Nvidia has made its most ambitious move yet to control the data center of the future. On Monday, the company said it would pay $6.9 billion to buy Mellanox Technologies, which sells networking and storage equipment to many of the same corporations that use Nvidia's chips to speed up artificial intelligence in servers. The deal follows a steep slowdown in its core graphics chip business, which started to stumble in the second half of last year.
The deal gives the Santa Clara, California-based company the ability to sell more of the chips used in the world's largest data centers. Nvidia could also offer integrated solutions combining Mellanox's devices with its own. Mellanox builds networking, storage and other hardware based on Infiniband and Ethernet used in high-performance computing, including interconnects that move data between server boxes. The deal is projected to close before the end of 2019.
Nvidia makes graphics processing units (GPUs) that contain thousands of cores that can be used to divide and conquer computationally intensive tasks. The chips were first developed to enhance video game graphics. But they also turned out to be ideal for running machine learning algorithms. Now Nvidia's processors are being stacked in data centers of major corporations and cloud computing firms, which use them to supplement computer chips from Intel and AMD.
Jensen Huang, Nvidia's CEO, sees the company's future further removed from the business that helped make it the largest seller of discrete graphics. Its chips have emerged as the gold standard for running a class of computations used in machine learning. That has threatened to push Intel, the largest maker of chips used in servers, into more of a supporting role. Nvidia's customers include cloud computing giants like Amazon and Microsoft, among others.
Mellanox sells switches and other networking products to many of the same customers. The company is a major vendor of Infiniband technology, one of the most commonly used interconnects in supercomputers, which can be used to connect servers to each other and storage with high bandwidth and low latency. Last year, Mellanox reported $1 billion in annual revenue for the first time as it took advantage of the shift to 100 Gbps Ethernet in data centers.
Nvidia is trying to take advantage of the end of Moore's Law, the chip industry's blueprint for building faster, smaller and cheaper chips, including Intel's Xeon and other CPUs. At the same time, the amount of compute required for artificial intelligence is growing exponentially. That has led to the increasing use of accelerators like Nvidia's Tesla, Microsoft's Brainwave and Xilinx's Versal, among others, which can take over workloads offloaded by the CPU.
Moving data from the processor to the accelerator, however, takes time. There are short but significant delays when sending data from the Intel's CPUs to Nvidia's GPUs, creating a performance bottleneck. Latency is also introduced when Nvidia's chips have to access the memory where machine learning algorithms are stored. Speedy interconnects are required to make sure that all these components can function efficiently.
The interconnect is important enough that Nvidia has started to design its own. Last year, Nvidia unveiled its first custom networking chip based on its NvLink technology, the NvSwitch, which enables 16 GPUs to be connected in the same server while sharing memory. But as more and more information is delivered to data centers—and fed into more and more advanced artificial intelligence algorithms—thousands of GPU may have to be combined to handle it all.
That is where Mellanox comes in. The company's Infiniband technology reduces the time it takes for computer nodes to communicate with each other in data centers. The company, which was founded in 2000, worked with Nvidia on the Department of Energy's Summit supercomputer, the fastest system in the world. The system uses more than 27,000 graphics chips stitched together in servers using NvLink, with each server connected via Infiniband.
Nvidia said that Mellanox would allow it to sell more integrated networking, storage and processor solutions, boosting performance and lowering costs for customers. Nvidia said the Mellanox deal will expand its total market opportunity to $61 billion by 2023. “I've believed for a very long time that Nvidia would go from a chip-level company to a server-level company to essentially a data center-scale company,” Huang said on a conference call to discuss the deal.
The combination comes as Nvidia's colossal growth over the last three years starts flagging. Nvidia reported slowing revenue growth in its latest quarter as demand for graphics chips used in gaming, the backbone of its business, has slowed down. That has saddled it with lots of unsold inventory, causing its profit margins to plummet. The company's gaming revenue in the 2019 fiscal year jumped 13 percent to $6.25 billion, or around 23 percent slower than in 2018.
Even though its biggest business is gaming, Nvidia's server business is booming. Last year, sales surged 52 percent to $2.93 billion despite growth getting clogged in the second half of the year by a slowdown in spending on cloud computing infrastructure. Customers delayed orders due to economic uncertainty, Huang said. Nvidia reported sales of $679 million in its latest quarter, an increase of 12 percent year over year versus a decline of 14 percent quarter over quarter.
The deal demonstrates how aggressively Nvidia will spend money to grow its business. The company said it would pay $125 per share, more than 14 percent more than Mellanox's share price at the end of last week. Colette Kress, Nvidia's chief executive officer, said that the deal would help bolster its profit margins—the selling price of its products minus the cost of production—and its earnings per share. The deal has already been approved by the boards of both companies.