Tech Titans in the Driving Seat
Article By : Dylan McGrath
Locked in an AI arms race and sitting on piles of cash, hyperscalers are beginning to set the agenda in semiconductor process technology.
Representatives from internet and cloud-computing giants Facebook, Amazon, Microsoft, and Google were at Semicon West earlier this month, roaming the show floor, sitting in on presentations, and kicking the tires on the newest fab tools and semiconductor processing techniques. It's not completely unprecedented that a few people from one of the big hyperscalers or the other attend a semiconductor show, but it's unusual when they all show up en masse like that. Something's up.
“Hardware is sexy again,” said G. Dan Hutcheson, the veteran semiconductor capital equipment analyst who is chairman and CEO of market watcher VLSI Research. “People tend to assume that it’s all software now. But, you know, software runs best on hardware.”
The hyperscalers' presence — and their deep pockets — was like a ray of sunshine for what might otherwise have been a dreary event for an industry in the midst of a significant downturn. The appearance of engineers from these firms in such large numbers was proof-positive that, despite the semiconductor industry’s unenviable current position on the supply-demand curve, there is a very bright light at the end of the tunnel: Many, many more chips will be made.
Glamour companies the likes of Google and Facebook generally don’t like to get their hands dirty. They mostly make their hay through sophisticated software algorithms developed by their teams of programmers. They certainly don’t build semiconductors, and they have no desire to start doing so anytime soon. Why, then, would these massive companies with billions of users take such keen interest in things as esoteric as the edge deposition layer thickness specification on a batch sputtering tool?
In recent years, the so-called hyperscalers have been sending more people to Semicon West and shows like it, demonstrating an eagerness to engage with the semiconductor manufacturing supply chain in ways that most system OEMs never have. Not only are they designing their own chips, but — failing to convince foundries to commit to longer-term roadmaps — they are rolling up their sleeves, rubbing elbows, and sharing ideas with equipment, materials, and design software vendors to help drive the direction of semiconductor process technology.
About three years ago, Hutcheson and his colleagues at VLSI Research began to take note of the growing presence of big tech giants like Google at Semicon West and other gatherings of the semiconductor manufacturing community. Dubbing them “hyper-verticals,” Hutcheson concluded that these firms were picking up the mantle from vertically integrated companies of the 1970s and 1980s like AT&T, General Electric, and General Motors, which maintained their own fabs and built their own chips. This breed of company all but died in the 1990s with disaggregation and the advent of the foundry-fabless model.
Hutcheson began asking around. He learned from contacts at places like Google and IBM that these companies could no longer get the advances that they needed just from software. In order to optimize their businesses, they needed to see all the way down into the devices and see how the chips were evolving.
Off-the-shelf devices, even high-end microprocessors from the likes of Intel, weren’t enough for these tech giants, Hutcheson said. “All the conventional chips that are sold as building blocks by chipmakers, it only gets them so far,” he said. “This is what I was seeing three years ago, that these guys were starting to say that hardware is the future. And when you hear a software guy say hardware is the future, it kind of wakes you up.”
The expanding hyper-vertical list
In the years since, other companies such as Facebook and Amazon have been added to Hutcheson’s hyper-verticals list. Heated competition to develop their own deep-learning accelerators has only thrown more fuel on the fire, because for the hyperscalers, processing transactions is money in the bank. The more transactions they process, the more ads they can serve, and the more revenue they rake in.
“People tend to think about the cloud as kind of this extension of IT,” Hutecheson said. “But the difference is that IT is a cost center in an organization. In a cloud company, it’s a profit center.”
Suppliers, of course, have taken notice of what their customers want. Even though their silicon will ultimately be made by foundries, the hyperscalers’ money talks. The combined revenue of Google, Amazon, and Facebook totaled more than $400 billion last year.
What the semiconductor equipment makers are doing — especially some of the more sophisticated equipment makers like Applied Materials and ASML — is trying to understand the needs of the hyper-verticals so that they know where to focus their tool development efforts, according to Hutcheson. “Now that these high-end companies are starting to develop their own requirements, it’s beginning to drive the equipment.”
Case in point is Applied’s introduction at Semicon West of new Endura physical vapor deposition (PVD) platforms developed specifically for emerging memory technologies being used in cloud computing as well as the internet of things. One of these tools, the Endura Impulse PVD system, features up to seven deposition chambers and is geared toward the production of emerging non-volatile memories PCM and ReRAM being used to provide high-performance, low-power embedded memories for the AI era.
“This tool is very process-specific, and you will see more of these type of announcements that are linked all the way up the supply chain,” Hutcheson said.
To Gary Dickerson, Applied’s president and CEO, the development of the new Endura systems is an example of Applied connecting with and working with a broader ecosystem of customers and partners to understand what types of technologies are needed and what types of process capabilities are possible. This, he said, is what will define which companies are successful going forward as the industry moves from being driven solely by the brute-force scaling of Moore’s Law to a new playbook for innovation that relies on new materials and new architectures.
“Now, we are thinking more about how we can co-optimize all of these different capabilities,” said Dickerson. “The new Endura systems are a great example of how we can take a combination of different technologies and engineer a structure that’s not possible today.”
Hutcheson agreed. “We’ve seen more and more connectedness.”
Call them hyperscalers or hyper-verticals, the AI silicon development of tech’s heavyweight gorillas will increasingly set the agenda for semiconductor process technology going forward. Behind today’s top tier of Amazon, Facebook, Google, and Microsoft sits a second tier of mostly Chinese companies such as Alibaba, Baidu, and Tencent. Beyond that sit dozens of other companies hoping to one day design their own deep-learning accelerators.
“All of the Tencents and a lot of other companies that are in Asia, they’ll be coming,” Hutcheson said.