AI Suddenly Got More Expensive.

In March 2026, Tencent Cloud raised prices for some AI models by over 400%. Alibaba Cloud also increased AI service prices for certain models on its Bailian platform twice within a month.

A friend in content creation did the math: He uses APIs to call large models and batch-generates short video scripts for a week (about 50 scripts). Charged by token consumption, the cost was about RMB 15 three months ago. In March, with many AI relay stations shut down and major cloud providers raising model prices, the same workload now costs nearly RMB 60.

In early 2024, China's average daily token calls were 100 billion. By March 2026, that number had grown to 140 trillion.

A 1,000-fold increase in two years.

What is a token?

Asking Qianwen, 'What's the weather like in Shenzhen today?' uses about 200 tokens (including the question and answer). Asking DeepSeek to 'write a 5,000-word project plan' uses about 8,000-10,000 tokens. Tokens are the basic unit of AI information processing and the billing unit for computing power. Think of it as AI's word counter—the more you use, the more computing power you burn.

So, why did AI suddenly get more expensive? Is it because providers want to raise prices?

01 The Truth Behind the Price Hikes

JPMorgan predicts that by 2030, China's AI inference token consumption will grow about 370 times compared to 2025. Goldman Sachs predicts global token consumption will grow 24 times in the same period.

The conclusion is clear: AI's computing consumption is entering exponential growth.

The reason? How we use AI has changed.

We used to ask AI questions, and with a few dozen tokens, the matter was settled. But now, more people are asking AI to 'do things for me.'

For example, 'Help me plan a 5-day trip, including flights, hotel recommendations, scenic routes, and food recommendations.' AI doesn't do this in one step—it requires multiple inferences, calls to multiple tools, and step-by-step execution. This is called an Agent task in the industry. An Agent task consumes 100-1,000 times more tokens than a simple Q&A.

From Q&A to execution, computing consumption differs by three orders of magnitude.

On the supply side, leading cloud providers like ByteDance, Alibaba, and Tencent have seen a sharp rise in AI infrastructure spending. From 2024 to 2025, spending on core AI chip components doubled from $22 billion to $52 billion. High-bandwidth memory (HBM) now accounts for about two-thirds of GPU material costs, up from about 50%. Expensive chips, expensive memory, expensive electricity—this has led to an across-the-board rise in computing costs.

So, price hikes aren't a choice by providers—demand is too strong, and capacity can't keep up.

This is very much like mobile communications around 2010. Smartphones exploded in popularity, but base stations were insufficient, making data expensive. Later, with large-scale 4G and 5G deployment, data became cheap. AI is now at the same crossroads.

Except, the ones building 'AI base stations' aren't the three major carriers—it's China's new infrastructure initiative: the computing power network.

02 The Black Hole of Computing Power

When many people think of businesses using AI, they imagine bosses who don't understand AI but demand employees use it to review reports or write plans.

That's short-sighted. The bulk of AI's computing consumption isn't there at all.

Manufacturing. Companies like Foxconn and CATL use AI for visual quality inspection—cameras capture product images, and AI instantly judges whether they're good or defective. Running 24/7, a single production line processes millions of images daily, requiring constant computing power.

Programmers. Tools like Tongyi Lingma and GitHub Copilot help developers write code and find bugs. Millions of developers in China now use AI-assisted programming daily, with each request consuming computing power.

Content industry. AI generates copy, video scripts, and product descriptions. Much of the copy behind the short videos you watch daily involves AI.

Healthcare. AI reads CT scans and detects early-stage lung cancer, with accuracy approaching or exceeding human interpretation in some scenarios.

Scientific research. AI accelerates drug discovery, drastically shortening the cycle from target identification to molecular screening from years to a fraction of the time.

Agriculture. Satellite remote sensing plus AI analysis identifies which fields lack water or have pests, guiding farmers precisely on fertilization and pesticide application.

These scenarios have one thing in common: they all consume massive computing power. And they're not niche—manufacturing quality affects every product, medical AI affects everyone's health, and agricultural AI affects food supply.

When computing power runs short, price pressures don't stop at cloud providers. They trickle down the chain to every ordinary person.

How to solve this? China's answer: build a 'high-speed rail network for computing power.'

03 The Weight of a Network

Water networks, new power grids, computing power networks, next-gen communication networks, urban underground pipelines, and logistics networks—these are the 'six networks' explicitly planned in the '15th Five-Year Plan.'

At their core, they're new infrastructure. Just as China invested heavily in high-speed rail and 4G networks 20 years ago to lower costs and barriers for society as a whole, the same logic applies now.

Why can computing power networks stand alongside power and water networks?

Because computing power has become the 'electricity' of the AI era.

A century ago, every factory generated its own electricity. Then the national grid was built, and factories could simply plug in—no need for on-site generators.

The computing power network follows the same logic. Today, every company buys its own GPUs and builds data centers—costly and inefficient. Once the computing power network is built, companies can purchase computing power on demand, like buying electricity—pay for what you use, no need to build your own power plant.

