DriveNets
Last reviewed
Jun 3, 2026
Sources
16 citations
Review status
Source-backed
Revision
v1 · 1,708 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 · 1,708 words
Add missing citations, update stale details, or suggest a clearer explanation.
DriveNets is an Israeli networking-software company that builds large telecom and AI networks out of standard "white box" hardware controlled by cloud-native software instead of traditional proprietary routers. Founded in December 2015 by Ido Susan and Hillel Kobrinsky and based in Ra'anana, the company first made its name with Network Cloud, a disaggregated networking platform adopted by tier-1 carriers such as AT&T. Since 2023 it has applied the same approach to a very different problem: building an open Ethernet fabric for the back-end networks that connect thousands of GPUs in AI clusters, positioning itself as an alternative to Nvidia's InfiniBand. In June 2026 DriveNets raised a $410 million Series D at an $8.5 billion valuation, one of the largest private rounds in Israeli tech history, with chipmaker AMD joining as a strategic investor.[1][2]
DriveNets was started by two veterans of the Israeli telecom-software scene. Ido Susan, the CEO, had earlier co-founded Intucell, a self-optimizing-network startup that Cisco bought for $475 million in 2013. Hillel Kobrinsky, who serves as chief strategy officer, founded the web-conferencing company Interwise, acquired by AT&T for $121 million in 2007.[3][4]
The two built DriveNets in stealth and self-funded it for nearly four years, working on a single idea: that telecom networks could be run the way hyperscalers run their data centers. Instead of buying large, fixed, vendor-specific routers, an operator could run a software network operating system across clusters of cheap merchant-silicon boxes and scale capacity by adding more boxes. The pitch to carriers was blunt. Traffic was growing relentlessly while revenue stayed flat, and stitching together gear from ten or more equipment makers was expensive and slow. DriveNets claimed it could cut a carrier down to roughly two suppliers and save around 40 percent on network cost.[3]
The core product, Network Cloud, separates network software from network hardware. A router has traditionally been a single integrated appliance: the chassis, the line cards, the forwarding chips, and the operating system all came from one vendor and were sold as one unit. DriveNets breaks that apart. Its operating system, DNOS, runs as cloud-native software on standard white boxes built around merchant chips from Broadcom, and a cluster of those boxes behaves like one very large logical router.[3][5]
This is an implementation of the Distributed Disaggregated Chassis (DDC) model defined through the Open Compute Project, which AT&T helped originate for its core network. A DDC fabric uses two box roles: leaf units that connect to the outside world, and spine units that interconnect the leaves. Internally the fabric chops packets into fixed-size cells and spreads them evenly across all available paths, so a single heavy traffic flow cannot congest one link while others sit idle. DNOS folds together functions that used to need separate appliances, including forwarding, quality of service, VPN services, security, and telemetry.[3][6]
AT&T became the flagship customer, running DriveNets software in its core network, and other large operators followed. By 2025 the company reported roughly 100 customers, with Comcast and Japan's KDDI among the named names, and a headcount of about 500 across Israel, the United States, Japan, and Romania.[2][7]
The growth of large machine learning models created a new networking problem that turned out to suit DriveNets well. Training a big model spreads the work across thousands of accelerators that must constantly exchange gradients over a dedicated "back-end" network. That network's job is to keep the expensive GPUs fed. As Susan put it, "the most expensive idle asset in the world right now is a GPU waiting on the network."[1]
For years the default back-end fabric was InfiniBand, the low-latency interconnect that Nvidia acquired with Mellanox in 2019. InfiniBand delivers excellent performance but comes from essentially a single vendor, which locks operators into one supplier for a critical and costly layer. DriveNets argued that Ethernet, the open and universally understood data-center technology, could match InfiniBand if it were re-engineered to behave deterministically rather than on a best-effort basis.[8]
That re-engineering is the heart of Network Cloud-AI, announced in 2023. It applies the same scheduled, cell-based DDC fabric from the telecom product to the AI back-end. DriveNets calls the technique Fabric-Scheduled Ethernet. Rather than letting switches forward traffic greedily and then react to congestion, the fabric slices traffic into cells, uses virtual output queues with receiver-based scheduling so a destination only pulls data it is ready to accept, and spreads everything across every link. The result is advertised as a lossless, congestion-free, jitter-free fabric that needs little hand-tuning and degrades gracefully when a link fails. A single Network Cloud-AI system is designed to connect up to 32,000 GPUs at 800 Gbps per port.[5][9]
In October 2024 DriveNets and the Taiwanese hardware maker Accton launched white boxes for the fabric built on Broadcom's Jericho3-AI chips. In a 512-GPU production cluster, the company reported its Ethernet fabric beat InfiniBand on standard NCCL collective-communication benchmarks by up to 18 percent, and by about 6 percent on average, evidence it cites for the claim that open Ethernet can now exceed the proprietary alternative at scale.[9][10]
DriveNets was self-funded until 2019 and has since raised about $1 billion across four venture rounds, with a large secondary sale layered in along the way.
