Nexthop AI raises $500M Series B as networking becomes its own AI category

For most of the last decade, networking gear was treated as a commodity layer under whichever GPU or cloud was capturing the headline. Nexthop AI is the latest sign that thesis is changing. The company has raised $500 million in Series B funding to scale open-source-derived switching infrastructure aimed specifically at AI data centers.

The headline number is striking, but the bigger signal is the round itself. Series B investors do not typically write half-a-billion-dollar checks for an arms-supplier company; they do it when they believe the supplier has carved out a category that other players cannot collapse into a feature.

#Why networking, why now

Training a frontier model is no longer compute-bound in any naive sense. The bottleneck has shifted to the fabric between GPUs — east-west bandwidth, deterministic latency, and the firmware that arbitrates collective operations. Hyperscalers have been quietly building this in-house for years. Nexthop is betting that everyone else — neoclouds, sovereign AI projects, GPU-rich startups — will not be able to.

The company has not detailed the unit economics, but the open-source NOS underpinning makes the pitch straightforward: hyperscaler-grade fabric without the hyperscaler price tag, on commodity silicon, with software the customer controls.

#What the round says about the wider market

After a tight 2024 and a recovering 2025, Series B and C activity in deep-tech infrastructure has decisively reopened in 2026. Sequoia, a16z, Lightspeed, Insight, and Coatue remain the names attached to most break-out rounds, and the verticals concentrating capital — AI, semiconductors, cybersecurity, robotics — share a common feature: defensible IP and long lead times to replicate.

Nexthop fits that pattern. The question is whether networking remains a venture-fundable category once the early customers wire themselves in, or whether — as has happened before with storage and database tooling — the hyperscalers eventually absorb the layer.