Artificial intelligence adoption accelerated sharply across enterprises in 2026, but soaring usage costs are prompting many businesses to rethink their technology choices.
This has driven growing interest in Chinese-developed AI models that offer comparable capabilities at a fraction of the price.
As companies expand AI deployments across software development, customer service and workflow automation, many are discovering that usage-based pricing has made AI bills far more unpredictable than initially expected.
While the price of individual AI tokens has generally declined, the overall cost of completing increasingly complex tasks has risen as AI providers move away from flat subscription models toward consumption-based pricing.
The mounting costs are encouraging developers, startups and even larger enterprises to experiment with lower-cost open-source models, many of which originate in China.
Rising AI bills reshape buying decisions
The cost challenge has become increasingly visible across corporate America.
According to reports, Uber exhausted its entire 2026 artificial intelligence budget in just four months after employees rapidly adopted AI coding tools, forcing management to introduce usage limits.
The experience highlights a broader shift unfolding across enterprises, where organizations are now scrutinizing AI spending much more closely and increasingly routing less demanding tasks to lower-cost models instead of relying exclusively on premium offerings from US companies.
Chinese models have emerged as some of the biggest beneficiaries of that trend.
A Citi note said leading Chinese models now charge as little as 18 cents per million tokens compared with roughly $4 per million tokens for top US frontier models while continuing to narrow the capability gap.
On OpenRouter, a platform that allows developers to access multiple AI models through a single interface, the four most widely used models are now all Chinese, with DeepSeek remaining the most popular.
DeepSeek, which emerged as the face of China’s AI ambitions last year, continues to charge only about 3% of the token price of OpenAI’s GPT 5.5, making it one of the cheapest frontier models available.
Competition among Chinese developers has also intensified, with companies including Alibaba, Moonshot AI and Zhipu AI racing to improve performance and benchmark rankings.
Chinese models rapidly gain market share
The growing appeal of Chinese models is reflected in adoption data.
The share of tokens consumed by US companies through Chinese models on OpenRouter has remained above 30% every week since Feb. 8 and has climbed as high as 46%.
That represents a dramatic increase from an average of 11% over the previous 12 months and just 4.5% during the first half of 2025.
Z.ai’s GLM 5.2 is becoming especially popular
Among the newest entrants, Z.ai, or Zhipu AI’s GLM 5.2, has emerged as one of China’s strongest challengers.
According to Artificial Analysis’ benchmark rankings, GLM 5.2 ranks fifth globally behind three Anthropic models and one OpenAI model while outperforming Google’s Gemini models.
If enterprises use the version hosted by Zhipu in China, the token cost is only about 15% of that charged by OpenAI’s comparable model.
The model has been adopted rapidly since its June release.
Harpreet Arora, head of agentic infrastructure at Vercel, told CNBC that GLM 5.2 experienced the fastest adoption of any model tracked by the company during 2026.
“In its first full week after launch, daily token volume grew about 27x, and the number of customers using it grew about 80x.”
Arora said pricing has become the decisive factor for many developers.
“Price is doing the work here,” Arora said.
“When a task doesn’t need the best model, teams are beginning to route it to the cheapest one that’s good enough, and the recent wave of models coming out of China is winning that trade.”
GLM 5.2 is regarded as particularly effective for software development and AI agents capable of interacting with other applications autonomously.
According to a New York Times report, Rehaan Ahmad, co-founder of Silicon Valley startup alphaXiv, said the performance gap has narrowed considerably.
The model also incidentally came out when executives in Silicon Valley were becoming worried that the Trump administration was leaning toward regulating the technology.
“With Fable restricted, the gap between the US and China is very slim,” Ahmad said.
The report also cited ArenaAI Chief Executive Anastasios Angelopoulos, who said Z.ai has become the world’s third most widely used AI technology platform.
Last month, Bloomberg reported that Zhipu is considering a share sale to raise several billion US dollars in Hong Kong, people familiar with the matter said, after soaring about 2,000% since its listing in January.
Zhipu is working with advisers on a potential placement that may take place as soon as next month, the people said, asking not to be identified because the information is private. A six-month lock-up from its IPO expires on July 8.
Last month, Chinese AI startup Zhipu was weighing a multi-billion dollar share sale after it skyrocketed around 2,000% since its January IPO, Bloomberg reported, underscoring the enduring optimism around the country’s AI stocks.
Big cloud providers broaden access; smaller businesses gain most
Major US cloud providers are also making Chinese models easier to access.
Microsoft, Amazon Web Services and Google Cloud already offer models from DeepSeek, Z.ai, MiniMax and other Chinese developers through their cloud platforms.
According to people familiar with the matter cited by Axios, Microsoft has also explored adding DeepSeek’s latest model to power one of its own AI products, which currently relies on OpenAI and Anthropic technologies.
For startups, the economics are becoming increasingly difficult to ignore.
Rest of World reported that San Francisco-based AI startup Lindy recently switched from Anthropic’s models to DeepSeek.
Founder Flo Crivello said the move saved the company millions of dollars.
“You don’t need God to write your email,” Crivello said on technology show MTS.
“If you can get those lower tiers of intelligence for a tenth of the price, it would be foolish not to do it.”
The publication also cited Dallas developer Ruben Garcia Jr., who spends about $500 each month on Claude and ChatGPT for sophisticated planning while paying another $200 for Chinese models including Minimax, Moonshot’s Kimi and Xiaomi MiMo, which handle roughly 90% of routine coding and voice-recognition work.
Kyle Chan, a fellow at the Brookings Institution, told Rest of World that enterprise adoption among large companies is already relatively mature.
“[AI] adoption at large companies is fairly saturated,” Chan said. “The growth market for Chinese companies would be medium-sized businesses that are starting to get into AI but are wary of the costs.”
Security concerns remain a hurdle
Despite the growing momentum, analysts say geopolitical and security concerns remain major obstacles to widespread enterprise adoption.
Many organizations remain reluctant to process sensitive data through AI systems hosted in China because of concerns over government oversight, censorship and potential export-control risks.
Z.ai was added to the US Commerce Department’s trade blacklist in 2025, while corporate filings show several shareholders are linked to a Chinese government agency overseeing the country’s defense industry.
Companies using Chinese models have also attracted political scrutiny.
Lawmakers have investigated Airbnb and AI coding startup Anysphere after disclosures that they used Chinese open-source models including Qwen and Kimi.
Airbnb Chief Executive Brian Chesky later clarified that the company was not sending customer data to model developers.
Poe Zhao, founder of the Beijing-based newsletter Hello China Tech, told Rest of World that regulated industries would remain cautious because of concerns over data security and geopolitics.
However, GLM’s open-weight architecture allows companies to deploy the model on their own servers or private cloud infrastructure without sharing data with Zhipu.
Val Bercovici, chief AI officer at WEKA, told Reuters that open-source AI models are increasingly delivering the right balance between performance and cost.
“90% as good at 10% of the price,” he said. “We don’t need to spend the premium tokens on every level of effort.”
Still, the pricing advantage may not last indefinitely.
According to The Wall Street Journal, OpenAI is considering significant price cuts as competition with Anthropic and lower-cost Chinese rivals intensifies in the enterprise AI market.
The post Cheap, capable, and controversial: why US companies cannot resist Chinese AI models appeared first on Invezz
