A quiet but consequential shift is underway in the economics of artificial intelligence: Chinese laboratories have released models that rival Western counterparts in capability while costing a fraction as much, arriving precisely as the industry pivots toward agentic AI—systems that act, not merely answer—where token consumption multiplies costs by thousands. The timing is not accidental. As Silicon Valley's subsidized pricing model strains under the weight of this new demand, and as China courts the Global South with affordable AI infrastructure, what began as a pricing competition is becomin
Chinese AI models poised to dominate agentic AI market with aggressive pricing
At that scale, even small cost differences become meaningful budget gaps.
Mark
Why does the cost per token matter so much more for agentic AI than for regular chatbots?
Mimi
Because agents don't just answer one question. They work on a task for days, making thousands of decisions, each one consuming tokens. A chatbot might use 2,000 tokens for a conversation. An agent might use 100 million for a single project. When you multiply that by the price difference—30 times cheaper or more—you're talking about thousands of dollars in savings per task.
Mark
So it's not really about which model is smartest anymore?
Mimi
Not for most customers. A model that's 90 percent as intelligent but costs one-tenth as much is the obvious choice when you're running agents at scale. The intelligence gap has to be significant enough to justify the cost. Right now, Chinese models are closing that gap while staying dramatically cheaper.
Mark
What's stopping Western companies from just lowering their prices?
Mimi
They've been subsidizing prices with venture capital for years. They can't sustain that if token usage explodes. Chinese companies are under pressure too—their computing costs are rising—but they're willing to operate on thinner margins to capture market share. It's a race to the bottom that Western companies may not be able to win.
Mark
Is this really about technology, or is it about who can afford to lose money longer?
Mimi
Both. China is also making a strategic bet on the Global South. If developing countries build their AI infrastructure on cheap Chinese models, those countries become dependent on Chinese technology. That's leverage that extends far beyond price.
Mark
Can Western companies catch up?
Mimi
They're trying. Meta and OpenAI are releasing cheaper models. Nvidia is building hardware that runs agents locally without per-token costs. But they're playing catch-up. Chinese models are already good enough and already cheap. The window to compete on price alone may be closing.
El Pulso
- The cost gap between Chinese and Western AI models is not marginal—it is existential for agentic workloads, where a single task can cost one-thirty-fourth as much on DeepSeek as on OpenAI or Anthropic.
- Major Western institutions are already defecting: Microsoft dropped Anthropic's Claude Code and is exploring DeepSeek to power its own agentic office assistant.
- Western AI companies are structurally exposed, having subsidized user prices with venture capital they cannot sustain as token consumption scales into the billions per day.
- Chinese models still carry reliability weaknesses—higher token burn on complex tasks, lower task-completion rates—but the newest releases are closing that gap rapidly.
- The West is responding with cheaper models and local hardware, but the window to compete on price in the Global South may already be narrowing as China embeds its AI infrastructure across developing-world institutions.
A quiet but consequential shift is underway in the economics of artificial intelligence: Chinese laboratories have released models that rival Western counterparts in capability while costing a fraction as much, arriving precisely as the industry pivots toward agentic AI—systems that act, not merely answer—where token consumption multiplies costs by thousands. The timing is not accidental. As Silicon Valley's subsidized pricing model strains under the weight of this new demand, and as China courts the Global South with affordable AI infrastructure, what began as a pricing competition is becoming a contest over who will shape the technological foundations of the next era of human organization.
The economics of artificial intelligence are shifting faster than Silicon Valley anticipated. In the past month, two Chinese labs—Z.ai and Moonshot—released models that perform nearly as well as OpenAI or Anthropic's best offerings, but at a fraction of the price. Silicon Valley start-ups are already running them.
The deeper disruption lies not in chatbots but in what comes next: agentic AI—systems that draft emails, conduct research, write code, and manage tasks autonomously over days. These agents consume tokens at a staggering rate. Where a simple exchange burns a few thousand tokens, a single agentic task can run into the millions. When researchers benchmarked 657 agentic tasks, Anthropic's flagship model cost nearly $1,000 to complete them. Z.ai's GLM-5.2 cost $270. DeepSeek, after a 75 percent price cut, completed the same suite for $14.
That arithmetic is already reshaping institutional decisions. Microsoft canceled its Anthropic coding subscription partly over token costs and is now exploring DeepSeek for its agentic office assistant. An AI research firm switched to Xiaomi's MiMo-2.5-Pro simply because its founder burns over 100 million tokens daily and the savings compound into meaningful budget gaps.
Western companies face a structural bind: their subsidized pricing, backed by venture capital and sovereign wealth funds, cannot survive the token volumes that agentic AI demands. Chinese firms face their own pressures—domestic chip shortages pushed cloud computing prices up 30 percent in ten days earlier this year—yet they continue undercutting Western prices and offering free tiers to capture users, even at a loss.
Chinese models still have weaknesses. GLM-5.2 burns more tokens on complex tasks than its Western rivals, and DeepSeek's flash model underperformed on task completion. But Moonshot's Kimi-K3, released last week, completed more tasks correctly than any model tested—at roughly one-third the cost of Anthropic's equivalent. The reliability gap is closing.
The West is responding: Meta released a model at $4.25 per million tokens, undercutting Anthropic; OpenAI launched a cost-efficient variant of GPT-5.6; Nvidia is preparing a $2,000 laptop that runs agentic workflows locally, bypassing per-token fees entirely. But that laptop price point is within reach of Western enterprises, not of developing nations.
