NVIDIA has quietly dominated this type of chip

When it comes to NVIDIA, we often think of their GPUs. Indeed, with its leading CUDA and hardware, NVIDIA's GPUs have become the top choice in the AI market. Through the acquisition of Groq's intellectual property and technology and investment in CPU technology, NVIDIA has gradually built a solid moat in computing power. However, there is another type of NVIDIA chips that have quietly dominated the market.

According to IDC data, NVIDIA became the top-revenue data center Ethernet switch provider in the first quarter of 2026. The chip layout behind it has contributed significantly.

The Quietly Rising "New Giant"

Due to the booming development of artificial intelligence and the construction of various data centers around the world, the market has witnessed significant growth. It is reported that the revenue of the Ethernet switch market reached $15.4 billion, a year-on-year increase of 39.8%. The revenue from hyperscale data centers and enterprise data centers alone reached $10 billion, a 61% increase compared to the same period in 2025.

In terms of market growth, the Americas region led with a year-on-year growth rate of 49.7%, followed by the EMEA region with a 32.2% year-on-year increase, and the Asia-Pacific region ranked third with a 25.9% year-on-year growth rate.

NVIDIA currently holds a 21.5% share of the data center Ethernet market with revenue of $2.1 billion, a year-on-year increase of 192.7%. The key factor driving this revenue growth is NVIDIA's Spectrum-X platform, which provides end-to-end network solutions, including BlueField DPU and NVIDIA LinkX cables. These products are designed for large-scale GPU clusters, which is the current direction of artificial intelligence development.

It is reported that through the integrated and collaborative design of GPUs and networks, NVIDIA has met the needs of hyperscale data centers and enterprises for AI factory network infrastructure. This structural transformation is reshaping the landscape of manufacturers in the entire data center network industry.

IDC further pointed out in the market that the demand for 400G and 800G deployments remains strong and will continue to grow in the next few years. In the first quarter of 2026, 800G switches accounted for 35.8% of the data center revenue share, while 200G and 400G switches accounted for 34.1% of the market share. These switches accounted for 70% of the global data center Ethernet revenue share.

With its AI-optimized Spectrum-X platform, this success highlights NVIDIA's growing dominance in the entire artificial intelligence infrastructure field, from GPUs and CPUs to high-performance networks. As data centers compete to expand to adapt to the AI era, NVIDIA is proving that it is not only a leader in the accelerator field but also a reliable end-to-end partner for the next-generation AI factories.

Thanks to such performance, NVIDIA has surpassed competitors such as Arista Networks, which held a 20.7% share of the data center Ethernet switch market in the first quarter of this year. Other major manufacturers include Cisco, Huawei, and HPE.

From relevant statements, it can be seen that this chip giant is paying more and more attention to network technology and regards it as a major growth driver. At the recent shareholders' meeting, NVIDIA CEO Jensen Huang said that Spectrum-X "is currently larger in scale than the sum of all other Ethernet network products."

At the earnings conference call in May, NVIDIA CFO Colette Kress said that the company's broader data center network revenue tripled compared to the previous year, reaching $15 billion. In contrast, in an earlier quarter, NVIDIA disclosed that its quarterly network business revenue was close to $11 billion, a year-on-year increase of as high as 263%.

From this, we can see the importance of the network business in the company's revenue.

An Acquisition That Changed the Landscape

NVIDIA's remarkable performance in the switch market is due to a key acquisition completed in 2019 - acquiring the Ethernet and InfiniBand interconnection giant Mellanox Technologies for approximately $6.9 billion.

At that time, this deal did not attract as much attention as the later AI GPUs. Many people thought that NVIDIA was just filling in its network product line. However, looking back today, it seems more like a strategic investment that will determine NVIDIA's competitive landscape in the next decade.

In the data centers at that time, GPUs were still just an accelerator in servers, and the network mainly played the role of "connecting devices." However, as the scale of large model training continues to expand, the number of GPUs has increased from hundreds to thousands and even tens of thousands. What really limits the efficiency of AI clusters is no longer just the GPUs themselves, but the data exchange ability between GPUs.

A popular saying in the industry well summarizes this change: "In the AI era, GPUs determine the upper limit of computing power, and the network determines the utilization rate of computing power."

For example, in a training cluster with tens of thousands of GPUs, if the network latency increases by a few microseconds, or congestion causes some GPUs to wait for data, the final loss is not just a few microseconds, but the cost of tens of thousands of GPUs idling simultaneously. For AI factories with construction costs of billions of dollars, such losses are unacceptable.

It is in this context that the value of Mellanox began to be truly realized.

As one of the most important manufacturers in the global high-speed interconnection field, Mellanox has long been deeply involved in InfiniBand and high-performance Ethernet switch chips. Its switch ASICs, network interface cards (NICs), smart NICs, and network software have accumulated profound advantages in the HPC field. After the acquisition, NVIDIA not only obtained a complete network product line but, more importantly, the ability to design GPUs, CPUs, DPUs, switches, and even optical interconnections in a unified manner.

The widely concerned Spectrum-X today is a representative of this integration ability.

It is not simply selling a switch but combining the Spectrum switch chip, BlueField DPU, ConnectX network card, LinkX high-speed interconnection, and software stack to form a complete AI network platform. Compared with the traditional "server from one company, switch from another, and network card from yet another" assembly model in data centers, NVIDIA has started to deliver a complete set of AI infrastructure to customers.

For hyperscale cloud providers, the greatest value of this model is not to reduce the number of procurement targets but to enable the collaborative optimization of GPUs, networks, and software, thereby improving the training efficiency of the entire AI cluster.

This is why more and more AI factories have started to adopt the "GPU + network" overall procurement model in recent years, rather than purchasing servers and switches separately.