| Round | Date | Amount | Lead investors | Valuation |
|---|---|---|---|---|
| Series A | Feb 2019 | $110M | Bessemer Venture Partners, Pitango | ~$400-500M[3][11] |
| Series B | Jan 2021 | $208M | D1 Capital Partners | undisclosed[7] |
| Series C | Aug 2022 | $262M | D2 Investments | over $2.5B[12] |
| Secondary | Oct 2025 | ~$650-800M | AT&T (15% stake) | $5B[13][14] |
| Series D | Jun 2026 | $410M | Bessemer, Atreides Management | $8.5B[1][2] |
The Series A, announced when the company came out of stealth in February 2019, was one of the largest first rounds in Israeli history. The Series C in August 2022 was led by a new investor, D2 Investments, with returning backers Bessemer, Pitango, D1 Capital, Atreides, and Harel.[12]
The AT&T transaction in October 2025 was unusual. It was a secondary sale, meaning AT&T bought roughly a 15 percent stake from existing shareholders such as Pitango, Bessemer, Harel, and Poalim Equity rather than putting fresh capital into the company. Reported figures ranged from about $650 million to $800 million, valuing DriveNets at around $5 billion and giving its decade-long investors a first major payout. The deal signaled AT&T's intent to act as a strategic partner, not just a customer.[13][14]
On June 1, 2026, DriveNets announced a $410 million Series D co-led by Bessemer Venture Partners and Atreides Management, with two new investors, AMD and Red Dot Capital Partners, joining alongside existing backers Pitango and D1 Capital. The round set an $8.5 billion valuation, which made DriveNets Israel's second most valuable private company at the time, behind storage firm VAST Data.[1][2][15]
The round reflected the surge in AI infrastructure spending. DriveNets said it had been cash-flow positive since 2025 and was sitting on more than $1 billion in signed business and backlog, much of it tied to its AI fabric. The company said it would use the money to scale inventory for its growing AI pipeline and to expand what it calls heterogeneous AI infrastructure, meaning fabrics that mix accelerators and gear from multiple vendors rather than a single stack.[1][2]
AMD's participation is the detail worth flagging. AMD sells Instinct GPUs that compete with Nvidia's, and an open Ethernet fabric that works equally well with any vendor's accelerators is exactly the kind of ecosystem AMD wants to encourage, since it weakens the pull of Nvidia's tightly integrated InfiniBand-plus-GPU bundle. DriveNets already listed AMD, Broadcom, Dell, and Supermicro among its hardware partners, so the investment formalized an existing relationship.[1]
DriveNets sits in a crowded field. On the telecom side it competes with the incumbent router vendors Cisco, Juniper Networks, and Nokia, which is why Israeli press has long framed it as a challenger to Cisco and Juniper. In AI back-end networking it competes most directly with Nvidia, whose InfiniBand and Spectrum-X Ethernet products dominate GPU clusters, and with merchant-switch sellers such as Arista Networks and Cisco that also push Ethernet for AI.[14][16]
DriveNets is not alone in betting on Ethernet. The industry-wide Ultra Ethernet Consortium, formed in 2023, is standardizing Ethernet enhancements for AI and high-performance computing, and most large network vendors now offer some AI-Ethernet line. DriveNets' particular wager is that a fully scheduled, disaggregated fabric, the architecture it already proved in carrier networks, is the cleanest way to get InfiniBand-class behavior out of open hardware. Whether that thesis holds as Nvidia pushes its own Ethernet products and the Ultra Ethernet specifications mature is the open question over the company's next chapter.[8][9]