There lies the geopolitical dimension. For governments across the Global South, the cost of running AI agents is not merely high—it is prohibitive. China has recognized this opening. At last week's World AI Conference in Shanghai, President Xi announced cooperation centers spanning nearly the entire Global South. If Chinese models become the foundation of AI ecosystems across the developing world, the consequences will extend far beyond any pricing spreadsheet.
The economics of artificial intelligence are shifting beneath our feet, and the shift is happening faster than Silicon Valley expected. Over the past month, two Chinese AI labs—Z.ai and Moonshot—have released models that perform nearly as well as anything OpenAI or Anthropic has built, but at a fraction of the cost. The models are already in use. Silicon Valley start-ups are running Z.ai's GLM-5.2. Moonshot's Kimi-K3 won't be far behind.
The real story isn't about raw intelligence. It's about what comes next: a world where artificial intelligence doesn't just chat with you, but acts on your behalf. These agents—systems that can draft emails, conduct research, write code, manage tasks across days—are exponentially more expensive to run than a simple chatbot. They consume tokens at a staggering rate. A token is the unit by which AI companies meter usage, roughly equivalent to a word. A standard question-and-answer exchange might burn a couple thousand tokens. A single agentic task can run into the millions. Some users are already burning tens or hundreds of millions of tokens per day. When the cost difference between models is measured in the thousands of dollars per task, price stops being a luxury concern and becomes the deciding factor.
Z.ai charges $1.92 per million output tokens for GLM-5.2. Anthropic charges $25 per million for its Opus 4.8. The gap widens when you measure real-world usage. Researchers at MIT, Stanford, and Google DeepMind estimate that agentic tasks consume up to 3,500 times as many tokens as a simple reasoning problem. When Artificial Analysis ran 657 agentic tasks—the kind of office and administrative work that companies might soon automate—Anthropic's Opus 4.8 cost nearly $1,000. Z.ai's model cost $270. DeepSeek, which announced a 75 percent price cut to its tokens, completed the same suite of tasks for $14. The math is brutal. An agent powered by DeepSeek costs as little as one thirty-fourth the price of an OpenAI or Anthropic equivalent.
This arithmetic is already reshaping decisions at major institutions. Microsoft recently canceled its subscription to Claude Code, the coding assistant from Anthropic, partly because of token costs. The company is now exploring DeepSeek to power Copilot Cowork, its agentic AI office assistant. Azeem Azhar, an AI agent pioneer who runs a research firm augmented by these systems, switched to MiMo-2.5-Pro, a model from Chinese conglomerate Xiaomi, because he burns more than 100 million tokens daily. The savings compound. At that scale, even small cost differences become meaningful budget gaps.
Western AI companies face a structural problem. OpenAI and Anthropic have heavily subsidized their user subscriptions, relying on venture capital and sovereign wealth funds to absorb the losses. They cannot sustain those prices if the world moves toward agentic AI. Chinese companies face their own pressures—rising computing costs as demand outstrips domestic chip supply, and intense market competition pushing them to cut prices further. In March, cloud providers including Alibaba, Huawei, and Tencent all raised computing prices by roughly 30 percent in ten days. Yet Chinese AI enterprises continue to undercut Western pricing and offer free service schemes to gain users. In 2025, SiliconFlow spent 24 percent more on marketing than its entire revenue.
The Chinese models do have weaknesses. GLM-5.2 burns more tokens on complex tasks than Western equivalents, eroding some of the price advantage. DeepSeek-V4 Flash performed poorly on Artificial Analysis's tests, completing fewer tasks correctly than competitors. But this gap is closing. Kimi-K3, released last week, completed more tasks correctly than any other model tested—at roughly one-third the cost of Anthropic's Fable 5. Once a Chinese model combines affordability with reliable performance, Western AI companies will face a genuine crisis.
Western responses are emerging. Meta announced on July 9 that its new model, Muse Spark 1.1, would cost $4.25 per million output tokens, undercutting Anthropic and OpenAI. OpenAI has released Luna, a cost-efficient version of GPT-5.6 designed to run cheaply on agentic tasks, though it sacrifices some performance—Kimi-K3 scores nearly 10 percentage points higher on Artificial Analysis's tests. Nvidia is positioning itself to reduce agent costs by releasing a laptop that runs agentic workflows locally, without paying per-token fees through an API. The price is expected to be around $2,000, within reach of Western enterprises but not of developing nations.
This is where geopolitics enters. For governments in the Global South, the cost of running AI agents is not merely expensive—it is prohibitive. China's leadership has recognized this. At the World AI Conference in Shanghai last week, President Xi Jinping announced that China would launch AI application cooperation centers within six regional intergovernmental organizations covering nearly the entire Global South. An op-ed in the People's Daily last year argued that the low cost and open weights of Chinese AI made it more accessible to developing countries against 'hegemonic' Western AI. If Chinese models become the foundation of AI ecosystems across the developing world, the geopolitical consequences will extend far beyond price.
Citas Notables
At the scale agents operate, even small cost differences compound into meaningful budget gaps.— Azeem Azhar, AI agent pioneer
China would launch AI application cooperation centres within six regional intergovernmental organisations that together cover almost the whole Global South.— President Xi Jinping, World AI Conference Shanghai
Quieres la nota completa?
Lee el original en The Strategist | ASPI's analysis and commentary site ↗
Technology