Former Google software engineer Linwei Ding was found guilty by a US court on Tuesday of stealing confidential information tied to Google’s AI infrastructure. Ding is said to have stolen AI secrets from Google for Chinese companies. He has been sentenced to 12 months minus one day in prison.
Documents transferred by Linwei Ding to Chinese companies included details about Google’s hardware infrastructure and software platform for training large AI models, including Tensor Processing Unit (TPU) and Graphics Processing Unit (GPU) systems. The custom TPUs serve as flagship AI technology for Google, allowing the company to train AI models without needing to rely on external companies for compute.
Experts believe that these documents could allow a recipient to “skip huge parts of the design process” for an AI supercomputer that Google may have spent millions of dollars to develop over a decade.
US District Judge Vince Chhabria said the punishment reflected the “seriousness” of the offence and described the conduct as a “systematic, brazen effort to steal Google’s property.” Ding will also serve two years of supervised release and pay more than $189,000 (roughly Rs 1.79 crore) in restitution along with a $25,000 (roughly Rs 23.75 lakh) fine.
Who is Linwei Ding?
Linwei Ding, also known as Leon Ding, is a Chinese national who came to the United States in 2010 to pursue a master’s degree. He later worked at several Silicon Valley technology companies before Google hired him as a software engineer in 2019. He became a lawful permanent resident in 2021.
Prosecutors claimed that his role gave him access to confidential information about the hardware and software infrastructure used in Google’s supercomputing data centres to train large AI models.
A jury in San Francisco found Ding guilty in January after an 11-day trial on seven counts of theft of trade secrets and seven counts of economic espionage. The espionage counts were later set aside by the judge.
How did Linwei Ding steal Google’s AI secrets?
According to prosecutors, Linwei Ding began transferring confidential material from Google’s network to a personal Google Cloud account on May 21, 2022, and continued until May 2, 2023. At trial, prosecutors claimed that Ding uploaded more than 500 confidential files over about a year.
A superseding indictment later alleged that this included more than 1,000 unique files, and prosecutors presented evidence concerning more than 2,000 pages of confidential information. Court papers also said that in December 2023, less than two weeks before leaving Google, Ding downloaded the material to his personal computer.
As per court documents, these files covered Google’s infrastructure and software for training large AI models, including details about TPUs and GPUs. Other information included software that allows chips to communicate and carry out tasks, and software that coordinates thousands of chips into a supercomputer. It also included details of Google’s custom-designed SmartNIC, a networking component used in its AI supercomputers and cloud networking products. Prosecutors said some of the information was meant to give Google an edge over cloud rivals and reduce its reliance on Nvidia chips.
The case also focused on Ding’s links to China while he was still at Google. As per prosecutors, months after the uploads began, an early-stage technology company in China offered him a chief technology officer role. He visited the company, Rongshu, in late 2022 and took part in investor meetings from around December.
In 2023, he founded Zhisuan, a China-based technology company that proposed to train large AI models. The indictment quoted a Zhisuan document circulated by Ding in November 2023 as saying, “We have experience with Google’s ten-thousand-card computational power platform; we just need to replicate and upgrade it and then further develop a computational power platform suited to China’s national conditions.”
Ding is said to have applied in late 2023 to a Chinese government programme aimed at attracting technological talent to China. Prosecutors claim he never told Google about his links to either company and, while in China pitching Zhisuan, arranged for an intern to swipe his Google identification card several times so it would appear he was working from his office. The judge later wrote that the evidence was not enough to prove that, during the May 2022 to April 2023 period, Ding knew or intended that his conduct would benefit the Chinese government.
This case comes at a time when the US and China are in an arms race when it comes to developing AI models. While the US has blocked Beijing from using the most advanced GPUs to train AI models, Chinese companies are increasingly looking at local ways to develop their own chips. Ox Alpha, a new AI model that went viral last week, is said to have been trained entirely on domestic-made chips. Ox Alpha is part of Z.ai’s GLM-5 family of models.
A social media account affiliated with Chinese state-run media has also warned the US ahead of joint talks on AI, stating that American companies cannot dictate AI rules for the world.
– Ends