In a sense, NVIDIA is no longer just selling GPUs but the entire AI data center.

However, according to an analysis provided by Semianalysis recently, in addition to winning customers with product strength, the company also seems to be relying on its influence. Semianalysis said that many new cloud executives they talked to believe that NVIDIA will retaliate if there are non-NVIDIA network devices in their clusters or if their cloud services offer AMD GPUs or TPUs. This is another reason for NVIDIA's excellent network business performance.

Although growing rapidly, NVIDIA has not stopped pursuing new technologies.

The Future of NVIDIA's Network Technology

When interviewed by networkworld in September last year, Gilad Shainer, the senior vice president of NVIDIA's network, said that data centers are evolving into a new type of computing unit, with the main computing unit shifting from CPUs to GPUs, and the functions also changing from being distributed among different components to an infrastructure supporting artificial intelligence workloads. The evolution of this infrastructure requires synchronized data transmission and involves at least four networks: the computing network, the vertical expansion network, the horizontal expansion network, and the access network.

"Today, the scale of data centers has changed. In the AI era, the data center itself has become a computing unit. We no longer ask 'How many CPUs can I buy?' but 'How can I design a data center that can run my workloads with the highest efficiency?' " Shainer said. "This transformation has fundamentally changed the way we design, connect, and optimize infrastructure. Giant data centers have become the new computing units, and the existing network architecture can no longer cope." he added.

Shainer previously explained in a NVIDIA blog: "What we need is a hierarchical design using cutting-edge technologies - such as co-packaged optics, which once seemed like science fiction."

NVIDIA has also integrated a set of algorithms into its Spectrum-X Ethernet platform. These algorithms can implement various network protocols, enabling the Spectrum-X switches, ConnectX-8 super network cards, and systems equipped with Blackwell GPUs to achieve long-distance connections without hardware changes. These Spectrum-XGS algorithms use real-time telemetry data (tracking traffic patterns, latency, congestion levels, and distances between sites) to dynamically adjust control parameters.

Developing and building Ethernet technology is a key part of NVIDIA's development roadmap. Since the first launch of Spectrum-X in 2023, NVIDIA has quickly developed Ethernet into its core R & D direction. At the same time, NVIDIA is still actively developing InfiniBand technology, which remains its core connection solution.

"InfiniBand was designed from the beginning for synchronous high-performance computing. It has features such as RDMA, which can bypass CPU jitter and support adaptive routing and congestion control," Shainer said. "It is the gold standard for large-scale AI training and connects more than 270 top supercomputers around the world. Ethernet is catching up, but traditional Ethernet designs - built for telecommunications, enterprises, or hyperscale clouds - are not optimized for the unique needs of AI," Shainer said.

"When we first started paying attention to the AI back-end network at the end of 2023, InfiniBand dominated the market with a market share of over 80%," Sameh Boujelbene, the vice president of the Dell Oro Group, wrote in a report previously. "Although InfiniBand dominates, we have always predicted that Ethernet will eventually win in large-scale applications. However, it is worth noting that the popularity of Ethernet in the AI back-end network is very fast. As the industry moves towards 800 Gbps and higher speeds, we believe that Ethernet now has a solid advantage and is expected to surpass InfiniBand in these high-performance deployments."

The 650 Group also predicted in a report: "In the next one or two years, with the popularization of 800G and the formation of 1.6T networks, Ethernet will become a more mainstream network technology, surpassing InfiniBand. The 800G cycle in the field of artificial intelligence will create new records in revenue and the number of ports."

For NVIDIA's scalable networks such as NVLink, optical devices are an important element in the connection because a large amount of bandwidth needs to be transmitted between the connected GPU silicon devices.

"We are working on increasing the computing density within a single rack so that the rack can use copper cables. Copper cables have zero power consumption, are reliable, and very cost-effective. As long as copper cables can be used, they should be used. But when customers need to expand the network to a greater distance, copper cables cannot be used because they cannot transmit such a long distance, and then optical fibers are needed," Shainer said.

Currently, NVLink can provide a two-way bandwidth of up to 1.8 TB/s per GPU, and each rack can support up to 72 GPUs. It is expected that in the next few years, faster and more capacious NVLink technology will develop rapidly to meet the needs of higher speeds and more communication between GPUs.

In the field of optical communication, NVIDIA provides pluggable optical modules for its Ethernet and InfiniBand network devices. However, NVIDIA is also fully entering the field of co-packaged optical (CPO) networks. CPO integrates network optical modules directly into the switch ASIC chip. It is expected that CPO technology will develop rapidly in the next few months and years to handle AI traffic and ultimately other network traffic that requires high performance.

Conclusion

An IDC report on the switch market, seemingly just a change in the ranking of manufacturers, actually reflects a fundamental change in the competition logic in the AI era.

In the past, GPUs, CPUs, switches, and network cards belonged to different markets; today, they are being redefined as the same set of AI infrastructure. What determines the outcome of the competition is no longer the performance of a single chip but who can truly integrate computing, networks, interconnections, software, and systems.

From the acquisition of Mellanox, to the creation of Spectrum-X, and then to the layout of NVLink, CPO, BlueField, and a complete network platform, NVIDIA has actually been answering the same question - how to make the increasingly large AI clusters run with the highest efficiency.

This is why switch chips have become NVIDIA's new growth engine.

In the next few years, as AI factories move from tens of thousands of cards to hundreds of thousands or even millions of cards, the importance of the network will only continue to increase. For NVIDIA, perhaps what really deserves attention is not how many GPUs it has sold but whether it is quietly becoming the one who sets the rules for the entire AI infrastructure.

This is a topic worthy of deep consideration for all industry participants.

This article is from the WeChat official account "Semiconductor Industry Observation" (ID: icbank). Author: Editorial Department. Republished by 36Kr with permission.