Jensen Huang, the chief executive officer of chipmaker Nvidia, has presented his decision to purchase semiconductor design firm Arm from Japan’s SoftBank as all about artificial intelligence.
“A.I. is the most powerful technology force of our time,” Huang said in a press conference on Monday and that by purchasing Arm for $40 billion, “together we will create the world’s premier computing company for the age of A.I.”
While chief executive’s often portray big M&A deals in this sort of hyperbolic language, in this case, there’s more to Huang’s claims than the usual marketing spin.
Earlier this year, Nvidia surpassed Intel to become the world’s most valuable semiconductor manufacturer. The company makes a specialized kind of computer chip known as graphic processing units, GPUs. These chips were initially designed to handle the high data volumes that need to be crunched for animation—making them the standard chip for use in computer gaming and computer-aided design.
But that same high-volume processing ability has also made them the go-to chip for artificial intelligence applications. The company’s GPUs represent more than 97% of the A.I.-specific computing infrastructure offered by big cloud service providers such as Amazon’s AWS and Microsoft’s Azure, according to a report last year from research firm Liftr Cloud Insights. Nvidia’s data center revenues overtook gaming-related sales for the first time ever in the most recent quarter, bringing in $1.75 billion.
What Nvidia’s purchase of Arm does is allow the Santa Clara, California-based GPU giant to extend its dominance in A.I. from data centers and dedicated servers down to mobile phones and small microprocessors, which are increasingly being used in connected vehicles and other networked devices—from street lamps to washing machines—that are part of the so-called Internet-of-Things (IoT).
Arm, which is headquartered in Cambridge, England, has long-dominated the market for the computer chips that power the world’s smart phones, with its designs accounting for more than 90% of the market. It is also a leader in the design of small, low-powered chips that enable smart sensors to be incorporated into everything from cars to toasters.
It was the company’s preeminence in these two areas that Masayoshi Son, the CEO of SoftBank, cited when he purchased Arm for $31 billion in 2016, saying at the time that he saw Arm as the centerpiece of a portfolio companies he was building around the promise of the IoT revolution.
Since then, Arm has enhanced the ability for mobile phones using its chip designs to run machine learning applications on the phone itself and it has also, this past year, pushed A.I. capabilities into even its smallest and most power-constrained microprocessors designs.
It also launched a business designed to help companies manage an ecosystem of millions of connected sensors and devices, in part through dedicated software that Arm would maintain. This “IoT Services” division is not being sold to Nvidia as part of the deal, and will remain owned by Softbank.
“Energy-efficient architecture”
Huang cited Arm’s experience in finding ways to make hardware that can run sophisticated A.I. programs while consuming far less energy as a key asset Nvidia was hoping to exploit across its portfolio of A.I.-dedicated chips in the future. He also said Arm would allow Nvidia to reach a whole new set of customers beyond its traditional datacenter and gaming markets.
“We would never have been able to reach the broadness and vastness of the ecosystem that Arm has created with this incredibly energy-efficient architecture they’ve created,” he said. “Energy efficiency is the single most important thing in computing going forward.”
Many analysts say Huang’s emphasis on Arm’s strength in energy-efficiency is smart. One criticism of Nvidia’s existing GPUs is how much electricity they consume. “Energy efficiency is a new superpower in semiconductor,” J.P. Gownder, vice president and principal analyst at technology analytics firm Forrester Research, said. He said there were two reasons for this: cost and climate. “Power costs and climate change come into play, particularly in large-scale deployments like data centers,” he said.
While the deal gives Nvidia sweeping power in A.I. computing for the moment, it may also provoke a challenge that could knock the company off that prized perch.
Unlike Nvidia, Arm is not a chipmaker. It simply designs chips and then licenses those designs to semiconductor manufacturers for a fee, along with a royalty payment whenever those chips are shipped to end customers. It’s a great asset-light, high-profit business model. And one of the reasons it has worked so well for Arm in the past is that the company was independent, allowing it to sell to both Apple and Samsung, as well as the likes of Intel and AMD, which also design their own chips too.
While Gownder suggested that Nvidia could potentially adopt Arm’s licensing model for designs of its GPUs too, it seems just as likely that many existing Arm customers—such as Samsung, Intel and Qualcomm—which compete with Nvidia for data center business may fear being dependent on an entity controlled by a rival for access to the latest technology.
Huang and Simon Segars, Arm’s CEO, have sought to preempt such concerns by saying that Arm’s business model will not change and that the U.K. company will continue to sell to anyone and everyone. “We see the virtue of fairness and openness and independence,” Huang said. “Continuing this business model is the best way to expand the reach of Arm and the value of Arm.” Segars said that he had spoken to Arm’s customers when rumors about Nvidia’s likely purchase of the company first surfaced earlier in the summer and sough to reassure them. “It is clear to me that Arm can maintain its business model, and the way it was operated,” he said.
“Huge opposition”
But many analysts are skeptical. “This will rightly face huge opposition, most notably from ARM licensees,” said Geoff Blaber, analyst at CCS Insight. “A huge diversity of businesses from Apple to Qualcomm are dependent on ARM and will be motivated to unite in opposition.”
Reuters quoted an unnamed South Korean chip industry source as saying the deal would raise concerns among Nvidia competitors such as Samsung and Qualcomm that Arm might hike licensing fees. “Arm customers may try to find alternatives to Arm for the longer term,” the source told the news service.
Already some chipmakers, resentful of having to pay licensing and royalty fees to Arm, have been moving to an open-sourced architecture called RISC-V, which was developed by researchers at the University of California at Berkeley and made freely available in 2014. Since then it has been adopted by dozens of semiconductor companies, including many in China, and Nvidia’s purchase of Arm could accelerate that trend, says Alan Priestley, an analyst with technology research firm Gartner. (Nvidia itself uses RISC-V in parts of its chipsets.)
Nvidia’s purchase of Arm also does nothing to patch its Achilles-heel when it comes A.I. computing. While GPUs currently dominate the market for A.I. applications, many think that a new generation of chips, designed with A.I. in mind from the start, may soon unseat them.
Already several end-users, most notably Google and Tesla, have created their own A.I.-specific chips that are designed to run neural networks—a kind of A.I. software loosely based on how the brain processes information—more efficiently than Nvidia’s GPUs. Established chipmakers from Qualcomm to Xlinx, as well as fast-growing startups like Graphcore and Cerebras, have created alternative A.I. chips featuring radically different layouts than those found in either Nvidia’s or Arm’s existing semiconductors. Intel, seeking to make up ground it’s lost to Nvidia in recent years, has purchased a number of these startups, including, most recently, Israeli startup Habana Labs for $2 billion.
Among the new designs seeking to replace the GPU are so-called “neuromorphic” chips that—like the neural networks they are designed to run—claim inspiration from the way connections are wired in the human brain.
“There are a wide range of different architectures being pursued for A.I. workloads, many of which may be more efficient than a GPU,” Priestley said. He cautioned that “A.I. is at a very early stage in it development,” and that newer hardware and software combinations could wind up displacing established market leaders.
Nvidia’s purchase of Arm keeps the company in the poll position for A.I. applications—but maybe not for long.
