The number making waves in tech this week is 46%. That's the share of enterprise AI traffic from US companies that flowed to Chinese-made models during a peak week in mid-2026. Depending on which headline you read, it either signals America losing the AI race or proves that price always wins in a commodity market.
Neither of those framings tells the full story. Here's what the actual data shows, what it doesn't, and the part of this story that most coverage isn't explaining — because it makes the picture considerably more complicated.
The Data, Exactly as It Stands
💡 Source: A CNBC investigation published July 7, 2026, drawing on platform-level usage data from OpenRouter — an API aggregation service that lets developers route requests across 400-plus AI models from more than 60 providers. The data reflects actual tokens processed in production, not surveys or downloads.
Here's what the numbers actually say:
- Chinese-origin AI models have accounted for at least 30% of enterprise token volume on OpenRouter every single week since February 8, 2026
- The peak week hit 46%
- The prior 12-month average was 11%
- In the first half of 2025, Chinese models held just 4.5% of OpenRouter traffic
- The crossover moment arrived the week of February 9-15, when Chinese models out-consumed US models on the platform for the first time
The provider breakdown within that Chinese share is specific. DeepSeek holds 17.6% of OpenRouter's routed tokens, amounting to 5.13 trillion weekly — the single largest vendor on the platform. Alibaba's Qwen follows at 13.9% with 2.77 trillion tokens weekly. Z.ai's GLM-5.2 recorded the fastest new-model adoption on Vercel's infrastructure in 2026, with daily token volume growing roughly 27-fold and customer count growing roughly 80-fold in its first full week of availability.
Why This Is Happening — The Price Gap Is Staggering
"Price is doing the work here," Harpreet Arora, head of agentic infrastructure at Vercel, told CNBC. "When a task doesn't need the best model, teams are beginning to route it to the cheapest one that's good enough, and the recent wave of models coming out of China is winning that trade."
The price gap he's describing isn't close:
| Model | Price (per 1M input tokens) | Origin |
|---|---|---|
| DeepSeek V4 Flash | $0.14 | China |
| GLM-5.2 (Z.ai) | ~$0.50 | China |
| GPT-5.5 (OpenAI) | $5.00 | USA |
| Claude Sonnet 5 (Anthropic) | $3.00 (from Sept 1) | USA |
Run those numbers across a full production workload and the gap gets brutal. One analysis tracking full-workload costs puts Claude at around $4,811 and GLM-5.2 at about $544 for the same job — roughly a 9x cost difference. For a startup burning six figures a month on AI tokens, that's not a rounding error — it's runway.
Real companies are acting on it. AI startup Lindy moved 100% of its traffic from Anthropic's Claude to DeepSeek and told CNBC it will save millions. Coinbase runs 1,200 AI agents on Chinese models and cut its AI spend in half. Airbnb uses Chinese models through US-based service providers for non-sensitive workloads. Uber burned through its entire annual AI budget in just four months of 2026, prompting the company to implement $1,500-per-engineer-per-month spending caps.
The Capability Gap Has Narrowed Enough to Matter
Price alone doesn't explain the shift — the models have to be good enough for the work. A year ago, Chinese models were meaningfully weaker than US frontier models on most benchmarks. That's no longer consistently true.
GLM-5.2 from Z.ai scored 62.1% on SWE-bench Pro, a widely-used coding benchmark — above GPT-5.5's 58.6% and within one percentage point of Anthropic's Opus 4.8 on the same test. US-built models still lead on vision, multimodal, regulated, and national-security workloads — but for the bulk of what most enterprises actually do with AI day-to-day (bulk classification, drafting, code scaffolding, internal tooling), the performance gap has closed enough that the cost difference becomes the deciding factor.
The Detail Most Coverage Is Missing: The 46% Number Is More Complicated Than It Looks
⚠️ The 46% figure is real, but it tells a different story depending on how you count. That number represents tokens in a peak week, through OpenRouter — a platform specifically designed to help developers route to the cheapest available model, which means it attracts a price-sensitive user base that isn't representative of all enterprise AI usage. Measured by revenue rather than token volume, the picture looks very different. Anthropic reportedly holds about 65% of Vercel's AI spend even while holding only 14.8% of its token volume. US labs are charging more per token and winning on higher-value workloads — Chinese models are winning on volume, not necessarily on the work that matters most. About 96% of inference teams still keep OpenAI or Anthropic somewhere in their stack. This is model routing, not model replacement.
The Fable 5 Shutdown Made This Worse — Accidentally
One development directly accelerated the Chinese model surge that isn't getting enough attention in the coverage: the June 12 export control order that suspended Anthropic's Fable 5 and Mythos 5 models for 18 days.
