China’s internet giants order $5bn of Nvidia chips to power AI ambitions

China’s internet giants are rushing to acquire high-performance Nvidia chips vital for building generative artificial intelligence systems, making orders worth $5bn in a buying frenzy fuelled by fears the US will impose new export controls.

Baidu, ByteDance, Tencent and Alibaba have made orders worth $1bn to acquire about 100,000 A800 processors from the US chipmaker to be delivered this year, according to multiple people familiar with the matter. The Chinese groups had also purchased a further $4bn worth of the graphics processing units to be delivered in 2024, two people close to Nvidia said. 

The A800 is a weakened version of Nvidia’s cutting-edge A100 GPU for data centres. Due to export restrictions imposed by Washington last year in a bid to choke Beijing’s technological ambitions, Chinese tech companies are only able to buy A800s, which have slower data transfer rates than A100s.

As hype has intensified around AI over the past year, Nvidia’s GPUs have become the hottest commodity among the world’s biggest tech companies to provide computing power for the development of large language AI models.

Chinese internet groups are racing to stockpile the A800 chips over fears that the Biden administration is contemplating new export restrictions that would capture even Nvidia’s weakened chips, as well as a wider GPU shortage driven by overwhelming demand.

Washington on Wednesday announced a ban from next year on some US investment into China’s quantum computing, advanced chips and artificial intelligence sectors.

“Without these Nvidia chips, we can’t pursue the training for any large language model,” said one Baidu employee who declined to be named. 

The companies are developing their own homegrown large language models after the success of ChatGPT, the breakout chatbot launched by Microsoft-backed OpenAI eight months ago. 

ByteDance has numerous small teams working on various generative AI products, including an AI chatbot codenamed Grace, which is currently undergoing internal testing, according to two people close to the company.

ByteDance tested a generative AI feature for its social media app TikTok earlier this year, named TikTok Tako, which licenses OpenAI’s ChatGPT.

Two employees with direct knowledge of the matter said ByteDance had already stockpiled at least 10,000 Nvidia GPUs to support its ambitions. They added that the company had also ordered almost 70,000 A800 chips to be delivered next year, worth about $700 million.

Alibaba plans to plug all its products into its large language model, including its online shopping platform Taobao and its mapping tool Gaode Map. Meanwhile, Baidu is making its own ChatGPT-like project, a generative AI chatbot called Ernie Bot.

Baidu, ByteDance, Tencent and Alibaba declined to comment. Nvidia said in a statement: “Consumer internet companies and cloud providers invest billions of dollars on data centre components every year, often placing orders many months in advance.”

Earlier this year, as excitement about AI gathered momentum, most Chinese internet giants had less than a few thousand chips in stock that could be used to train large language models, according to multiple employees at Chinese tech companies. As demand has grown since then, so too has the cost of those chips. 

“The price of the A800 in the hands of distributors has risen by more than 50 per cent,” said one Nvidia distributor.

Tencent Cloud in April released a new server cluster — computing power for others to rent — wielding the Nvidia H800 GPU, a version of its latest H100 model adapted for China, which can power large language model training, autonomous driving and scientific computing.

Alibaba Cloud has also received thousands of H800 chips from Nvidia, and many customers have already contacted the company seeking cloud services powered by these chips to help them replicate ChatGPT in China, according to two people close to Alibaba. 

ByteDance provides cloud computing devices with stockpiled Nvidia chips A800 and A100. It also launched a platform in June for companies to trial different large language model services.