Stay informed with free updates
Just sign up for artificial intelligence myFT Digest — delivered straight to your inbox.
China’s artificial intelligence groups have posed “AI-in-a-box” products that companies can run on their own premises as a threat to AI cloud computing services offered by the country’s major technology groups such as Alibaba, Baidu and Tencent. are on sale.
Huawei has signed deals with more than a dozen AI startups to bundle and sell its large-scale language models with its AI processors and other hardware. Its partners include groups such as his Beijing-based Zhipu AI and language specialist company iFlytek.
The Chinese group is introducing the box to bring advances in generative AI to on-premises or private cloud setups, which account for about half of the country’s cloud market.
Huawei estimates that the market for so-called “all-in-one machines” in China will reach RMB 16.8 billion ($2.3 billion) this year. Minsheng Securities analysts predict that the government AI box market could reach RMB 450 billion by 2027.
iFlytek founder Liu Qingfeng said at last year’s AI Box product launch event that the company’s “all-in-one machine has the highest performance, is safe, controllable, and ready to use out of the box.” Ta.
The move capitalizes on Chinese companies’ concerns about data protection, unlike the way AI is being commercialized in Western countries.
This trend could derail the ambitions of tech giants, which have invested heavily in building AI infrastructure and large-scale language models that can be sold as services to customers via the cloud. Baidu’s Robin Li has laid out a vision to run hundreds of AI apps on the company’s basic model.
The proliferation of AI-in-a-box may also increase the fragmentation of China’s cloud market.
When China first rolled out cloud computing, Huawei established itself as a leading provider of private clouds, especially for wealthy governments and state-owned groups. Alibaba and Tencent ended up competing primarily to provide computing power to other Internet groups.
This became an issue in 2021 when the Chinese government cracked down on its technology customer base, upending the business models of major customers such as online education companies. Since then, Alibaba’s cloud division has averaged single-digit quarterly growth.
“Baidu and Alibaba are focusing on public cloud, but the ecosystems in China and the US are very different,” said a local investor. “They will have to adapt to survive.”
Baidu announced last week that generative AI and underlying models contributed RMB 324 million in revenue in the first quarter. The company has also started selling all-in-one machines, but analysts say the tech giant doesn’t have much of an advantage in a market where it’s more difficult to leverage its own data centers with massive processing power. It has said.
Dylan Patel, principal analyst at research group Semi-Analysis, said that while some Chinese companies may want on-premises AI, there is little value compared to using APIs to connect to public clouds or large-scale language models. He said that it would be inefficient.
“Usage will be very sporadic, which means all of the very expensive AI hardware will not be properly utilized,” he said.
A research note from a Chinese investment bank cited security concerns among Western AI groups, including a case in which OpenAI’s ChatGPT accidentally shared users’ search history and a Samsung employee allegedly leaking trade secrets to a chatbot. flaws are highlighted.
“Organizations need to be able to protect their data, and building a private cloud is the way to prevent valuable data from being leaked,” iFlytek’s Liu told potential customers.
Kent Huang, a banker at technology advisory firm China Renaissance, said the all-in-one box helped circumvent a shortage of computing power in China caused by the U.S. government’s export restrictions on advanced chips.
Procurement records show large Chinese state-owned groups favor private cloud AI.
A smart city service provider in Chengdu recently made a RMB 2 million bid to run China Unicom’s large-scale language model on Huawei’s all-in-one machine hardware.
China National Nuclear is also considering building an on-premises system that bundles models with hardware with more than 10 billion parameters, or variables used to train the system and ultimately shape its output. Startup Zhipu AI is interested in the project, according to people familiar with the matter.
According to Minsheng Securities, Zhipu charges RMB 1.8 million annually for a 12 billion parameter model equipped with eight Huawei Ascend 910 chips. The Ascend 910 is Huawei’s most advanced AI chip with the processing power to train large models and process “inference”, or respond to user queries.
Huawei, Baidu, Alibaba and Tencent declined to comment.
