DriveNets and AMD Challenge Nvidia’s AI Infrastructure Dominance
2 November, 2026
The Israeli company is deepening its partnership with AMD to build a more complete AI infrastructure offering, while developing networks that connect AI clusters across distant data centers
By Yohai Schwiger
The competition between AMD and Nvidia in artificial intelligence infrastructure extends far beyond the performance of their processors. Turning thousands of accelerators into a single computing system requires a network that allows them to exchange data and synchronize efficiently. This is where Israeli company DriveNets aims to make its mark, deepening its partnership with AMD and playing a role that, in the broader infrastructure ecosystem, resembles that of Mellanox within Nvidia’s offering.
“Nvidia takes great pride in providing a complete solution. It doesn’t just hand you a GPU and leave you to figure out the rest,” Dudi Cohen, Vice President of Product Marketing at DriveNets, told TechTime. “It provides the processors, the networking solutions, and the professional services. AMD was missing the piece that connects everything together. Our solution fills that gap.”
The partnership gained a financial dimension in June, when AMD participated as a strategic investor in DriveNets’ $410 million funding round. The size of AMD’s individual investment was not disclosed.
On the engineering side, the companies published a joint reference architecture — essentially a detailed, validated blueprint for building an AI computing cluster. It specifies which components to use, how to connect them, and how to configure the system for optimal performance.
The design combines AMD Instinct MI355X accelerators and Pensando Pollara network interface cards with DriveNets’ AI Fabric networking infrastructure. It also incorporates software optimizations that coordinate accelerator operation and data exchange, allowing the processors and network to function efficiently as a unified system.
“We worked in AMD’s labs and jointly published a reference architecture. If you want to build a system based on AMD processors, this is the document you should follow,” Cohen explains. “It covers AMD’s computing hardware, our networking infrastructure, how everything connects, and how the system is supposed to operate. It also includes performance measurements we conducted together.”
The tests covered model training, inference, inter-accelerator communication, and network resilience. In certain inference configurations, the companies reported throughput improvements of approximately 3%–5% and a 10%–15% improvement in time to first token, compared with publicly available results for equivalent AMD hardware configurations.
These were internal tests conducted under specific conditions, rather than evidence of a general performance advantage over Nvidia. The two companies have also established a joint laboratory to evaluate customer workloads.
DriveNets’ key differentiator is its combination of multi-vendor Ethernet infrastructure and end-to-end system integration. The company does not develop its own switching chips. Instead, it primarily uses Broadcom silicon, building hardware, software, and services around it.
“We’ve created an alternative to InfiniBand, which has long been the performance benchmark in this market,” Cohen says. “We provide a complete solution: hardware, software, services, and optimization tools that tune parameters not only in the switches but also in the servers. We look at the entire system.”
When a Data Center Runs Out of Power
Beyond networking within individual data centers, DriveNets is developing what it calls Scale Across: technology that connects accelerators located at geographically separate facilities into a shared computing cluster.
According to Cohen, the primary driver is the growing shortage of available electrical power.
“Data centers aren’t really measured in square meters. They’re measured in megawatts — by how much electricity you can deliver to them,” he explains. “At some point, you run out of power at a particular site. You want to expand your accelerator cluster, but you can’t. One option is to find another site, perhaps 50 kilometers away, and continue expanding there.”
The challenge is particularly acute in AI training, where large numbers of accelerators work together on the same task. Distance introduces latency, while congestion and packet loss can slow down the entire workload.
DriveNets addresses these challenges through deep packet buffering, traffic scheduling, and load balancing, aiming to maximize link utilization while minimizing the impact on performance.
“This is very different from conventional data center interconnection,” Cohen says. “The requirements for bandwidth, latency, packet loss, and quality of service are much more demanding. The goal is to create what functions logically as a single system, even though it is physically distributed across multiple locations.”
A commercial implementation has already been demonstrated in Project Redwood, developed with US-based WhiteFiber. The project connects two facilities approximately 84 kilometers apart.
The companies reported a network capacity of 111.2 terabits per second and latency of 0.9 milliseconds.
Interestingly, Cohen notes that the project uses Nvidia processors rather than AMD hardware — illustrating that DriveNets continues to serve customers across competing AI ecosystems.
From Telecom Networks to AI Infrastructure
DriveNets was founded in 2015 by Ido Susan and Hillel Kobrinsky. Its original business focused on separating networking software from hardware, using software to make multiple standard switches operate as a single large router.
Its customers include AT&T, Comcast, and KDDI.
“About two and a half to three years ago, we realized that our technology was also well suited to AI systems,” Cohen recalls. “The underlying technology and expertise are shared, but we have separate development and sales teams. In some respects, the new business operates like a startup within the company.”
Looking ahead, Cohen sees a growing opportunity as alternatives to Nvidia gain traction.
“Until recently, most of our customers were using Nvidia, simply because Nvidia accounted for most of the market,” he says. “Lately, we’re seeing a significant influx of AMD customers. Within the next six months, we expect to have more AMD-based projects than Nvidia-based ones. As additional players gain market share, our opportunity grows.”
Photo: Dudi Cohen, DriveNets. Credit: Karina Shalmaev. Image edited.
Posted in: News
