Mellanox
Last reviewed
Jun 3, 2026
Sources
16 citations
Review status
Source-backed
Revision
v1 · 2,111 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 3, 2026
Sources
16 citations
Review status
Source-backed
Revision
v1 · 2,111 words
Add missing citations, update stale details, or suggest a clearer explanation.
Mellanox Technologies, Ltd. was an Israeli-American semiconductor and networking company that designed high-performance interconnects for servers, storage systems, and data centers. Founded in 1999, Mellanox grew into the leading supplier of InfiniBand silicon and a major vendor of high-speed Ethernet, building the network adapters, switches, cables, and programmable processors that wired together many of the world's fastest supercomputers and largest cloud data centers. In 2019 Nvidia agreed to buy the company for roughly $6.9 billion, prevailing over a field of larger rivals that reportedly included Intel, Xilinx, and Microsoft. The deal closed on April 27, 2020 and turned Mellanox into the foundation of Nvidia's data-center networking business, a unit that would prove central to the company's strategy of selling whole "AI factories" rather than individual chips. [1][2][3]
Mellanox was founded in May 1999 in the northern Israeli town of Yokneam Illit, with financial and sales operations established in Santa Clara, California (later relocated to Sunnyvale). The founding team came largely out of Intel and Galileo Technology, an Israeli chip company that Marvell acquired in 2000. The co-founders included Eyal Waldman, who served as president and chief executive officer for the company's entire 21-year independent life, along with Michael Kagan, Shai Cohen, Roni Ashuri, Eitan Zahavi, Evelyn Landman, and several others. Kagan, a Technion-trained engineer and former Intel architect, became Mellanox's chief technology officer. [4][5]
The company was created specifically to build silicon for InfiniBand, a then-emerging switched-fabric interconnect aimed at replacing the aging PCI bus and at linking servers and storage with very high bandwidth and very low latency. InfiniBand did not yet have a settled name when Mellanox began, and the standard's early commercial prospects were uncertain. Several large backers, including Intel and Microsoft, scaled back their InfiniBand ambitions in the early 2000s, but Mellanox stayed the course and focused the technology on the high-performance computing (HPC) market, where its combination of throughput and latency was hard to match. That bet paid off as scientific computing clusters proliferated: by the late 2000s, Mellanox interconnects were appearing throughout the TOP500 list of the world's most powerful supercomputers. [4][6]
Mellanox held its initial public offering in February 2007, listing on the NASDAQ under the ticker MLNX and raising about $102 million; the company was also listed on the Tel Aviv Stock Exchange until it delisted there in 2013. Oracle later became a significant shareholder, at one point holding roughly a tenth of the company. [4]
Mellanox built end-to-end "interconnect" portfolios, meaning it supplied every piece needed to move data across a cluster: the silicon, the network adapter cards installed in servers, the top-of-rack and director switches, the optical and copper cables, and the software stack. A distinctive feature of its design philosophy was Virtual Protocol Interconnect, which let a single adapter speak either InfiniBand or Ethernet, giving customers flexibility between the two fabrics.
| Product line | Category | Role |
|---|---|---|
| ConnectX | Network adapters (NICs / HCAs) | Multi-protocol adapters supporting InfiniBand and Ethernet, with hardware offloads such as RDMA |
| Quantum | InfiniBand switches | High-radix switches carrying InfiniBand traffic at speeds reaching 200 Gb/s (HDR) and beyond |
| Spectrum | Ethernet switches | Switch silicon and systems for cloud and enterprise Ethernet networks |
| BlueField | Data processing units (DPUs) | Programmable Arm-based system-on-chip combining a network adapter with on-board compute for storage, security, and network offload |
| LinkX | Cables and transceivers | Copper and optical interconnects that complete the end-to-end fabric |
The ConnectX adapter family was the company's flagship and the basis of much of its revenue, progressing through successive generations (ConnectX-3, -4, -5, -6 and so on) that roughly doubled bandwidth each cycle. Mellanox was an early and aggressive promoter of Remote Direct Memory Access (RDMA), a technique that lets one computer read or write the memory of another without involving either machine's CPU, dramatically cutting latency for distributed workloads. Its RDMA over Converged Ethernet (RoCE) work later became an important ingredient in Nvidia's AI-oriented Ethernet products. [6][7]
Mellanox expanded both organically and through acquisition. In 2011 it bought the Israeli scale-out fabric company Voltaire for about $8.75 per share, or roughly $208 million, strengthening its InfiniBand switching and software. In 2013 it acquired the silicon-photonics firm Kotura and the optical-interconnect company IPtronics to push into integrated optics. The most consequential purchase came in February 2016, when Mellanox completed its acquisition of EZchip Semiconductor for approximately $811 million. EZchip, an Israeli network-processor specialist, had itself earlier acquired the many-core processor startup Tilera. That lineage of programmable, Arm-based network silicon fed directly into the BlueField DPU line, which Mellanox introduced in 2016 and which Nvidia would later make a centerpiece of its data-center roadmap. [4][8]
By the time of the Nvidia deal, Mellanox was a substantial business in its own right. It reported revenue of about $1.09 billion for fiscal 2018, the first time it exceeded $1 billion in a year, and employed roughly 2,900 people, the bulk of them in Israel, where it was one of the country's largest technology employers. [4][9]