When the US government ordered Anthropic to cut off global access to its flagship models with no transition period, enterprises that had built workflows around those models had to find alternatives overnight. Nikkei Asia reported that Chinese AI usage by US firms surged in the immediate aftermath. The policy created the very outcome it was presumably designed to prevent: it pushed companies toward less-regulated, non-US alternatives at exactly the moment when those alternatives had become competitive enough to absorb the workload.
The models were restored on July 1. But the data shows that at least some of the users who switched during that window didn't switch back.
Now Both Governments Are Moving at Once — In Opposite Directions
This is the part of the story that most coverage is burying in a single paragraph, and it's actually the most consequential development of the week.
On the same day CNBC published its 46% investigation — July 7 — Reuters reported that China's Ministry of Commerce was in talks with Alibaba, ByteDance, and Z.ai about restricting overseas access to their most advanced models. On July 8, US lawmakers confirmed they are investigating Airbnb and Cursor's parent company Anysphere over their use of Chinese AI.
Both governments are now moving to control which AI models their companies can use — from opposite directions. The US is pressuring companies that use Chinese models. China is considering restricting its own models' export. If both moves proceed, the developer market that has spent six months optimizing its stack around cheap Chinese open-weight models could find itself squeezed from both ends simultaneously, with no transition period on either side.
✅ The practical question for any company using AI: are you building single-model dependencies into your production stack, or are you routing across providers in a way that lets you swap models when regulatory or geopolitical conditions change? Airbnb's approach — routing Chinese models through US-based service providers for non-sensitive workloads, while keeping US models for customer-facing and regulated data — is the version of this answer that multiple lawyers and security teams have already signed off on. It's not perfect, but it's a model for how to use cost savings without betting the entire stack on access that two governments are actively trying to control.
What This Actually Means for the AI Industry's Biggest Bets
OpenAI and Anthropic both filed for IPOs in June 2026, at reported valuations of $852 billion and $965 billion respectively. Those numbers assume both companies hold pricing power and market share over time. The 46% data point challenges the pricing-power assumption directly.
CNBC ran the explicit warning in May: "Cheap AI could derail OpenAI and Anthropic's IPOs." OpenAI is on track for an operating loss close to $14 billion in 2026, with profitability not expected until the end of the decade. When your mid-market is being undercut 36 to 55-fold on price, and performance parity is arriving faster than expected, a private valuation in the hundreds of billions starts to rest on a narrower foundation than it looked like a year ago.
Frequently Asked Questions
What does "46% of US enterprise AI traffic goes to Chinese models" actually mean?
It means that during a peak week in mid-2026, 46% of the tokens routed by US companies through OpenRouter — an API aggregation platform — went to AI models built by Chinese companies including DeepSeek, Alibaba Qwen, and Z.ai's GLM. It's a token-volume figure, not a revenue or market share figure, and it reflects OpenRouter's price-sensitive user base more than all enterprise AI usage.
Why are US companies using Chinese AI models?
Primarily cost. Chinese open-weight models are 60% to 90% cheaper per token than leading US models, according to OpenRouter's data team. For high-volume workloads — agent loops, bulk classification, code scaffolding, internal tools — the cost difference can translate to millions of dollars annually. The performance gap has also narrowed: GLM-5.2 outscores GPT-5.5 on at least one major coding benchmark.
Is it legal or safe to use Chinese AI models in the US?
Currently legal for most commercial applications, but the regulatory picture is shifting. Congress is actively investigating companies that use Chinese AI, and the US government has already shown it will restrict access to specific models (Fable 5, Mythos 5) when it deems them a security risk. Data handling is the primary concern — Chinese models accessed through US-based API gateways with appropriate data handling agreements represent a lower-risk profile than direct API calls to Chinese servers.
Are US AI labs actually losing to Chinese AI?
On token volume for commodity workloads, yes. On revenue, not yet — US labs still capture a majority of enterprise AI spending because they dominate high-value, customer-facing, and regulated workloads. About 96% of teams using Chinese models still keep OpenAI or Anthropic somewhere in their stack. This is model routing for cost optimization, not wholesale replacement.
Which Chinese AI models are most popular with US companies?
DeepSeek leads in raw token volume on OpenRouter at 17.6% of all routed tokens. Alibaba's Qwen follows at 13.9%. Z.ai's GLM-5.2 recorded the fastest adoption growth of any model on Vercel in 2026, with 27x daily token volume and 80x customer growth in its first full week of availability.
Last updated: July 12, 2026. Data in this article sourced from CNBC's July 7, 2026 investigation, OpenRouter platform analytics quoted by CNBC, Financial Times reporting cited by Dealroom, and Reuters' July 7 reporting on China's Ministry of Commerce review. Company statements (Airbnb, Coinbase, Lindy) sourced from CNBC. Expert quotes attributed to Harpreet Arora (Vercel), Justin Summerville (OpenRouter), and Vercel CEO Guillermo Rauch.
0 Comments