{"id":19882,"date":"2026-07-08T17:07:21","date_gmt":"2026-07-08T11:22:21","guid":{"rendered":"https:\/\/eng.dragonmedia.com.np\/?p=19882"},"modified":"2026-07-08T17:07:21","modified_gmt":"2026-07-08T11:22:21","slug":"cost-and-efficiency-drive-rising-use-of-chinese-ai-models-by-us-companies","status":"publish","type":"post","link":"https:\/\/eng.dragonmedia.com.np\/?p=19882","title":{"rendered":"Cost and Efficiency Drive Rising Use of Chinese AI Models by US Companies<\/strong>"},"content":{"rendered":"

Dragon Media News Desk<\/strong><\/p>\n

US companies are increasingly turning to artificial intelligence models developed by Chinese firms as they expand AI-powered products and services, with industry analysts identifying lower costs and competitive performance as the main factors behind the shift.<\/p>\n

The trend does not necessarily indicate a broad preference for Chinese technology over American systems. Instead, it reflects growing pressure on businesses to control the rapidly rising cost of running AI applications at scale.<\/p>\n

Recent models developed by Chinese companies, including DeepSeek, Zhipu AI and Alibaba\u2019s Qwen, have narrowed the performance gap with leading systems from US firms such as OpenAI and Anthropic in areas including coding, reasoning, tool use and autonomous agent tasks.<\/p>\n

At the same time, the cost of using some advanced American models has become a significant concern for companies processing large volumes of data and operating AI agents that may complete hundreds or thousands of steps for a single task.<\/p>\n

In AI systems, text, code and other forms of information are divided into basic processing units known as tokens. Companies are generally charged for both the tokens sent to a model and those generated in response. As AI agents perform increasingly complex and repetitive tasks, token consumption can rise quickly, making small differences in price commercially significant.<\/p>\n

Data cited by US media outlet CNBC showed that the share of tokens used by American companies on Chinese AI models through OpenRouter remained above 30 percent every week after February 8, reaching as high as 46 percent.<\/p>\n

That compared with an average of about 11 percent over the previous 12 months and only 4.5 percent during the first half of 2025.<\/p>\n

OpenRouter is a platform that allows developers to access models from multiple providers through a common interface and select systems based on cost, performance and specific operational requirements.<\/p>\n

Its data indicate growing use of Chinese models from DeepSeek, Zhipu AI, MiniMax, Xiaomi and Alibaba\u2019s Qwen family, particularly for coding, tool use and autonomous AI-agent applications.<\/p>\n

However, OpenRouter data do not represent the entire US artificial intelligence market. Major companies may access models through direct commercial agreements, their own data centres or cloud platforms operated by Amazon, Microsoft and Google.<\/p>\n

The figures should therefore be viewed as an important indicator of changing behaviour among developers and startups, rather than a complete measure of corporate AI use across the United States.<\/p>\n

One prominent example is Lindy, a US-based AI automation startup that reportedly shifted all of its traffic from Anthropic\u2019s Claude to DeepSeek in June.<\/p>\n

The company\u2019s chief executive said the move was expected to save millions of dollars over the following months. Lindy reportedly used a version of the Chinese-developed model hosted by an American service provider on infrastructure located in the United States.<\/p>\n

The case illustrates how companies can use Chinese open models without necessarily sending their data to servers operated in China.<\/p>\n

Chinese models have gained a strong pricing advantage through a combination of technical innovation, open-source development and aggressive commercial strategies.<\/p>\n

DeepSeek\u2019s models have attracted attention for offering comparatively low application programming interface prices while maintaining competitive capabilities across many enterprise tasks.<\/p>\n

Zhipu AI has similarly promoted its GLM model family as a lower-cost alternative to leading US systems. The company has claimed that some of its models perform close to advanced Anthropic systems on agent-related benchmarks while costing substantially less.<\/p>\n

Benchmark results published by model developers should nevertheless be interpreted cautiously, as performance can vary across testing environments and real-world business applications.<\/p>\n

A major factor behind the lower operating cost of Chinese models is the growing use of mixture-of-experts architecture.<\/p>\n

Instead of activating every part of a large model for each request, mixture-of-experts systems activate only the components most relevant to a particular task. This can reduce the amount of computing power needed for inference while preserving much of the model\u2019s overall capability.<\/p>\n

Chinese developers have also focused on methods such as multi-head latent attention, sparse attention, mixed-precision training and cache-based processing to improve efficiency.<\/p>\n

US restrictions limiting China\u2019s access to the most advanced graphics processing units have also pushed Chinese AI laboratories to place greater emphasis on algorithmic optimisation and efficient use of available hardware.<\/p>\n

Industry experts say these constraints have encouraged Chinese developers to design systems that can achieve strong results with fewer computing resources.<\/p>\n

Companies including DeepSeek, Zhipu AI and Alibaba have consequently focused on reducing the computing power and number of tokens required to complete practical tasks.<\/p>\n

For enterprise users, the result is often lower operational cost without a major decline in performance for routine applications.<\/p>\n

Many business tasks do not require the largest or most expensive available model. Customer service, document analysis, content classification, standard coding support and administrative automation can frequently be handled by smaller and more efficient systems.<\/p>\n

The rapid growth of AI agents has made cost efficiency even more important. An agent may plan a task, consult external tools, review results and correct errors through numerous stages before completing an assignment.<\/p>\n

In such cases, the difference between two models\u2019 token prices can have a substantial effect on monthly expenditure.<\/p>\n

As a result, many companies are adopting mixed-model strategies. Less expensive systems are used for routine work, while more powerful and costly models are reserved for complex reasoning or specialised tasks.<\/p>\n

Open-source and open-weight Chinese models also offer companies greater control over deployment.<\/p>\n

Businesses can install some models on their own servers, adapt them for specific requirements and decide where their data are processed. Compatibility with software interfaces used by OpenAI and Anthropic can also make migration from American systems relatively straightforward.<\/p>\n

Analysts say the global AI market is therefore moving beyond a competition based solely on which company produces the most powerful model.<\/p>\n

The more important commercial question is increasingly how much computing capacity, time and money are required to complete a task reliably.<\/p>\n

Chinese models are becoming more competitive not only because of their technical capabilities, but also because of lower token prices, efficient architecture, flexible deployment and broader availability.<\/p>\n

Their expansion nevertheless raises continuing concerns over data security, privacy, legal jurisdiction, service reliability and the handling of sensitive information.<\/p>\n

These issues are particularly important in defence, government, finance, healthcare and cybersecurity, where organisations must know where a model is hosted, where data are stored and which country\u2019s laws apply.<\/p>\n

For this reason, some US companies are choosing Chinese-developed open models hosted in American data centres rather than using services directly operated from China.<\/p>\n

The latest trend suggests that the global artificial intelligence market is evolving toward a multi-model and multi-provider system rather than remaining dependent on a small number of closed platforms.<\/p>\n

Growing use of Chinese models by US companies is also likely to place pressure on OpenAI, Anthropic and other American developers to lower prices, improve efficiency and offer customers a wider range of models for different cost and performance requirements.<\/p>\n

For Chinese AI companies, however, maintaining long-term international growth will require more than low prices. They will also need to demonstrate service reliability, data protection, transparency and compliance with international regulatory expectations.<\/p>\n

The increasing presence of DeepSeek, Zhipu AI and Qwen in the US market shows that cost-effective inference and practical deployment have become central to the next stage of global AI competition.<\/p>\n

\n\t\t\t
\n\t\t\t <\/div><\/div>\n\t\t