Two days before delivering the keynote at Google I/O, the company's annual State of the Union address, Sundar Pichai is worried about losing his voice.
Sitting at the coffee table inside his remarkably spartan office at company headquarters, the Google CEO speaks softly, even by his standards. Step by careful step, he explains the major themes that will run through his keynote speech, and at first, they seem less weighty than they should, considering they represent the near-future of the world's largest internet business. That's partly because Pichai isn't feeling too well and partly because those themes aren't all that different from the themes that drove last year's speech. Google, he keeps saying, is now an "AI-first company." But about halfway through this chat, he reveals that the company will unveil a new chip specifically designed to power the new wave of artificial intelligence. However softly he speaks, this news will echo loudly across the tech industry.
Yes, Google unveiled an AI chip at last year's conference. But this is something different. According to Pichai, Google's new chip is far more powerful and far more versatile than the one that came before, designed not only to run the latest machine learning services but to train them as well. No one else has really built a chip like that, though others, including Intel, are scrambling to build such tech. And this time, Pichai says, Google will share its AI chip with the whole world. As he sees it, this is part of the company's wider effort to democratize artificial intelligence, to put it in the hands of everyone.
"One of the most exciting things we all can do is demystify machine learning and AI. It's important for this to be accessible by all people," he tells me in the I/O greenroom just after his Wednesday morning keynote, stage music pumping on the other side of the wall, the excitement and verve back in his voice. "That journey matters."
Still, when I ask if Google will sell this chip on the open market, he says no. That would democratize things a little too much. Rather than sell the chip itself, Google will offer its new silicon to the world via a cloud computing service. Anyone will be free to use it---but only if they use the Google Cloud. "We view our state-of-the-art machine learning and AI technologies as a differentiator for the cloud platform," he says. "I think it will help propel the platform forward." In other words, he hopes the new chip and the new service will set Google's cloud business apart from services offered by its main rivals, including Amazon and Microsoft, the unnamed competitive threat underlying his I/O keynote.
Between its two AI labs---Google Brain, based at company headquarters in Silicon Valley, and DeepMind, a London AI startup Google purchased a little more than three years ago---Google is leading the new wave of artificial intelligence research and development so rapidly changing entire industries and economies. And it's leading an equally enormous shift in the way the world builds and operates data center hardware. But all that is old news. In this year's State of the Union, what stands out is how forcefully Google is working to share its technology with others and, in no small way, leverage all of these strengths in the hope of reaching the top of the cloud computing market, which could ultimately be the most lucrative tech market of them all.
"I have always felt that you build a product or a platform---particularly a platform---and you do it well, there will be a lot of revenue to go with it," Pichai says. "If you are running cloud at scale, that means you are powering the data platforms that will transform many industries. Will there be an economic opportunity there? I absolutely think it will be big."
The overwhelming share of Google's revenue today comes from online advertising. But the company believes cloud computing---where computing power is rented over the internet to businesses and software developers---could one day bring in far more. Indeed, research firm Forrester says the market will reach $191 billion by 2020. The trouble is that after years of dithering, Google's cloud business trails Amazon and Microsoft in this rapidly growing market. Pichai hopes the company's lead in AI can propel its fortunes forward.
Google built its new chip as a better way of serving its own AI services, most notably Google Translate, says Jeff Dean, the uber-engineer who oversees Google Brain, the company's main AI lab. Even so, Pichai says that the company has only started using the chips internally, even as the company is preparing to share access to them via the Google Cloud. That shows how much importance the company places on the cloud market: Pichai and company are not waiting around.
Of course, this new chip is no guarantee of success in the cloud. Most companies train their AI models today using nVidia GPUs, graphics processing chips that also happen to work really well for machine learning. Choosing to use Google's new chips via the cloud could require is real change in the way businesses and developers do things. "The question is: how easy is it?" says Chris Nicholson, the founder and CEO of a machine learning startup Skymind. "Do you have to be a Google engineer to understand the stack?"
Pichai acknowledges that there may be learning curve. Buy he also believes that developers are willing to make the changes that can significantly push technology forward. "The pace with which these things shift can be very fast," he says. "When the web came, Javascript development didn't exist. The whole development chain didn't exist. But it happened."
What's more, Pichai knows that GPUs will continue to play a big role in AI research and operations. Google will offer its new chip via the cloud, but it will offer access to GPUs and possibly other chips for AI, as well. "This is not either/or," he says. "It's about advancing the overall optionality." That's Silicon-Valley-speak for "choice." And it makes good sense. Pichai wants to give businesses whatever they need, even as he gives them something no one else can.
Ultimately, Pichai says Google is doing what it is best at: building platforms that others can use to build whatever they like. "We have always cared about computing platforms. I have spent most of my time here working on them," he says. "It is the most leveraged way to drive innovation." The TPU service is yet one more platform that will likely feed so many others in the months and years to come. After all, it's a way of running any machine learning service—the core technology a host of other industries will use to transform how their own businesses work.
Pichai is particularly interested in Google's AutoML research project, a way of building artificial intelligence using artificial intelligence. He says it reminds him of Inception. "I tell my engineers: 'We must go deeper,'" he quipped during his keynote. His larger point is that automation will soon be an enormously important way of creating AI---and he envisions Google offering this option to the rest of world as its own online service, driven by the AI chips the company just unveiled.
Meanwhile, as Google also revealed this morning's keynote, its engineers are building new software---TensorFlow Lite---that will run machine learning models on individual smartphoness, eliminating the need to connect to the internet to perform every AI-powered task. Google intends to open source this technology, too, turning it into yet another platform for the masses. These localized versions of AI will still feed off the new TPU service---and feed into it too. Before running AI models on the phones, developers must train them in the data center.
Still, amid all the talk of democratizing AI and creating platforms for everyone, let's not forget that cloud computing is also a rather direct way of making money, a resource that also happens to be enormously important to the company's future. Much like last year, Pichai is intent on painting Google as an AI-first company. And there's a lot of truth to that spin. But he also knows that AI is a path to another breakthrough: a new way of funding everything his company does.
