NVIDIA Executive: “Thanks to Mellanox, We Are the World’s Largest Networking Company”
“A lot of people do not understand that NVIDIA is the world’s largest network provider today. This came from processes that NVIDIA identified, בראש ובראשונה the recognition that to expand and develop artificial intelligence and keep pace with its progress, you need systems that can scale quickly, which is why we acquired Mellanox,” says Dion Harris, head of data center for supercomputing and artificial intelligence, in a conversation with Globes on the sidelines of the Computex computing conference in Taiwan.
The communications division of NVIDIA, which is managed from Israel and is based on Mellanox’s operations, reached an annual revenue run rate of $40 billion last month, with annual growth exceeding 200%. “To advance in AI, the network becomes critical,” Harris explains. “Thanks to it, we have built the complete solution we offer companies today.”
The days of the “agent revolution”
Harris is NVIDIA’s point man vis-à-vis the cloud giants and artificial intelligence companies. His role is to understand each company’s needs, help them design their server farm architecture, and identify future developments that will keep NVIDIA relevant against customers such as Meta or OpenAI, and in competition with AMD and Intel.
Last week, NVIDIA launched “Vera” at the Computex conference in Taiwan, a core processor intended for home computers and AI servers, directly challenging Intel, which simultaneously launched its advanced Xeon 6 chip. Following the announcement, many in the industry wondered where the strategic partnership between Intel and NVIDIA, announced only last September, had gone.
Harris rejects the claims: “Intel is still part of our strategy, and we continue to support the x86 architecture. A customer who wants to combine one of our graphics processors with an Intel processor will be able to do so, just as we allow this with ARM technology.”
Harris explains the rationale behind the development: “The new workloads in server farms, stemming from the ‘agent’ revolution, require a new type of processor. The reason we created the new processor is to answer a new market. If not for that need, we would have continued relying on Intel’s or ARM’s existing technologies.”
The server farms of the cloud giants need ever-increasing computing power. They are no longer only about training models, but also about running a variety of applications simultaneously. This phenomenon, known as “heterogeneous server farms,” requires the use of chips from different suppliers. The entry of players such as Intel, AMD and SambaNova raises the question of whether this represents share being taken away from NVIDIA.
Harris: “In the era of AI agents, systems generate content from scratch. In this reality, priorities change, efficiency in token generation becomes critical. Such efficiency cannot be achieved with a single type of processor or graphics card. It requires complex communication between processors and layers of software and hardware, deployed across different sites and sometimes supplied by external vendors. To meet these requirements, a combination of graphics accelerators is needed alongside rapid scalability and a network infrastructure that connects all the server farms. The goal is to ensure efficient data processing, from a single server rack to connectivity between remote servers.”
This is how organizational resilience is built in a reality of war and uncertainty. For your attention, Globes strives for a diverse, substantive and respectful discourse in accordance with the ethical code set out in the trust report under which we operate. Expressions of violence, racism, incitement or any other inappropriate discourse are filtered automatically and will not be published on the site.
