Nexthop AI scores $110m
- Nexthop AI came out of stealth on March 25, 2025 with $110 million, aiming to build custom networking gear for hyperscale AI cloud operators. - The company is led by former Arista executive Anshul Sadana and says it can cut hyperscalers’ product cycles by 6 to 12 months. - The bet is that AI bottlenecks are shifting from chips alone to networking, power, optics, and cluster efficiency.
AI infrastructure is starting to look less like a pure chip story and more like a plumbing story. That is the backdrop for Nexthop AI, which launched from stealth on March 25, 2025 with $110 million to build custom networking systems for the biggest cloud companies. The pitch is simple — training and running giant AI models now depends on moving data between huge clusters of GPUs fast, cheaply, and without wasting power. That sounds boring next to model demos, but turns out it is exactly where a lot of the pain is. ### What does Nexthop actually make? Nexthop is not trying to be another foundation-model company. It builds networking hardware to a customer’s specs, hardens the network operating system the customer wants to use, and ties in optical and electrical interconnects from different suppliers. Basically, it is selling a custom stack for hyperscalers that do not want off-the-shelf networking if that means giving up efficiency. (nexthop.ai) ### Why is networking suddenly the problem? A giant AI cluster is only as useful as the network connecting the GPUs inside it. If the links are slow, hot, power-hungry, or hard to tune, expensive accelerators sit around waiting. Sadana’s argument is that AI networking has become a different engineering problem from ordinary datacenter networking — with signal integrity, latency, thermal management, and power per port all becoming first-order constraints. (nexthop.ai) ### Why not just buy from Cisco or Arista? That is the gap Nexthop says it sees. Traditional vendors sell broadly useful products. Hyperscalers increasingly want systems shaped around their own racks, supply chains, software preferences, and deployment timelines. Lightspeed framed this as a build-versus-buy problem that no longer fits cleanly into either bucket. Nexthop’s answer is a JDM-style model — more like an extension of the customer’s engineering team than a standard box vendor. (networkworld.com) ### Why is Anshul Sadana the key name? Because this is a credibility play as much as a technology play. Sadana spent 17 years at Arista, including as COO and chief customer officer, and before that worked at Cisco. That matters because hyperscale networking is a trust business. If you are asking the world’s biggest cloud operators to let you into a critical layer of their stack, experience in hardware design, manufacturing, supply chain, and customer deployment is not a nice extra — it is the whole entry ticket. (lsvp.com) ### What is the concrete promise? The clearest one is speed. Sadana said companies building these systems themselves can compress product-development cycles by 6 to 12 months by working with Nexthop. That is a big claim, but you can see why investors like it. In an AI buildout, shaving even a few months off deployment can matter more than squeezing out a tiny benchmark gain later. (networkworld.com) ### Why did investors write such a big check? Because the market behind this is huge if the thesis is right. Nexthop’s launch materials point to hyperscalers spending billions on GPU and networking deployments, with more than two gigawatts of capacity being added annually. Lightspeed also framed the opening as part of a cloud-switching market that could reach $75 billion by 2029, including a $35 billion cloud segment. (networkworld.com) The idea is classic picks-and-shovels — sell into the AI boom without betting on which model wins. ### What is the catch? Custom infrastructure is hard to scale. The more tailored the product, the harder it can be to repeat the business cleanly across customers. And hyperscalers are tough buyers — they have leverage, deep internal teams, and brutal performance requirements. Nexthop is betting that AI’s pace of change makes those customers want a specialist partner anyway. (nexthop.ai) ### Bottom line? Nexthop’s launch matters because it shows where the AI infrastructure bottleneck is moving. Chips still matter most at the headline level, but the next fortunes may be made in the layers that keep giant GPU clusters fed, linked, cool, and productive. Nexthop is trying to be one of those companies. (nexthop.ai) (networkworld.com)