Mellanox became an acquisition target after years of pressure from activist investors who argued the company was undervalued. The process effectively began in 2018: Marvell reportedly approached Mellanox that year but was rebuffed, and the hedge fund Starboard Value pushed for changes. Mellanox engaged bankers to run a sale, and through late 2018 and early 2019 a competitive auction took shape among several of the largest names in chips and cloud computing. Intel was reported to have bid in the range of roughly $5.5 billion to $6 billion in January 2019, while Xilinx and Microsoft were each associated with offers near $5 billion. [2][10]
Nvidia ultimately outbid the field. On March 11, 2019 the two companies announced a definitive merger agreement, signed the prior day, under which Nvidia would acquire all outstanding common shares of Mellanox for $125.00 per share in cash, valuing the company at approximately $6.9 billion in enterprise value. The price represented a premium over Mellanox's recent trading levels (its shares had closed around $114 on March 8, 2019, valuing it near $5.9 billion) and exceeded the reported rival bids. At the time it was by far the largest acquisition in Nvidia's history. [1][2][3]
For Nvidia's chief executive Jensen Huang, the logic was that the unit of computing was shifting from the individual server to the entire data center. "The emergence of AI and data science, as well as billions of simultaneous computer users, is fueling skyrocketing demand on the world's datacenters," he said, framing Mellanox's networking as the complement to Nvidia's accelerated computing. Waldman, for his part, said the two firms shared "the same vision for accelerated computing" and described the combination as a natural extension of a longstanding partnership, the two companies having already collaborated on supercomputing systems. [1]
The transaction required clearance from antitrust regulators on several continents, and although it raised few competitive concerns in most markets, it became entangled in the broader US-China technology tensions of 2019 and 2020. Approvals came in from the United States, the European Commission, and Mexico, leaving China's State Administration for Market Regulation (SAMR) as the final and slowest hurdle. SAMR conducted a lengthy review focused largely on conglomerate effects, that is, the risk that Nvidia might tie or bundle its GPUs with Mellanox's networking gear to disadvantage competitors. [11][12]
On April 16, 2020, after more than a year, SAMR granted conditional approval. The conditions required the combined company to keep selling Nvidia GPUs and Mellanox interconnects separately on fair, reasonable, and non-discriminatory (FRAND) terms, to preserve interoperability with rivals' products, to refrain from tying or bundling, and, in a novel provision, to give Chinese customers the option to procure and stockpile up to a year's supply of the two product lines. With the last clearance secured, Nvidia completed the acquisition on April 27, 2020. The final transaction value was reported at approximately $7 billion. [11][13][2]
The table below summarizes the deal's key milestones.
| Date | Event |
|---|---|
| 2018 | Activist pressure builds; Mellanox rejects a reported Marvell approach and runs a sale process |
| January 2019 | Intel reportedly bids roughly $5.5 billion to $6 billion; Xilinx and Microsoft also linked to offers |
| March 10, 2019 | Nvidia and Mellanox sign the definitive merger agreement |
| March 11, 2019 | Acquisition announced publicly at $125 per share, about $6.9 billion |
| April 16, 2020 | China's SAMR grants conditional approval, the final regulatory clearance |
| April 27, 2020 | Deal closes; Mellanox becomes part of Nvidia (about $7 billion) |
| November 2020 | Founder and CEO Eyal Waldman departs |
After the close, Mellanox was rebranded as the core of "Nvidia Networking," and its leadership was woven into Nvidia. Michael Kagan, Mellanox's longtime CTO, became chief technology officer of Nvidia in 2020, a notable elevation of an executive from the acquired company. Eyal Waldman, however, chose to leave roughly six months after the deal closed, in November 2020, saying that after building a company over more than two decades he did not want to be "number two." [5][14]
The Mellanox product families carried forward and accelerated under Nvidia's ownership, often renamed to fit the parent's branding. The InfiniBand switch line continued as Nvidia Quantum, with the Quantum-2 generation delivering 400 Gb/s connectivity and in-network computing features such as the SHARP collective-operations engine. The Ethernet line evolved into Spectrum-X, an Ethernet platform tuned specifically for AI training and inference traffic that drew directly on Mellanox's RoCE and congestion-control heritage. The BlueField DPU became a strategic pillar, positioned by Nvidia as a third major processor category alongside the GPU and CPU, offloading networking, storage, and security from the host so that GPUs could devote themselves to computation. [6][15]
Mellanox's importance to the AI-infrastructure era is hard to overstate. Training a large neural network is not a single-chip problem but a distributed one, spread across thousands or tens of thousands of GPUs that must constantly exchange gradients and parameters. At that scale, the network becomes a primary determinant of performance: a slow or congested fabric leaves expensive accelerators idle. By acquiring Mellanox, Nvidia gained the ability to design the interconnect in concert with the GPU, ensuring that bandwidth, latency, and collective-communication patterns were matched to its hardware. Combined with Nvidia's own scale-up NVLink fabric, the Mellanox-derived Quantum InfiniBand and Spectrum-X Ethernet products gave the company an end-to-end story spanning a single server up to data-center-wide clusters, which it began to market as integrated "AI factories." Networking grew from a rounding error into a multibillion-dollar segment of Nvidia's revenue in the years that followed. [6][15]
The acquisition's regulatory legacy also lingered. In December 2024, amid escalating US restrictions on chip exports to China, SAMR reopened scrutiny of the deal, and in 2025 it issued preliminary findings that Nvidia had breached the conditions attached to the 2020 approval. The renewed probe underscored how a networking takeover negotiated in 2019 had become a flashpoint in the geopolitics of artificial intelligence hardware. [12][16]