From sector to sector, Chinese companies have come to dominate global industry through their ability to export mass quantities of low-cost products. The world of AI may prove harder to conquer with the same trick.

Chinese firms involved in AI, from tech giants like Alibaba and Tencent to newer model developers like Zhipu have this year had to hike prices for their services, in response to rising costs. 

As they do so, it is becoming clear that the dynamics that have helped China Inc. take the lead in industries from solar panels to electric vehicles do not always translate neatly to AI. In traditional manufacturing, economies of scale mean the more a firm produces, the lower the cost of making each additional unit becomes. In turn, the company can keep bringing down prices for customers, gaining market share while preserving its profit margins. 

But as AI consumers demand more from the technology, it is getting harder for firms to keep their costs down, with the chips that power AI systems consuming both ever more electricity, and water to keep them cool. Excessive user demand can also shorten the lifespan of the chips, meaning firms have to spend more to replace them; and it can impose higher costs on firms as they expand data centers or lease computing power from other providers. 

“The prevailing assumption was that AI might follow the same export playbook as [online retailers] Shein and Temu: compress costs, scale globally, and let price do the work,” says Ivy Yang, founder of the New York-based consultancy Wavelet Strategy. “But now the illusion of cheap AI is breaking down.”

One reason why AI costs are rising is that users are moving on from making simple queries on platforms like DeepSeek or ChatGPT. Now they are looking to use AI agents that can carry out tasks like coding, research, and workflow automation.

That kind of energy-intensive activity can raise per-user infrastructure costs by anywhere between 10 to 100 times, says Poe Zhao, a China tech analyst and founder of the Hello China Tech newsletter.

Unlike chatbots that generate one response and wait for the next prompt, agents can figure out how to complete a multi-stage task themselves. But without humans supervising each step, AI agents that lack enough data to work from often make errors, meaning they have to keep restarting tasks — compounding computing costs with every misstep. That makes AI fundamentally different from earlier e-commerce business models, says Zhao.

“In ride-hailing or food delivery, platforms could subsidize demand while betting that scale, market share, and future monetization would eventually absorb the cost,” he says. “In AI, heavy usage does not dilute costs, but intensifies them.” 

Similar issues are playing out in the United States. Anthropic, for example, recently closed a pricing loophole that had been costing the company heavily. Subscribers to its Claude AI model had been connecting it to third-party AI tools like OpenClaw, using them to run workloads that consume far more computing power than normal human-to-AI chat interactions — raising the company’s costs even as it kept subscription fees low.

That case drew wide attention in China tech circles last month after the head of Xiaomi’s large language model team, Fuli Luo, shared it on X as a warning to her peers “against racing to the bottom on pricing.”

In China, some major companies do offer heavily discounted subscription plans, but they have become harder to sign up to. Alibaba, for example, offers a monthly AI subscription for programmers costing around 200 yuan ($28.5) for up to 90,000 requests. But access to the discount is distributed only in limited daily batches that disappear within minutes. 

A token is not like a piece of furniture, where quality and price can be assessed at a glance. Intelligence only creates value when embedded into workflows, and its cost depends entirely on how it is used. The same model can be either highly efficient or wildly wasteful depending on how it is deployed.

Kyle Chan, a research fellow at the Brookings Institution

Several Chinese AI companies have begun raising prices too. Since late March, Alibaba, Baidu, and Tencent have each announced multiple price increases for their AI services, with those for some of Tencent’s offerings, such as its large language model Tencent HY2.0 Instruct, rising as much as fourfold.

LLM startups like Zhipu have meanwhile raised access costs to its models several times this year, in an example of how companies are finding it hard to keep costs down as user demand rises, rather than benefiting from economies of scale. When the company hiked its subscription price by 30 percent in February, it said it was having to invest more “in computing power and model optimization” simply to maintain its service levels. 

In the generative video space, meanwhile, ByteDance’s Seedance raised price three times over April alone. The third of those rises saw its annual subscription fee rise to 2599 yuan from 3099 yuan ($371 to $442).

Bytedance did not respond to a request for comment.

One of China’s most famous AI firms has bucked the trend, however. On April 24, DeepSeek launched its long anticipated new flagship model, pricing its V4 Pro model at $3.48 per million output tokens, a fraction of the $30 per million output tokens OpenAI charges for GPT-5.5. (A token is the basic unit of data that LLMs use to read and generate language, and has become the standard pricing unit in the AI industry).

The company also suggested that its costs could fall further once Huawei starts selling its Ascend 950 chips at scale in the second half of this year: those are the most advanced AI chips developed in China to date, and DeepSeek V4 has optimized its models to run efficiently on them. 

Yet Kyle Chan, a research fellow at the Brookings Institution, says DeepSeek may be an outlier. Unlike tech giants such as Alibaba, which have to maintain sprawling businesses ranging from cloud services and e-commerce to logistics networks, DeepSeek is focused on one thing: large language models. Also, unlike the now publicly-owned Zhipu, it faces less shareholder pressure to be profitable — it is still mostly funded by the hedge fund run by its founder Liang Wenfeng.

Chan adds that DeepSeek could gain a further cost advantage if Huawei’s Ascend 950 chips successfully ramp up production, lowering the company’s inference costs as it expands its use of domestic hardware. “There might be some players that end up leaving the market if they can’t keep up,” he says.

Chip giant Nvidia’s chief executive Jensen Huang has long pointed to the broader shift in priorities underway for AI firms — from developing the most powerful models towards controlling costs — repeatedly describing AI as a new kind of “factory”.

“The input is electrons, the output is tokens. In the middle is Nvidia,” he said recently on the Dwarkesh Podcast, adding that tokens generated by Nvidia’s chips are the cheapest in the world, thanks to “extreme optimization in code and chip design.”

Some experts say China’s experience with improving efficiency across manufacturing supply chains could yet give it an advantage in AI as it “industrializes”.

It is not just about lower costs, but the ability of firms across the supply chain to coordinate to compress costs, says Zhao.

What sets AI apart from traditional manufacturing is the value of its output is far harder to measure, however. 

“A token is not like a piece of furniture, where quality and price can be assessed at a glance,” says Chan. “Intelligence only creates value when embedded into workflows, and its cost depends entirely on how it is used. The same model can be either highly efficient or wildly wasteful depending on how it is deployed.”

---