NVIDIA's $10B+ Revenue Networking Business: How the $6.9B acquisition of Mellanox in 2020 defined AI in the data center
Discussing NVIDIA’s entrance into the networking business and how bundling networking+GPUs as an end-to-end solution led to dominance in the data center
In Silicon Valley, there is much debate on the best acquisitions of all time — from Google’s purchases of Android ($50M, 2005) and YouTube ($1B, 2006); to Facebook’s acquisition of Instagram ($1B, 2012); to Microsoft’s acquisition of Github ($7.5B, 2018) and LinkedIn ($26.2B, 2016). NVIDIA’s acquisition of Mellanox ($6.9B, completed April 2020) deserves inclusion in this elite set [link to announcement].
Mellanox, a prominent supplier of high-performance computing and networking technology, has been integral to NVIDIA’s rise to dominance in the data center. With its GPUs and Mellanox’s Networking, NVIDIA has since 20X-ed its Data Center business to $70B+ run rate across 35K customers and all the hyperscalers.
NVIDIA entered the networking business in 2020 with the $6.9B acquisition of Mellanox completed in April 2020. Unfortunately for Mellanox shareholders, the acquisition was paid in cash which was preferred by shareholders. NVIDIA’s share price has increased by more than 10X since. Today with Mellanox’s products, NVIDIA is the market leader in data center networking, with a $10B+ revenue product line that is tripling — far in excess of competitors in both scale and growth rate.
With the Mellanox acquisition, NVIDIA began its foray into networking, complementing its market leading expertise in graphics processing units (GPUs). This integration of networking and GPUs created massive performance improvements in computing workloads like AI that require high bandwidth and low latency — offering an integrated data center solution no competitor in either GPUs or networking has. The integration of Mellanox's networking with NVIDIA’s GPUs created an end-to-end computing and networking solution, leading to its HGX offering (8 Nvidia GPUs with integrated networking). HGX is purposefully designed for these highly demanding computing workloads, unhindered by networking constraints. Today, its HGX offering and its Networking products in Ethernet and Infiniband are top drivers of growth across the company.
In less than 4 years, NVIDIA has leveraged Mellanox to build a $10B+ revenue business in Networking that is tripling — entirely from scratch! Further, it has leveraged integrated networking to build the market leading HGX chip that is a top seller. NVIDIA’s advances in networking also dramatically expanded its software product offerings and dramatically increased its competitive moat.
The acquisition by NVIDIA marked a significant milestone, as it not only expanded NVIDIA's capabilities in the data center space but also underscored the growing importance of networking technologies in the era of AI and big data.
We discuss the history of Mellanox and its acquisition, the implications of bundling compute and networking in the data center, NVIDIA’s Networking business today, and the competitive landscape in AI networking.
Mellanox History — 1999-2020
Mellanox Technologies, founded in 1999, has its roots in Israel and originated as a semiconductor company specializing in interconnect networking technologies for data centers and high performance computing. The Company quickly became notorious for its innovative approaches to improving data center efficiency and performance. Notably, the Company built on open protocols Infiniband and Ethernet, which included adapters, switches, software and silicon — all key components in managing data flow and communication in complex computing environments.
While Mellanox did not create Ethernet (first developed by Xerox in the 1970s) or Infiniband (developed in the late 1990s by the InfiniBand Trade Association), the Company significantly contributed to the development and adoption of these technologies. In particular, Mellanox has been instrumental in advancing InfiniBand technology, providing a range of InfiniBand products including adapters, switches, and cables, and helping to establish InfiniBand as a leading standard for high-performance interconnects in data centers.
Pictured: Mellanox co-founder and CEO Eyal Waldman with Nvidia co-founder and CEO Jensen Huang, 2019
In 2020 at the acquisition, Mellanox was a public company with $1.3B of revenue (+40% YoY) with 65% gross margins and ~2,800 employees mainly in Israel [announcement]. As we detail later, NVIDIA massively accelerated the trajectory of Mellanox post acquisition — nearly 10X the scale while nearly 10X-ing the growth rate! Today its Networking business is above $10B of revenue and tripling. Incredibly, this understates the impact of networking on NVIDIA’s rise in the data center. The Mellanox acquisition also created the opportunity to sell integrated GPU/networking chips like NVIDIA’s HGX product, which have become top sellers with market leading compute performance.
Prior to its IPO in 2007, Mellanox had raised $89M in capital — initially from Sequoia Capital, Bessemer, USVP, Clear Ventures, Jerusalem Venture Partners. It later added strategic investors IBM and Intel Capital who were also customers. In 2010, Oracle acquired a 10.2% stake in the Company which it later sold to value investor Starboard in 2018. Multiple strategic acquirers besides NVIDIA put in bids to acquire Mellanox, including Intel (who at the time was the largest player in the data center market), Microsoft, and Xilinx (acquired by AMD for $49B in stock in 2022, which set the record for the largest chip acquisition). The acquisition took a while to be approved by regulators, but since networking was a new product offering for NVIDIA versus a direct competitor — it was ultimately approved. Further, NVIDIA was far from its dominant stature today in 2020 as it was smaller than Intel, Cisco and others at the time
Since the acquisition, Waldman has played a key role in managing NVIDIA’s presence in Israel, which today is more than 10% of its total workforce and perhaps its largest R&D hub outside of Silicon Valley. Outside of Mellanox and NVIDIA, Waldman is known internationally for his efforts to foster peace between Israelis and Palestinians by hiring over 100 Palestinian engineers for Mellanox and bringing Apple and other companies to open design centers in Rawabi. Sadly, both his daughter and her boyfriend were murdered in the Hamas attacks on October 7 [link]. Despite the deep personal loss, Waldman is still committed to economic development and jobs for Palestinians, as well as peace in the region. Waldman has since left NVIDIA but is quite involved in the Israeli tech community and on the board of Check Point.
