in Europe must label AI-generated content. If this content is generated by a Chinese model without appropriate traceability, the double non-compliance (unlabeled AI + undocumented provider) exposes the company to enhanced sanctions.

❓ Frequently Asked Questions

46% of tokens, does that mean 46% of companies?

No. It is 46% of the token volume, not the number of companies. A handful of high-volume large companies (e-commerce, logistics, fintech) can account for a disproportionate share of the tokens. The majority of US companies still use US models, but the biggest consumers are switching first because the economics are most visible at their scale.

Is GLM-5.2 really better than GPT-5.5 on SWE-bench Pro?

On SWE-bench Pro specifically, yes: 62.1% vs. 58.6%. But SWE-bench Pro measures real software ticket resolution, not general capability. In overall agentic, GPT-5.5 dominates at 98.2 against unpublished scores for GLM-5.2 on this specific benchmark. The key point is that GLM-5.2 outperforms GPT-5.5 on a specific and critical benchmark, for 1/40th of the price.

Is the advisor model architecture complicated to set up?

Moderately. It requires a routing middleware (like OpenRouter itself, or an internal proxy), a confidence scoring system, and regression testing to calibrate the escalation threshold. For a team of 3-5 ML engineers, it's 2 to 4 weeks of work. ROI is typically achieved in one month of production.

Will Chinese models be banned in the United States?

Nothing indicates this to date. The US voluntary standards are not binding. The July 2026 investigation could lead to recommendations, but an outright block of open-weight models would face considerable legal hurdles (First Amendment, research freedom, lack of existing legal framework).

Which host to self-host GLM-5 or Kimi K2.6?

Any standard GPU infrastructure works. The weights are public and the MIT license does not restrict the infrastructure. For companies that want rapid deployment without managing the infra, solutions like Hostinger for web hosting coupled with dedicated cloud GPU instances do the job for POCs and small production volumes.

✅ Conclusion

46% of US enterprise tokens going to Chinese models is not a statistical anomaly: it is the predictable result of a 60 to 90% price gap at near-equivalent performance, and enterprises are doing what they have always done — follow the money. If your AI pipeline isn't testing at least one Chinese model in an advisor position by the end of the quarter, you are probably overpaying for an identical result.