A year after DeepSeek’s R1 model sent global tech stocks into freefall, the real story isn’t another breakthrough model. It’s that 1.4 billion people have become the world’s largest AI testing ground — and what they’re learning is reshaping how AI gets deployed everywhere.

China’s daily AI token usage hit 140 trillion in March 2026, up from 100 billion at the start of 2024 — a 1,000-fold increase in two years, according to data published by China’s National Data Administration at the China Development Forum. On OpenRouter, the world’s largest AI model API aggregation platform, Chinese models accounted for 61% of total token consumption among the top ten models in February 2026.

This isn’t a Western company building a product and hoping people use it. This is a nation turning AI into infrastructure.

🔍 THE BOTTOM LINE

China isn’t just competing with Silicon Valley on model benchmarks. It’s running a parallel AI civilisation at a scale the West hasn’t matched — and the usage patterns emerging from 1.4 billion people will define how AI works for everyone, whether we like it or not.

The Numbers That Matter

• 140 trillion daily tokens — China’s AI usage in March 2026 (National Data Administration)
• 61% of OpenRouter’s top-10 token volume — Chinese models in February 2026
• 5.16 trillion vs 2.7 trillion tokens — Chinese vs US models on OpenRouter, week of Feb 16-22
• 4 of top 5 most-used models globally — were Chinese (OpenRouter, February 2026)
• 7% annual growth — China’s pledged AI spending increase (government commitment)
• 155.2 million weekly active users — ByteDance’s Doubao chatbot alone (QuestMobile)
• 100 million+ monthly active users — Alibaba’s Qwen

If you think AI adoption is accelerating in the West, look at those numbers again. ByteDance’s Doubao has more weekly active users than the entire population of Japan. And it’s one app in one market.

Not Just DeepSeek — A Seven-Front Operation

The most common Western mistake is treating China’s AI story as a DeepSeek narrative. DeepSeek matters — its R1 model proved you could train frontier AI for roughly $6 million instead of $100 million, using mixture-of-experts architecture to activate only 37 billion of 671 billion parameters per inference. But the ecosystem is far bigger than one lab.

Alibaba’s Qwen has generated over 100,000 derivative models on Hugging Face — the largest open-weight ecosystem on the platform, surpassing Meta’s Llama. Even US companies are adopting Qwen: Airbnb uses it for customer service chatbots, drawn by comparable capability at lower cost. Qwen’s app now integrates shopping, food ordering, and payments directly into the chatbot interface. You don’t leave the AI to buy something — the AI is the shop.

ByteDance controls the most popular AI app in China. Its Doubao chatbot’s 155 million weekly users dwarf anything in the West. ByteDance is simultaneously a model developer, a distribution platform (through TikTok/Douyin), and an infrastructure provider — investing $8.8 billion in regional data centre infrastructure across Southeast Asia. No Western AI company has that kind of vertical integration.

Baidu pivoted from closed-source to open after DeepSeek forced the issue. CEO Robin Li had argued closed models would dominate; within weeks of R1’s release, Baidu opened portions of its own models. Its Qingduo creative platform went from producing roughly 20 ad creatives per hour to more than 2,000 after integrating DeepSeek technology.

MiniMax listed on the Hong Kong Stock Exchange in January 2026, share price doubling on day one. Its M2.5 model topped OpenRouter’s weekly call volume within a week of release — contributing 1.44 trillion tokens in a single week.

Moonshot AI’s Kimi K2.5 can dispatch up to 100 “agent avatars” to work in parallel, increasing complex task processing efficiency by three to ten times. Kimi’s cumulative revenue in less than a month after K2.5’s release exceeded its total earnings for all of 2025.

That’s not a one-horse race. It’s not even a two-horse race. It’s a field with enough players that competitive dynamics themselves drive rapid improvement — the same dynamic that propelled Silicon Valley in the 2010s, but at China’s scale.

The Cost Advantage Is Structural

Chinese AI models run at roughly one-sixth to one-quarter the cost of comparable American systems, according to a RAND report published in early 2026. DeepSeek’s API pricing sits at approximately $0.028 per million tokens — roughly 1/180th of equivalent Western models.

This isn’t because Chinese companies are running a loss leader. It’s structural: cheaper engineering talent, government-subsidised compute, ruthless efficiency driven by chip restrictions that forced innovation in training methodology, and a domestic market large enough to amortise fixed costs across hundreds of millions of users.

The chip restrictions backfired, but not in the way hawks predicted. Instead of crippling China’s AI capabilities, they forced the ecosystem to get efficient. DeepSeek’s entire approach — mixture-of-experts, reduced precision training, multi-round training loops — was born from having to do more with less. Now that “more with less” is a competitive advantage.

