Netris, the network-automation startup that helps GPU cloud operators go live faster, just closed a $15 million Series A led by Andreessen Horowitz. The story is interesting not because of the headline number, but because of what the company does and, more tellingly, what it deliberately does not do.
This is a company automating the unglamorous layer — switch configuration, multi-tenancy, day-2 operations — for the wave of AI neoclouds that have appeared in the last two years. And it does it without a single line of machine learning in the product.
The Problem: Going Live Is the Hard Part
Everyone wants to be a GPU cloud right now. But buying the GPUs, switches, and storage is only the start. Actually getting a cluster configured, running, and ready to sell capacity for training and inference can take months — and every week of delay is expensive GPUs sitting idle, burning depreciation instead of revenue.
Large incumbents (Equinix, NTT, Digital Realty, Oracle, the hyperscalers) solved this by hiring ranks of network engineers or building their own automation. A small neocloud operator rarely has that luxury. They are experts at sourcing hardware and finding customers, but the network plumbing in between is where they stall.
Netris’ pitch: software that runs on the network switches and a control platform that connects to them, automating setup, configuration, and operations — plus hardware-layer network abstraction and isolation for multi-tenancy.
Why “Hardware-Accelerated SDN” Matters
Here is the detail that should catch any platform engineer’s attention. Netris CEO Alex Saroyan argues that classic SDN (software-defined networking) is falling short for AI workloads:
“For AI, software is not okay, because the amount of traffic is so high, everything must be hardware accelerated. So you need something like SDN, but completely hardware accelerated.”
That is the right instinct. AI clusters move enormous east-west traffic between GPUs — collectives like all-reduce and all-gather are bandwidth- and latency-bound, and they saturate links constantly. A control plane that punts packets through a slow software path becomes the bottleneck. Pushing the data plane into the switch ASIC (where the traffic already flows) is how you keep line rate while still getting programmatic, automated control.
The platform is vendor-agnostic, working across both Nvidia and AMD server environments and standard data-center networking gear. That interoperability matters because neoclouds are not monogamous about silicon — they mix vendors to manage cost and supply.
The Counterintuitive Bit: No AI in the Product
The most refreshing part of the announcement is the founder’s explicit stance: there is no AI in Netris’ product.
“AI is not deterministic, right? Sometimes it likes to do things on its own. It’s good for creative work, but for changing many thousands of switch configurations, you don’t need to be creative. You need to be very persistent and repeatable.”
That is a platform-engineering truth dressed up as a funding-story quote. Network configuration is a known-procedure problem: the desired state is precise, the blast radius of a mistake is enormous, and “creative” is the last adjective you want near a production switch. Deterministic algorithms that are persistent and repeatable beat a probabilistic model that might occasionally feel inspired to renumber a VLAN.
It also reflects where LLMs actually help and where they do not. We happily use AI for incident summaries, runbook drafting, and root-cause assistance. We do not let it be the source of truth for the configuration that keeps the fabric up. Netris automated the part that should never be creative.
Traction: 35+ Clusters, ~1M GPUs
This is not a slideware raise. Netris says it is live at more than 35 GPU clusters worldwide, roughly a million GPUs total, operated by names like Lightning AI, Foxconn, Visionbay, HPE, TensorWave, and Telus. Notably, Nvidia was impressed enough by an early demo two years ago to recommend Netris to its own customers — an endorsement that matters more than the check.
a16z partner Guido Appenzeller joins the board. The capital goes to hiring engineers and sales, adding more hardware-vendor support, and extending the automation algorithms.
What This Means for Platform Teams
If you run or are building an internal AI platform, three takeaways:
- The network is a first-class platform concern, not plumbing. As soon as you operate more than one tenant on shared GPUs, the boundary between “infra” and “product” runs through your switch config. Multi-tenancy, isolation, and quota enforcement at the hardware layer are exactly what separates a science project from a platform.
- Automate the deterministic parts ruthlessly. Netris’ “no AI” stance is a lesson in scope discipline. Encoding the repetitive, high-stakes, well-specified tasks as deterministic automation frees your senior engineers for the genuinely ambiguous problems — and removes the single most common cause of outages: a hand-edited config drift.
- Vendor-agnosticism is a moat for the operator, not just the vendor. If your automation assumes one chipmaker or one switch OS, supply shocks and price swings become existential. Building platform abstractions that tolerate mixed hardware is how neoclouds stay flexible.
FAQ
What is Netris? A network-automation company whose software runs on network switches and offers a platform that automates configuration and operations for AI neocloud operators.
How much did Netris raise and who led it? A $15 million Series A led by Andreessen Horowitz (a16z), announced June 25, 2026.
Does Netris use AI in its product? No. The company deliberately uses deterministic algorithms for switch automation, arguing that configuration changes must be persistent and repeatable rather than creative.
What is a neocloud? A smaller, GPU-focused cloud provider that sells AI training and inference capacity — distinct from the hyperscale incumbents.
Why hardware acceleration instead of SDN? AI clusters push extreme traffic volumes; a software data plane becomes the bottleneck. Offloading to the switch ASIC keeps line rate while preserving automated control.
Who uses Netris today? Over 35 GPU clusters (~1 million GPUs) including Lightning AI, Foxconn, Visionbay, HPE, TensorWave, and Telus; Nvidia has referred customers to the company.
The neocloud gold rush is real, but the winners will be the ones who get from “rack of GPUs” to “selling capacity” fastest. Netris is betting that the bottleneck is the network — and that the right tool for that job is boring, deterministic automation, not another AI wrapper. For platform engineers, that is a bet worth watching.