How will this network operate? The core logic is 'Data Computing from East to West.'

China has planned eight national computing hubs. Four in the east—Beijing-Tianjin-Hebei, Yangtze River Delta, Guangdong-Hong Kong-Macao Greater Bay Area, and Chengdu-Chongqing—have concentrated users and need millisecond-level responsiveness. Four in the west—Inner Mongolia, Guizhou, Gansu, and Ningxia—have cheap green energy and abundant land.

AI inference tasks requiring real-time response stay in the east. Less urgent AI training tasks are scheduled to the west, running on cheap green energy.

This scheduling isn't just theoretical—it's already happening. Zhejiang and Xinjiang use time-zone differences to 'shift peaks and fill valleys,' reducing token electricity costs by 18%. Western green data centers achieve a PUE (Power Usage Effectiveness) below 1.15, far below the national average.

In the era of large model inference, the ratio of inference computing demand to training computing demand is expected to reach 3:1. This means future computing networks will be designed more for real-time responsiveness—exactly the value of 'Data Computing from East to West.'

04 Computing That Reaches for the Stars

Computing power networks aren't just ground-based.

On May 14, 2025, China launched 12 space computing satellites aboard a Long March 2D rocket from the Jiuquan Satellite Launch Center, forming the world's first large-scale space-based intelligent computing satellite constellation.

These satellites carry 'AI brains'—high-performance intelligent computers and distributed operating systems. Traditionally, satellites had to transmit images to the ground for processing, taking hours. Computing satellites can analyze data in orbit, compressing disaster emergency response to seconds.

Early identification of wildfires and floods, precision agriculture, ocean monitoring—satellites can 'compute while flying,' transmitting only key results to the ground, drastically reducing data transmission bottlenecks.

An even more ambitious concept: building large-scale computing centers in space, turning space into 'offshore data centers.'

By late 2025, Elon Musk announced SpaceX's plan to expand the Starlink V3 satellite cluster and build space-based data centers in orbit, submitting a formal application to the FCC in early 2026 for a million-satellite network. Jeff Bezos also predicted in October 2025 that gigawatt-scale data centers would be built in space within 10-20 years. Theoretically possible, but technically still far from practical use. China's current focus is the first step: networking computing satellites.

Ground-based networks address 'insufficient computing power,' while space-based networks solve the bottleneck of massive satellite data being untransmittable or unprocessable once transmitted. These two paths, though at different stages, aim in the same direction: making computing power ubiquitous.

05 Trillion-Dollar Investments

During the '15th Five-Year Plan' period, total investment in the computing power network is expected to reach RMB 2-2.5 trillion.

In 2026 alone, total investment in the 'six networks' will exceed RMB 7 trillion, with annual investment in the computing power network no less than RMB 400 billion.

As of early 2026, the national computing interconnection platform has connected 578 resource pools from 155 enterprises across 31 provinces (autonomous regions, municipalities), including 316 EFLOPS of intelligent computing resources and 720,000 GPU accelerator cards.

Funding sources are clear. In 2026, the central government will allocate over RMB 2.55 trillion in policy-based funds to support the 'six networks,' with over RMB 800 billion in ultra-long-term special bonds available for the computing power network, featuring 20-50 year terms, low interest rates, and no increase in local hidden debt.

Market-driven funds are also flowing in. ByteDance, Alibaba, Tencent, Baidu, and the three major carriers are all increasing investments related to the computing power network.

Signs of inclusivity are emerging. An intelligent technology company in Nantong successfully applied for a RMB 15,000 computing voucher subsidy, with models like 'computing power banks' and 'computing power supermarkets' being piloted in multiple regions.

All this points to one direction: continuously lowering AI computing costs to make AI accessible to more people.

06 Silent Inclusivity

What will change in your life once the computing power network is built?

First, your cost of using AI will drop.

Today, when you subscribe to Doubao Premium or use Qianwen's paid version, the fees reflect computing power prices. Once the computing power network is built, AI service prices will likely continue to decline, just as mobile data evolved from RMB 10/MB to today's unlimited plans.

Second, more people will be able to start AI businesses.

Today, launching an AI application requires buying dozens of GPUs, with upfront costs often in the millions. Once the computing power network is widespread, 'computing vouchers' and pay-as-you-go models will drastically lower barriers. With a good idea, you won't need to raise funds for equipment first.

Third, more AI capabilities will appear where you don't notice them.

Precision agriculture increases grain yields, stabilizing food prices. AI healthcare makes early screening widespread, lowering medical exam costs. Intelligent transportation reduces congestion, shortening commutes. These won't be branded as 'AI,' but you'll feel life improving.

China's goal in building this network isn't just to solve 'insufficient computing power.' We must also recognize that AI shouldn't be a weapon only for big companies—it should be infrastructure accessible to everyone, like electricity.

Today's 140 trillion tokens may become 1,400 trillion when the computing high-speed rail is complete. And by then, the cost of using AI may be just a fraction of today's.