斜杠中年斜杠中年AI × 沟通 × 商业 × 人生 × 健康
AI Practical Guide

Chinese AI Models Now Consume 8x More Tokens Than US Models Globally

2026-11-19Updated: 2026-11-1910 min readWesley Chong
#DeepSeek#Chinese AI#OpenRouter#GLM#MiniMax#global AI usage
Chinese AI Models Now Consume 8x More Tokens Than US Models Globally|AI Practical Guide 封面图

Summary

The numbers are no longer subtle. According to OpenRouter data analyzed by Chinese financial media, the week of November 9–15, 2026 saw Chinese AI models consume 119.1 trillion tokens globally, while US models consumed 14.83 trillion. That is an 8:1 ratio, and it has held for 29 consecutive weeks.

This is not a spike. It is a structural shift.

The Global Usage Rankings

For the week of November 9–15, the top 5 most-used models on OpenRouter were:

  1. DeepSeek V4.1 Flash (China) — 59.2 trillion tokens, #1 for four consecutive weeks, +6% week-over-week
  2. Tencent Hunyuan Hy4 (China) — 8.97 trillion tokens
  3. MiniMax M3.1 Flash (China) — 13.7 trillion tokens
  4. Zhipu GLM-5.3 Flash (China) — 13.2 trillion tokens
  5. (US model)

Four of the top five are Chinese. DeepSeek alone consumed more tokens than the entire US model ecosystem combined — and then some.

The total global token volume for the week was 211 trillion, up 0.48% from the prior week. Chinese models accounted for 56.4% of all tracked usage.

Why This Is Happening

The answer comes down to three factors, in order of importance:

1. Price. DeepSeek V4.1 Flash charges approximately $0.15 per million input tokens at off-peak rates, with cache hits as low as $0.003 per million. For comparison, GPT-6 Astra costs around $15 per million input tokens — 100x more. When a model is 100x cheaper and 80% as good, the economics are brutal.

2. Quality convergence. These are no longer toy models. Kimi K3.1 Pro, with 3 trillion parameters, ranks first among Chinese models on the Artificial Analysis intelligence index and approaches the level of top closed-source models. DeepSeek V4.1 Flash and Zhipu GLM-5.3 Flash are competitive on coding, reasoning, and tool-use benchmarks.

3. Open-source availability. Most of these models publish their weights. Developers can self-host, fine-tune, and build products without per-query fees. The open-source strategy has turned Chinese models into the default infrastructure layer for cost-sensitive applications worldwide.

The Enterprise Spending Story

The usage data aligns with spending data. Menlo Ventures' State of Enterprise AI 2026 report found that 89% of US enterprise open-source model spending flows to Chinese models:

  • Kimi (Moonshot AI): 61%
  • DeepSeek: 15%
  • Zhipu (GLM): 13%

Total global enterprise AI spending in 2026 is projected at $176.8 billion. The fact that nearly nine in ten open-source dollars from US companies go to Chinese models tells you where the value proposition sits.

This is not a geopolitical story. It is an economic one. When a product is dramatically cheaper and good enough, procurement follows the price.

The Revenue-Sharing Shift

A related development: Kimi K3.1 Pro officially launched on Microsoft Foundry on November 17, 2026 — the first time Microsoft Cloud has enabled a revenue-sharing model for an open-weight model. Previously, cloud providers hosted open-weight models and earned compute margins without sharing revenue with the model creators. Kimi became the second Chinese model company (after AWS) to negotiate revenue sharing with a global cloud giant.

This signals that Chinese model companies are no longer content to be free infrastructure. They want a cut of the downstream revenue — and the cloud giants are agreeing because the usage volume is too large to ignore.

The Capacity Constraint

The demand is so high that it is creating supply problems. DeepSeek began limiting its App and web version usage in mid-November: long conversations inactive for 7+ days are no longer supported, single input text was shortened to approximately 40,000 characters, and daily conversation quotas were capped. Users hitting the limit see: "Server busy, usage quota reached."

This is a telling signal. When a model company has to throttle free consumer access because compute cannot keep up with demand, you know the usage numbers are real, not inflated.

What This Means for You

If you are building AI-powered products or evaluating model providers:

  1. Test Chinese models. If you have not evaluated DeepSeek, Qwen, or Kimi in the last three months, your cost benchmark is stale. The quality has moved fast.

  2. Optimize for cache hits. The gap between standard pricing and cache-hit pricing is enormous (DeepSeek: $0.15 vs $0.003 per million tokens). Prompt caching is no longer optional — it is the difference between sustainable and unsustainable unit economics.

  3. Watch the revenue-sharing trend. As Chinese model companies negotiate revenue shares with cloud platforms, the hosting economics will shift. Self-hosting open weights may become more attractive than paying cloud margins.

  4. Plan for capacity limits. DeepSeek's throttling is a reminder that cheap models have real compute constraints. Build fallback routing into your architecture — if one model is rate-limited, your system should gracefully switch to another.

The global AI landscape in late 2026 is not a duopoly between OpenAI and Anthropic. It is a multi-polar market where Chinese models dominate on volume and cost, US models lead on frontier capability, and the gap between the two is narrowing every month. The last click — the decision about which model to use for which task — still belongs to a person. But that person now has more options than ever, and ignoring the Chinese side of the menu is leaving money on the table.

FAQs

What does the 8:1 token usage ratio mean?

In the week of November 9-15, 2026, Chinese AI models on OpenRouter consumed 119.1 trillion tokens while US models consumed 14.83 trillion. This ratio has held for 29 consecutive weeks, meaning Chinese models are being used approximately 8 times more by global developers.

Which Chinese models are most used globally?

DeepSeek V4.1 Flash ranked #1 for four consecutive weeks with 59.2 trillion tokens. The top 5 included four Chinese models: DeepSeek V4.1 Flash, Tencent Hunyuan Hy4, MiniMax M3.1 Flash, and Zhipu GLM-5.3 Flash.

Why are Chinese models getting so much usage?

Extremely low pricing is the primary driver. DeepSeek V4.1 Flash costs roughly $0.15 per million input tokens, with cache hits as low as $0.003. Combined with competitive quality, this makes them the default choice for cost-conscious developers worldwide.

Are US enterprises actually paying for Chinese models?

Yes. Menlo Ventures reported that 89% of US enterprise open-source model spending flows to Chinese models: Kimi (61%), DeepSeek (15%), and Zhipu (13%). Total global enterprise AI spending in 2026 is projected at $176.8 billion.

分享这篇文章 / Share Article
Wesley Chong

Author

Wesley Chong

Software developer, digital consultant, and Toastmasters speaker from Kluang, Malaysia.

Sharing practical ideas for AI, communication, business, personal growth, and healthy living.

Related Reading