Chinese AI shakes up global markets

In short: Chinese-developed AI models are becoming more advanced and cheaper to deploy, which is threatening US dominance. Financial markets are rattled, adding to concerns about whether massive investment in AI will generate adequate returns. What's next? Investors will keenly watch results from AI giant Nvidia later in the month. When China's DeepSeek sent shock waves through the technology world early last year, it raised a confronting question for America's tech giants, could China build cutting-edge artificial intelligence for a fraction of the cost? More than a year on DeepSeek is no longer the only strong challenger. In recent weeks, Chinese AI developers Z.ai (formerly known as Zhipu AI) and Moonshot AI have released new models that have made a splash on the global tech sector. Moonshot AI's release of Kimi K3 in particular, shook Wall Street, triggered accusations from Washington that the startup was using proprietary US technology to develop its models and even won praise from competitor OpenAI's president. Markets are rattled Despite a strong finish to the month, key AI bellwether indexes have been volatile. The Nasdaq 100 briefly entered correction territory earlier in the week, while this year's world beating South Korean Kospi fell as much as 38 per cent in July before recovering some of its losses. Many analysts say the recent sell-off in tech has been largely fuelled by concerns about whether the hundreds of billions of dollars of spending on AI by the technology giants will generate adequate returns. One key reason analysts fear those eye-watering investments may not be recouped is growing competition from cheaper Chinese owned rivals. "The broad thinking was the US was about six to nine months ahead of Chinese competitors but after developments recently, we now know that Chinese AI, Chinese infrastructure, are as good as some of the US frontier models, and that is providing another shake-up for Wall Street and those markets which are associated with Wall Street, such as in South Korea and Japan," said IG Market analyst Tony Sycamore. "The pullback is being most acutely felt in South Korea, but if we were to see the US tech market continue to accelerate lower, that would certainly flow through to Australians via the returns which their super funds are earning,"he said. Australian businesses react to the shift For some Australian businesses, particularly startups and companies building AI into their products and supply chains, the fierce international competition is changing the economics of AI. Jacky Koh, co-founder and co-CEO of Sydney-based startup Relevance AI, has had a front-row seat to this shift. His company helps businesses (like Canva, KPMG and Autodesk) build AI agents, systems that can essentially carry out tasks on their own, rather than simply respond like a chatbot. The models powering those agents broadly fall into two camps. Open, or, more specifically, open-weight models are free to download and most of the leading open models now come from Chinese developers. Self-hosting allows businesses to avoid paying a developer each time the model is used, although they must still cover the computing power required to run it. Leading proprietary models, mostly developed by American companies, remain controlled by their developers and are far more expensive to use. Until recently, Mr Koh said those agents were overwhelmingly powered by closed source models from US companies such as OpenAI, Anthropic and Google. But his firm is now increasingly directing more work towards models it can download and operate itself, including Chinese-developed 'open' models GLM and Deepseek. "Before the start of this year, I would say honestly, about five, if not 7.5 per cent of traffic were mainly in open weight models. "As of right now, we're closer to 20, if not 25 per cent." The economics are hard to ignore A major driver is cost. As businesses move beyond asking chatbot questions and deploying AI agents that carry out multiple steps autonomously, the number of tokens they can consume can rise dramatically. Tokens are the currency of the AI economy — every question AI reads and answers uses them — and businesses pay for that usage. Deepak John Joseph, who runs a product engineering business, Pupa Clic Technologies, says some businesses are coming to him because those bills have ballooned. "We have customers explicitly asking, for example, we currently consume $20,000 worth of tokens per day, and how do we drastically reduce this as our volumes are hitting through the roof and we need to really bring this down." he said. "We are looking at almost 80 per cent of savings in using hosted models in compared to the cloud hosted variants." For Australian businesses, the emerging picture appears to be firms increasingly mixing models, reserving the most powerful, expensive systems for difficult jobs and routing simpler work to cheaper alternatives. Will Liang, from Sydney-based AI advisory firm Amplify AI, calls the approach "model routing". "You use the top tier large language model to figure out what are the list of tasks you need to do to complete this objective, and then route different tasks to different models. "Some of the simpler tasks to more a cheaper model, as example." Concerns about Chinese AI There have been concerns about bias in Chinese-developed models, particularly around politically sensitive subjects. "It won't, for example, tell you very much about Tiananmen Square or tell you things maybe that are unfavourable to the Chinese regime," said UNSW AI Institute chief scientist Toby Walsh. However, Mr Liang says for most Australian businesses, a bigger barrier to adopting Chinese AI is the lack of trust and uncertainty around regulation. "We do get a lot of pushback (to using Chinese-developed AI), especially in healthcare as well as on the finance-based domains," confirmed Mr Joseph. To increase security, Trivium China partner in technology, Kendra Schaefer says firms should ensure that Chinese developed models are accessed through "trusted domestic cloud providers, or installed on premises under their own control". While open weight models mean they can be downloaded locally, she also cautioned it does not equate to freedom from security risk. "Like any locally installed software, a locally installed model is a vector for compromise. "Every time you download software onto your computer you accept the risk that it contains a malicious backdoor," she warned. She says it is not a China specific issue but rather an issue around cybersecurity risk. "It calls for the same response: investment in safety protocols and protective mechanisms purpose-built for locally hosted AI models."