More than seven years after exiting China, Google is taking the boldest steps yet to come back. And it’s not with a search engine.
Instead, Google’s ingress is centered around artificial intelligence. The internet giant is actively promoting TensorFlow, software that makes it easier to build AI systems, as a way to forge business ties in the world’s largest online market, according to people familiar with the company’s plans. It’s a wide pitch targeting China’s academics and tech titans. At the same time, Google parent Alphabet Inc. is adding more personnel to scour Chinese companies for potential AI investments, these people said.
“China is a tremendous opportunity for any company because it is by far the single largest homogeneous market,” said Kai Fu Lee, who headed Google’s China operations before the company left in 2010. The market dwarfs any other, given how many Chinese people are online, and data from that “can be used to advance products, especially those relating to artificial intelligence,” he added.
A more active Google in China does not guarantee a profitable Google in China. The company’s primary mechanism for cashing in on its AI tools, its cloud-computing business, can’t be accessed by developers in China without overseas servers or technical tricks to work around the country’s Great Firewall, laws and technology that control the domestic internet and block some foreign websites. Google also faces stiff homegrown competition, mainly from search nemesis Baidu Inc., in the race to create the most popular foundational tools for inventions like voice-controlled speakers and self-driving cars.
Still, Google is clearly interested in re-igniting its business in China. It pulled its search engine and many other services from the mainland in 2010 over government censorship. In the years since, Google has explored several paths for re-entry, including activating its mobile app store there, with little success. China has become the biggest market for smartphones running Google’s Android software, but without Google services.
“I’m committed to engaging more in China,” Sundar Pichai, Google’s chief executive officer, said in a recent interview. “We’ll thoughtfully figure out how to engage deeper, and I don’t know what the answers are.” A Google spokesman declined to comment.
Rather than another splashy product launch, Google’s latest China strategy is a grassroots effort focused on getting developers in the country trained and hooked on its AI building blocks. It’s similar to the way business software startups get employees using their services before corporate IT departments notice. Once the tools become popular, companies often accept the technology and sign up for full service.
In the past month, several of Google’s U.S.-based engineers have given at least three detailed briefings at TensorFlow developer events in Beijing and Shanghai. Two of those were invite-only, with attendees asked not to record, photograph or even blog about the sessions, according to people familiar with the gatherings. Google said it supports developers using TensorFlow anywhere in the world, and isn’t focusing on China specifically.
The latest tactic fits with Pichai’s mantra that Google be “AI-first” — an effort to re-orient its web services from a world where people type on screens to one where they talk to an array of devices. TensorFlow, which Google began offering free in 2015, is a cornerstone. The tools have become wildly popular with developers and inspired imitators. This year, Google’s cloud service began renting access to a new chip optimized for TensorFlow.
It’s hard to find a place as fertile for AI as China. The country has one of the fastest growing TensorFlow developer communities in Asia, despite the fact that Google’s cloud services are unavailable there. The Chinese government has made AI a national priority. Scores of Chinese companies are deploying machine-learning systems — AI software that automatically adjusts to data — to update banking services, identify faces in crowds and control drones.
Matroid Inc., a machine-learning startup in Palo Alto, California, hosted a conference on TensorFlow in March. Jeff Dean, a revered Google engineer, spoke and posted his presentation slides online. Within an hour, the slides were translated into Mandarin and went viral on the Chinese social network WeChat, said Matroid’s CEO Reza Zadeh.
Chinese developer Jiang Jun attended a TensorFlow event on Oct. 24 in Shanghai featuring Google employees. He’s a senior engineer at Ele.me Inc., a food-delivery app valued at $6 billion and backed by Alibaba Group Holding Ltd. Much of Ele.me’s systems are built on TensorFlow. Outside China, a startup like this might use servers from Google’s cloud business to run its app. But China’s Great Firewall blocks Google servers, so Jiang’s team has modified the code in some of TensorFlow’s tools so the software no longer tries to access files from Google’s servers and instead runs on Ele.me’s domestic servers.
“Before this year, Google didn’t pay too much attention to doing activities in China because, although they know China has a large market size, they know they can’t do a lot of things because of the firewall,” Jiang said after the TensorFlow event. “So when Google comes to China to introduce TensorFlow it is, in my opinion, more pure because it cannot get that much money.”
Even if Google’s cloud business was allowed to put servers in mainland China, local rivals like Alibaba now sell cheap cloud-computing products that would make it hard for the U.S. company to turn a profit, he added.
“But all of us developers are always waiting for Google to come to China to introduce more TensorFlow technologies and products,” he said. “Google’s cloud solutions are so cool and its tools are so convenient.”
Beijing-based Wang Xiaoyu said TensorFlow was a vital tool for her podcast startup CastBox.FM. Developing her own tools would’ve required a team of 20 expensive machine-learning specialists. Instead, she turned to TensorFlow and hired a single Chinese PhD graduate with TensorFlow experience capable of producing the same results. Her company is now worth about $60 million with more than 8 million users downloading her app.
To use TensorFlow, CastBox.FM has U.S. servers that store data on its non-Chinese users. Wang and her colleagues have this information piped into China for analysis. This addresses new government rules that require domestic user data to be stored in the country. It’s trickier for other startups with more Chinese customers to use TensorFlow, Wang said.
“Users in China simply use the best and the one with the most support,” Qiang Yang, a computer science professor at the Hong Kong University of Science and Technology, said, referring to the popularity of TensorFlow.
As interest has grown, Google is going on a hiring spree, recently posting AI job openings in several Chinese mega-cities. The company has started sending more staff there, too. In recent months, members of Google’s corporate-development team and Alphabet’s private equity arm, CapitalG, have met with AI companies based in China, according to two people familiar with the situation. One person described these efforts as “information-gathering” and said neither investment arm has decided to enter financing rounds recently. These people asked to remain anonymous talking about private company matters. In 2015, Google invested in Mobvoi, a Chinese startup run by former Google engineers that places AI-powered chatbots inside smartphones, cars and other devices.
Google’s latest China effort may fail like its previous attempts — and the country’s response to the company’s most-public event there earlier this year isn’t encouraging. An AI system developed by DeepMind, Alphabet’s AI research lab, trounced China’s human champion in a game of Go, offending senior officials and helping spark a government-funded push to dominate the technology.
That hasn’t deterred Google. The company has pitched Alibaba and Tencent Holdings Ltd. on using TensorFlow, an attempt to spread the software through the rank’s of China’s largest tech firms, people familiar with the effort said.
TensorFlow has been downloaded more than 7.9 million times so far, from both its own servers and outside sources, and the company is pleasantly surprised by the early uptake. Ricky Wong, an investor who often works in China, analyzed the location of the first 5,000 developers to access the tools and found more came from Beijing than all of Silicon Valley.
Early interest helps Google in China, but Baidu introduced its own AI toolkit, called PaddlePaddle, last year. The spread of Baidu’s tools among developers has outpaced Google’s this year, according to one person familiar with Google’s internal figures. Baidu declined to comment.
For some AI researchers in China, Baidu’s success reflects a loyalty to local offerings and caution with reliance on foreign tools, said Jiebo Liu, an AI expert who studies China at the University of Rochester. “They might use TensorFlow for prototyping,” he said. “But if they want to put something in product, they use their own.”