Bundling Compute and Networking for AI
With outstanding foresight, the acquisition announcement from NVIDIA’s management laid out the 3 key aspects to providing datacenter-scale computing across the full stack:
Compute — led by NVIDIA’s GPUs, with competition from AMD and Google/Broadcom’s TPUs
Networking — now led by NVIDIA’s Mellanox Networking, with competition from Arista, Cisco, Broadcom, Juniper, HPE and Aruba, Intel and Dell EMC
Storage — fragmented market with players like Dell, NetApp, Pure Storage, Huawei, HPE, IBM, Micron and Vast
Today, NVIDIA’s products dominate both compute and networking. With Mellanox “we have every major cloud service provider and computer maker as customers”, giving the company dominance in every aspect of AI computing other than storage.
With the integrated approach, NVIDIA not only has massive performance improvements on compute workloads — but also significantly lower total cost of ownership versus purchasing separate GPU and networking solutions.
Will NVIDIA acquire/build a storage solution to own all three key pieces end-to-end? Hard to say but thus far NVIDIA has opted for partnerships in the competitive storage market. If it did, it would acquire a vendor stronger in flash-based storage like Vast as that is a key technology for AI workloads.
(Source: IDC Estimates)
NVIDIA’s Networking Unit — Today
Not only has NVIDIA’s acquisition of Mellanox and its now $10B revenue Networking business been successful on its own front; it’s also been massive in differentiating its GPU offerings and software libraries. For customers, NVIDIA now offers an end-to-end platform with integrated GPUs and networking — notably, its HGX platform (bundled GPUs and networking) which is credited as its biggest growth driver in the data center.
This integration for high bandwidth, low latency workloads are ideal for performance in AI compute and reducing bottlenecks in data transfer. Further, integrated networking significantly improves scalability across GPUs and reduces total cost of ownership with managing separate compute and networking infrastructures across vendors.
Traditional Ethernet has significant bandwidth challenges in AI performance, while NVIDIA’s offerings are built from the ground up for AI workloads and achieve near 100% effective bandwidth and 1.7X performance improvements over Ethernet. In-network computing drives massive performance improvements versus competitive GPU products like AMD GPUs and Google/Broadcom TPUs.
In addition to NVIDIA’s acquisition of Mellanox, the Company also acquired Cumulus Networks in 2020, which Wing was fortunate to have backed. While Mellanox provided high-performance networking hardware, Cumulus added the network operating system layer, creating a more comprehensive networking solution stack. Cumulus Networks specialized in developing a Linux-based network operating system (NOS), Cumulus Linux. This NOS allows for the use of standard Linux tools and practices on networking hardware, bringing a level of software flexibility and programmability to network switches that was traditionally reserved for servers.
This synergy allowed NVIDIA to offer integrated, end-to-end networking solutions with software-defined networking. Today, the company has incredible demand for high performance use cases, particularly hyperscalers like Microsoft:
“You might have noticed that NVIDIA and Microsoft are building one of the largest AI infrastructures in the world. And it is completely powered by NVIDIA’s InfiniBand 400 gigabits per second network. And the reason for that is because that network pays for itself instantaneously. The investment that you’re going to put into the infrastructure is so significant that if you were to be dragged by slow networks, obviously, the efficiency of the overall infrastructure is not as high. And so, in the places where we focus networking is really quite important.”
Further, the integration of networking and compute has allowed NVIDIA to sell fundamentally new and novel products like NVIDIA’s HGX, the AI supercomputing platform that includes 8 NVIDIA GPUs as well as interconnect and networking for plug-and-play performance. NVIDIA chips carry the value of the full-stack, not just the chip, which its large consumer internet and enterprise customers particularly like for high performance, lower cost of ownership and ease of use.
The Future of AI Networking
Though NVIDIA has built a $10B+ business in just 3 years with Mellanox in networking, the battle for networking is far from over. The Company faces competition from Arista and Broadcom in its data center networking business, as well as Cisco, Juniper and Intel in other networking workloads that are less demanding.
NVIDIA has become the largest player at over $10B of annualized revenue and tripling, a growth rate that is nearly 8X higher than any other player — far above the fastest growing direct competitors in Arista (~30%) and Broadcom (~12%). The synergy between NVIDIA's GPUs and Mellanox's networking products allows for optimized performance in handling large datasets and complex computations, which is crucial in modern data centers serving AI workloads and vastly differentiate its offerings. Further, customers only have to work with one vendor versus separate GPU and networking providers.
Going forward, we expect NVIDIA to continue to gain market share in the ~$65B annual Network Infrastructure market, which itself continues to grow. Today, its $10B networking business represents just a fraction of this market and is capturing a disproportionate share of growth from AI networking.
“And so, that kind of tells you about the growth of high-performance networking. It is an indexed to overall enterprise and data center spend but it is highly indexed to AI adoption.” — Jensen Huang on Q3 Earnings call
This was a great article, super informative and a nice history lesson. Love the attachments :) - shared with friends!
Excellent article!