DriveNets: The AI networking fabric for heterogeneous AI

The buildout of AI infrastructure has produced a new generation of compute companies, yet no new networking companies have emerged. This isn’t because networking doesn't matter in AI, but because for years, there was no real market to invest in. NVIDIA’s stack was bundled and proprietary, and cloud-based networking gear was good enough. 

Every major infrastructure shift catalyzes a networking giant. Cisco defined the broadband era, Arista defined cloud, and we believe the AI era will produce the next defining one. This conviction is why we’re leading DriveNets' $410 million Series D—a leader in AI networking solutions and a company we’ve backed since their Series A—to accelerate inventory build-out against more than $1 billion in secured business and rising demand for open, multi-vendor AI networking.

Performance comes from a systems approach

NVIDIA customers think they’re purchasing the best-performing chips, but what they’re really buying is an end-to-end systems performance guarantee that no one has been able to match because the competition has only ever been silicon companies. For years, AI compute was dominated by training and building the models. Training rewarded homogeneous, integrated infrastructure, which is precisely what NVIDIA provided. The market wasn't ready for a disaggregated stack and heterogeneous compute before, but it is now.

The industry calls this the "inference flip": the moment inference became the largest and most challenging part of AI compute. Every time someone uses an AI product, a cloud provider charges the company behind it per token. Reasoning prompts consume 25 times the tokens of a simple query, while agentic workloads require 500 times as many. The data center has effectively become a token factory, and the only things that matter are throughput and efficiency. Homogeneous infrastructure built for a single workload type can’t serve a factory running dozens of them, and from different types simultaneously. The monolithic inference stack is fracturing, and the opportunity for an AI networking giant is clear.

When the network becomes the bottleneck

When NVIDIA acquired Mellanox in 2020—another Israeli networking company we invested in back in 2000—it gave NVIDIA the systems expertise to deliver a fully validated, end-to-end compute platform. It also closed the market. But three things have changed simultaneously since then: 

• New and diverse workload demands made NVIDIA's homogeneous stack insufficient
• Serious compute alternatives from AMD and others made leaving NVIDIA viable
• The economics of running a token factory at scale made staying wedded to NVIDIA prohibitively expensive

As a result, customers have begun embracing alternatives—and the systems/networking problem they're escaping into is formidable. Networking solutions built for the cloud can’t be retrofitted. AI data center traffic shifted from north-south to east-west, GPU to GPU, and it needs to be synchronized and simultaneous (unlike client-to-server traffic that has tolerance for delay). As a result, interconnect requirements have grown three times faster than GPU count, and the network is now the new bottleneck.

Optimizing the full AI stack end-to-end

True end-to-end performance optimization for large-scale AI clusters requires supporting scale-up, scale-out, and scale-across networking architectures, along with front-end and storage connectivity, which DriveNets does well.

However, what makes this hard to replicate is how deep DriveNets went. Rather than adding a software layer on top, they rewrote the low-level compute and collective communications libraries that govern how GPUs coordinate during training and inference—changing how traffic is generated in the first place. Their first announced validated end-to-end reference architecture is with AMD. And rather than managing congestion at the switch, DriveNets manages it at the lowest level of the stack, delivering dramatically better throughput and lower latency under the conditions AI workloads routinely produce.

An N-of-1 opportunity

Since DriveNets isn’t tied to any compute vendor, it can work with all of them: NVIDIA, AMD, ARM, Cerebras, and whoever comes next. And because it’s fundamentally a software company, it can run on any switching silicon as well.

The central irony of the heterogeneous compute era is that the more fragmented the inference compute layer becomes, the more critical it is to have a single unified networking stack beneath it. There may be a dozen compute platforms competing for this market, while only one networking vendor is capable of building validated, end-to-end systems across all of them—scale-up and scale-out, silicon-agnostic, and deep enough in the stack to actually matter. That is what makes DriveNets an N-of-1 company.

Our relationship with DriveNets dates back more than 15 years, when we backed DriveNet’s CEO and Co-founder, Ido Susan’s previous networking startup, Intucell, as a sole investor. We’re excited to continue this relationship and further anchor our conviction in Ido and the rest of the DriveNets team.