Every major platform shift in computing has produced a new networking giant, from the rise of the internet to the cloud-era emergence of merchant silicon. Today, as hyperscalers rebuild their infrastructure around AI, networking is once again up for grabs. We believe that Anshul Sadana and Nexthop AI are building the defining networking company for the AI era, and a16z is excited to be a major investor in Nexthop’s $500 million Series B.
The AI Bottleneck
One of the core bottlenecks for the rapid development and diffusion of AI capabilities is the massive shortage of GPUs. The customer demand is real, the capabilities are rapidly advancing, the spend is enormous, and clusters are being rented months or years before they go live. But GPUs don’t operate in isolation. Both training and inference workloads are massively distributed, with thousands of GPUs communicating constantly, exchanging gradients, synchronizing weights, and moving data at unbelievable speed.
And all of that communication flows through switches—the hardware that routes data between every GPU, server, and storage node in a data center. A single AI cluster can require thousands of switches organized across multiple network layers, from the scale-up fabric linking GPUs within a server rack to the scale-out network that ties an entire data center together. As clusters grow and each new generation of GPU raises the bar, each switch needs to handle more bandwidth at lower latency with greater reliability. Even today’s state-of-the-art switches can’t keep up. At every layer, networking is the bottleneck between GPU capacity and GPU output.
The Next Networking Era
Today’s networking market was built for a different era. AI demand is pushing customers to deploy each new generation of networking hardware faster than ever before, racing through 400G and 800G with 1.6 terabits/second now in development. Deployments span air-cooled and liquid-cooled environments, each with different design requirements. The specs are multiplying, the pace is accelerating, and engineering complexity is compounding. And that’s even before emerging technologies like co-packaged optics reshape switch design entirely.
Despite this, the supply side of data center switching has been consolidating. The cost of engineering, design, and integration is ballooning with every new generation. And no vendor today combines everything hyperscalers need: relentless hardware innovation with deep support for the open-source software and open standards they’ve already committed to. Enter Nexthop.
Nexthop builds high-performance Ethernet switches designed natively for AI traffic patterns rather than cloud traffic patterns. The industry has historically evolved enterprise networking products for the cloud, but Nexthop started from scratch: hardware and software co-designed for the demands of AI-scale networks, purpose-built for cloud providers running open-source operating systems.
This requires best-of-breed software for critical network functions like telemetry, congestion control, and load balancing, while maintaining an open architecture that lets hyperscalers extend and integrate into their own environments. Historically, open-source support has been treated as an afterthought. Nexthop inverts that: world-class hardware engineered around open-source from the start, combining the efficiency of highly customized products, the innovation of proprietary software companies, and the deep customer relationships of enterprise networking vendors.
Built for This Moment
Anshul spent over fifteen years at Arista, where he was integral to building the defining networking company of the last era. He is a technical leader who understands what hyperscalers need better than almost anyone in the industry—and those hyperscalers trust him deeply. There is no one better positioned to build this company.
Beyond that, the team at Nexthop is one of the strongest hardware and software teams in the industry. The relationships, technical depth, and operational experience required to win hyperscaler switching programs are extraordinarily rare. Nexthop’s team has all three, and customers have noticed.
Building a networking company that can serve hyperscalers at this scale requires getting everything right: the hardware, the software, the relationships, and the pace of execution. It’s why companies like this almost never get started. It’s also why, when they do, they can become generational.
The shift to AI is a massive platform change, and the opportunity for a new networking leader is at least as large. We’re thrilled to partner with Anshul and the Nexthop team as they build the networking company that the AI era demands.
- Investing in QuiverAI Yoko Li, Guido Appenzeller, and Martin Casado
- Investing in Temporal Sarah Wang, Raghu Raghuram, and Stephenie Zhang
- Investing in Inferact Matt Bornstein, Jason Cui, and Raghu Raghuram
- Investing in Mirelo Guido Appenzeller and Justine Moore
- Investing in Unconventional Matt Bornstein, Guido Appenzeller, and Raghu Raghuram
Shangda Xu
is a partner on the Growth investing team, focused on enterprise technology companies.
- Investing in QuiverAI Yoko Li, Guido Appenzeller, and Martin Casado
- Investing in Temporal Sarah Wang, Raghu Raghuram, and Stephenie Zhang
- Investing in Inferact Matt Bornstein, Jason Cui, and Raghu Raghuram
- Investing in Mirelo Guido Appenzeller and Justine Moore
- Investing in Unconventional Matt Bornstein, Guido Appenzeller, and Raghu Raghuram
- Investing in QuiverAI Yoko Li, Guido Appenzeller, and Martin Casado
- Investing in Temporal Sarah Wang, Raghu Raghuram, and Stephenie Zhang
- Investing in Inferact Matt Bornstein, Jason Cui, and Raghu Raghuram
- Investing in Mirelo Guido Appenzeller and Justine Moore
- Investing in Unconventional Matt Bornstein, Guido Appenzeller, and Raghu Raghuram