The infrastructure powering today's GPU clouds—those sprawling farms of chips training everything from chatbots to protein-folding models—runs on network plumbing that most people never think about. Alex Saroyan thinks about little else.
His company, Netris, just closed a $15 million Series A led by Andreessen Horowitz, the Santa Clara startup announced late last month. The round brings Guido Appenzeller onto the board—a notable get. Appenzeller, a partner at a16z and the former chief technology officer for cloud and networking at VMware, co-founded Big Switch Networks before it sold to Arista in 2020. He knows networking fabric the way some people know their morning commute.
What drew him to Netris, apparently, was scale. The startup claims its automation software now runs across 35 GPU clusters managing roughly a million graphics processing units—the chips that have become the new oil in Silicon Valley's AI gold rush. Whether that figure represents discrete deployments or cumulative capacity across customer environments, the company didn't specify. But in a market where GPU access has become a competitive advantage, even the plumbing matters.
Traffic Jams at Machine Speed
Netris automates the tricky bits of networking across the different fabrics that AI infrastructure relies on: Ethernet, InfiniBand, and NVIDIA's NVLink. Traditional software-defined networking, Saroyan argues, buckles under AI workloads. "Traditional SDN is insufficient for AI due to traffic volumes," he told TechCrunch shortly after the funding closed.
The distinction he's drawing is technical but significant. Most SDN approaches route traffic through software controllers that can become bottlenecks when data moves at the speeds AI training demands. Netris uses what it calls hardware-accelerated automation instead—offloading enforcement to the switches themselves while a central controller orchestrates configuration. Agents on each switch communicate back via encrypted gRPC, a detail that matters in multi-tenant environments where one customer's training run shouldn't peek into another's.
The platform integrates with NVIDIA's Unified Fabric Manager for InfiniBand and Network Fabric Manager for NVLink, threading through Kubernetes and Terraform APIs in the process. It's infrastructure-layer work, decidedly unsexy, but increasingly critical as enterprises and cloud providers race to stand up GPU capacity.
Customers and Claims
Named deployments include Lightning AI, Hewlett Packard Enterprise, TensorWave, TELUS, and Foxconn's Visionbay unit. Lightning AI recently standardized on Netris for a 10,000-GPU build using NVIDIA's GB300 architecture, a configuration the startup highlighted at GTC earlier this year.
Partnerships have followed: Red Hat, Mirantis, and Spectro Cloud in recent months, each positioning Netris at the intersection of Kubernetes orchestration and the physical switches underneath. The company also claimed as of March to have captured 12 percent of the so-called neocloud market—startups like CoreWeave and Lambda Labs that rent GPU capacity by the hour—within ten months of launching targeted outreach. Independent verification of that market share figure proved elusive, though the claim suggests aggressive growth in a crowded field.
Netris has been moving fast. Product updates shipped in February, March, and May added features like VictoriaMetrics integration and broader fabric support, the kind of iterative velocity that venture investors tend to reward.
What the Money Buys
The Series A follows earlier seed and pre-seed rounds that brought in several million dollars between 2018 and 2021, though Netris hasn't broken out exact figures. The company reported 800 percent year-over-year growth in annual recurring revenue in its own press materials, a figure reiterated in coverage by TechCrunch and Forbes, though from what base remains unclear.
Saroyan said the fresh capital will fund hiring in engineering and sales, expand support for additional hardware vendors, and bankroll new algorithmic development. Perhaps more urgently, it buys time to convert customer traction into durable market position before larger networking incumbents—Cisco, Juniper, Arista—decide the AI infrastructure layer is worth defending.
Appenzeller wrote in a16z's announcement that the firm was drawn to Netris in part because multi-fabric networking for AI remains genuinely hard, the kind of technical complexity that creates moats if executed well. The firm closed a $15 billion fundraise across five funds in January, with infrastructure remaining a stated priority.
For now, Netris occupies a niche: the messy intersection of commodity hardware, bleeding-edge AI workloads, and the old-school discipline of network engineering. Whether that niche expands into a platform play or gets subsumed by bigger vendors will depend, as these things often do, on execution speed and customer lock-in. Saroyan and his co-founders—Tigran Martirosyan and Arsen Arakelyan, both with him since the 2018 founding—are betting they can move faster than the incumbents can wake up.
It's a familiar wager in Silicon Valley. Sometimes it even works.
