By Asa Fitch
Nvidia Corp., the computer-graphics giant that made a successful bet on chips purpose-built for artificial intelligence, is facing new threats to its market dominance as rivals enter a multibillion-dollar race to power everything from customer-service chatbots to robotic lawn mowers.
Qualcomm Inc., long the leading designer of silicon for smartphones, unveiled a new AI chip last month for use in cloud computing that it plans to sell next year. Graphcore, a U.K. chip-making startup that recently secured $200 million in new funding from investors including Microsoft Corp., is testing its chips and plans to release them commercially in the second half of this year.
Underlining the commercial prospects for AI hardware, Hewlett Packard Enterprise Co. on Friday reached a deal to acquire Cray Inc., a company that makes high-end equipment for supercomputers that are increasingly being deployed for AI research.
Demand for chips specialized for AI is growing at such a pace the industry can barely keep up. Sales of such chips are expected to double this year to around $8 billion and reach more than $34 billion by 2023, according to Gartner projections.
That growth is creating opportunities for everyone. But with increasing competition, it will be harder for a single company to stay on top of the race for performance, said Linley Gwennap, the president of the Linley Group, a Silicon Valley research outfit.
"When you have so many different companies, there's a new chip coming out every month, and if that one happens to grab the lead, how long are they going to stay there?" he said. "There is going to be a lot of this leapfrogging going on."
Nvidia owns about three-quarters of the market for AI chips, based on Gartner estimates. Jensen Huang, Nvidia's chief executive, said some of the most sophisticated AI algorithms recently had been developed using Nvidia hardware.
He said Nvidia also is working to make its portfolio more flexible for customers by customizing its robust general-purpose chips with AI-specific circuitry -- and developing software to speed things up further. "I think that's a really great advantage," he said.
AI software enables computers to mimic human activity such as having conversations, making recommendations and avoiding obstacles on the street. Such systems require machines to perform a gargantuan number of simple calculations at the same time. That kind of math has traditionally been done by the central processing unit, or CPU -- the brain of a computer. Yet CPUs, which make their calculations one after another, aren't as good at the task as chips with lots of small brains making calculations at the same time.
Nearly a decade ago, Nvidia recognized graphics processing units it sold for videogame enthusiasts and other niche buyers -- chips that calculate how to display animated pixels on a computer screen -- were well-suited to AI because they work largely in parallel.
The Santa Clara, Calif., company plowed investment into adapting its chips for AI, positioning itself for a boom. Nvidia booked almost $3 billion in revenue last year from selling AI chips used in data centers. Its stock price has risen more than eightfold over the past five years.
On Thursday, Nvidia reported quarterly earnings that fell 68% but beat analysts' expectations. A slowdown in data centers where the company's AI chips primarily reside was one factor that led to the decline. Cloud companies that made significant purchases last year are getting the most out of existing chips before buying new ones, Mr. Huang said.
Chip companies aren't the only ones building AI silicon. Cloud-computing giants like Amazon.com Inc. and Alphabet Inc., huge buyers of Nvidia's chips, have started to design their own to meet their customers' specific needs. Their size -- Amazon's cloud unit had $25.66 billion in revenue last year -- give them ample resources to fund the high cost of development.
"We can aggregate a bunch of customer demand on any workload type and justify investing in one of the specialized chips," said Peter DeSantis, vice president of global infrastructure at Amazon Web Services. In November, the Amazon cloud unit unveiled a custom processor called Inferentia optimized for a form of artificial intelligence called machine learning. The processors are to be deployed on AWS; Amazon doesn't plan to offer them to other companies.
Intel Corp., already the biggest supplier of chips for computer servers, has gained a sizable foothold in AI chips for data centers by tweaking some of its most powerful CPUs to do better at AI tasks.
Smaller companies, too, are building AI chips from the ground up. They include startups Mythic Inc., Graphcore and Habana Labs Ltd., which have drawn a surge of venture-capital interest after many years during which early-stage investors shunned chips.
Nigel Toon, CEO of three-year-old Graphcore, said the company is hiring hundreds of people in anticipation of its first commercial products.
The advances that the coming wave of chips enable won't be immediately apparent to most consumers. But people's interactions with digital assistants like Amazon's Alexa and Apple Inc.'s Siri should become more fluid, and devices will be able to tailor their machine brains specifically to their owners -- such as allowing a video security system to recognize a house's occupants so it can distinguish them from intruders.
Despite the rapid progress seen in recent years, AI chip development still has a long way to go before it reaches maturity, said Naveen Rao, head of Intel's AI chip business.
"It's a very interesting point in time, kind of like the early 1900s and the introduction of flight, where the basic principles of flight were discovered," he said. "Once you know these basic ideas, then you iterate on them."
Write to Asa Fitch at email@example.com