Citizens Queue for AI Assistants

The AP reports that Chinese citizens are physically queuing outside tech companies to install AI assistants on their phones. Government leaders have pledged 7% annual growth in AI spending. This is state-coordinated, mass-adoption AI in a way the West simply doesn’t do.

In China, AI isn’t a productivity tool for knowledge workers. It’s consumer infrastructure — embedded in shopping, government services, education, healthcare. When your WeChat-equivalent has a native AI assistant that can book appointments, order food, file taxes, and help your kid with homework, adoption isn’t a choice. It’s just how things work.

This matters because usage at scale generates data that improves models, which attracts more users, which generates more data. The flywheel is spinning in China in a way it isn’t anywhere else — not because Chinese models are necessarily smarter, but because they’re embedded in daily life for a billion-plus people.

Why This Matters for NZ and the West

Three implications that don’t get enough attention:

1. The “AI race” framing is wrong. The West frames competition as who builds the smartest model. China is competing on who uses AI the most. By volume, China already wins. The question isn’t whether GPT-5 beats DeepSeek V4 on benchmarks — it’s whether a model used by 155 million people weekly generates more practical capability than a model used by 10 million people who pay $20/month.

2. Chinese AI is already in Western products. Airbnb uses Qwen. OpenRouter data shows Western developers routing to Chinese models for cost reasons. The idea that you can regulate Chinese AI out of your ecosystem is already false — it’s in your supply chain.

3. NZ isn’t in this conversation at all. While China coordinates national AI strategy and the US scrambles to respond, NZ has no equivalent framework. When Norway joined the Pax Silica AI supply chain alliance, we noted NZ wasn’t in the room. We’re still not. The AI adoption patterns being established right now — in both the US and China — will become the default infrastructure that small nations inherit, not negotiate.

The Surveillance Question

Any honest account of China’s AI adoption has to acknowledge the context. This is mass AI adoption in a surveillance state. The same AI assistants queuing citizens are enthusiastic about are integrated with a social credit system and government monitoring infrastructure. The data flywheel that makes Chinese AI better also makes the state’s surveillance capabilities better.

That’s not a reason to dismiss what China has achieved technically. But it is a reason to be clear-eyed about the trade-offs. When your AI assistant can file your taxes, it also knows your income. When it books your healthcare, it knows your medical history. In China, that data flows to the state as easily as it flows to the model.

Western AI adoption may be slower and more fragmented, but the privacy and governance frameworks — imperfect as they are — create a different relationship between citizen and AI. Whether that’s enough to matter is an open question.

❓ Frequently Asked Questions

Q: Are Chinese AI models actually better than Western ones? Not consistently. They’re cheaper and more widely used, which creates different advantages. On pure benchmarks, Western frontier models still lead in many areas. But the gap is narrowing fast, and cost-competitiveness means Chinese models win on value-for-money in most commercial applications.

Q: What does this mean for NZ? NZ risks becoming a passive consumer of AI infrastructure designed elsewhere. The adoption patterns being set now — by both the US and China — will become the defaults that small nations inherit. NZ needs its own AI strategy, not just reliance on whichever superpower’s tools are cheapest.

Q: Should NZ companies use Chinese AI models? Some already do, through platforms like OpenRouter. The cost advantage is real. But companies need to consider data sovereignty, the US-China tech Cold War (using Chinese models could violate future regulations), and the fact that “free or cheap” often means you’re the product.

Q: What happened to DeepSeek’s R2 model? R2, the anticipated successor to R1, has been delayed. Reports cite training difficulties on Huawei Ascend hardware — illustrating the tension between China’s chip self-sufficiency push and the practical reality that switching from NVIDIA to domestic chips isn’t seamless.

🔍 THE BOTTOM LINE

China isn’t trying to out-Silicon-Valley Silicon Valley. It’s building a parallel AI civilisation where 1.4 billion people use AI as infrastructure, not as a subscription. The usage data coming out of that experiment will shape how AI works for everyone — including us. Ignoring it because we don’t like the political system doesn’t make it less real.

SOURCES

• AP News — “China’s mass use of AI is shaping its global reach” (May 6, 2026)
• Boston Herald — “The rapid embrace of AI in China may shape how it is used globally” (May 6, 2026)
• DigitalInAsia — “Inside China’s AI Machine: Models, Chips, and Strategy” (April 6, 2026)
• China National Data Administration — Token usage data, China Development Forum (March 2026)
• OpenRouter — Model usage statistics (February 2026)
• QuestMobile — ByteDance Doubao weekly active users data
• RAND Corporation — Chinese AI cost advantage report (2026)