Anshul Sadana spent 17 years at Arista, most of them as COO, and eight years at Cisco before that, where he worked on the high-speed switching and leaf-spine design most data centers still run on. When someone with that background starts a company and raises a $500 million Series B at a $4.2 billion valuation, the engineers he hires get interviewed the way Arista used to interview: on whether you understand what actually happens to a packet, not on whether you can reverse a linked list under time pressure.
Nexthop AI left stealth in March 2025 with $110 million and closed the Series B a year later, led by Lightspeed Venture Partners and Andreessen Horowitz. The company builds switching systems for hyperscalers and large cloud operators, both off-the-shelf and customized, running open-source network operating systems like SONiC and FBOSS on Broadcom silicon. Its early switches (the NH-4010 at 51.2 Tbps and the NH-4220 at twice that) target the fabric inside AI clusters, the network that carries GPU-to-GPU traffic while a model trains. That product focus predicts most of what you will be asked.
The role decides which way the loop leans
Nexthop had around 255 people by mid-2026, spread across silicon and hardware design, systems software, the network OS, network operations, and cloud-facing software. A dataplane or SONiC engineer gets pushed hard on C++ and how a route makes it to the ASIC. A silicon or hardware role goes toward buffer architecture, SerDes, and board bring-up. A network reliability engineer gets grilled on BGP behavior and failure modes at scale. Every loop shares a spine (coding, systems, and a values conversation), but the depth round is fitted to the job.
Here is the shape most candidates see across a full loop.
| Round | Who runs it | What it tests | Example question |
|---|---|---|---|
| Recruiter screen | Recruiter | Background, motivation, logistics | “Why Nexthop, and why now?” |
| Technical phone screen | Engineer | C/C++ coding plus basic systems reasoning | “Implement a longest-prefix-match lookup over a route table.” |
| Coding and systems (onsite) | One or two engineers | Datapath-flavored coding | “Track the best BGP path per prefix from a stream of updates.” |
| Networking depth (onsite) | Senior network engineer | Fabric behavior, RoCE, congestion control | “How can PFC cause head-of-line blocking?” |
| Architecture design (onsite) | Staff or principal engineer | Fabric design at scale | “Design the network for a 16,000-GPU training cluster.” |
| Discipline depth (onsite) | Domain expert | SONiC route path, or silicon and buffers | “Trace a BGP route from FRR to the ASIC in SONiC.” |
| Values and behavioral (onsite) | Senior leader or founder | Ownership, ambiguity, fit | “Tell me about something you owned end to end.” |
The coding rounds are systems coding
Expect C or C++, not JavaScript, and problems that resemble the actual work. A common one: parse a stream of BGP update messages and keep the current best path per prefix as routes are announced and withdrawn. Another: implement a longest-prefix-match lookup over a set of routes, then talk through how you would make it fast enough to sit in a datapath. You might get a plain data-structures problem, but the follow-ups drag it toward memory layout, cache behavior, and what breaks at ten million routes.
How you handle the follow-up matters more than whether you finished. If you write a clean trie for prefix matching and can then explain why a naive trie thrashes the cache and how a multibit or compressed trie helps, that reads as someone who has shipped forwarding code before.
The networking round is where people wash out
This is the round that separates candidates, and it is unforgiving if your networking knowledge is a mile wide and an inch thick. The questions start from real fabric behavior:
- “A link between a leaf and a spine goes down. Walk me through reconvergence. What decides how long traffic gets black-holed?”
- “Explain ECMP hashing. Why does one 400G elephant flow from a training job melt a single uplink while the others sit idle?”
- “RoCEv2 runs RDMA over a lossless Ethernet fabric. How do PFC and ECN keep it from dropping, and how can PFC cause head-of-line blocking or a deadlock?”
- “What problem does DCQCN solve, and why does AI training traffic need congestion control that plain TCP incast handling will not give you?”
If you can sketch the leaf-spine, reason about oversubscription out loud, and explain why incast from an all-reduce collective is brutal on switch buffers, you are in good shape. If you have only read about RoCE, they will find the edge of your knowledge in about two questions.
The design round: build the fabric for a training cluster
The system design interview is networking system design, not “design Twitter.” A representative prompt: design the network for a 16,000-GPU training cluster. How many tiers of switches, what radix per switch, what oversubscription ratio, and how do you stop a single failed link or switch from stalling a job where thousands of GPUs march in lockstep.
Strong answers get specific. You would talk about a rail-optimized topology so each GPU’s NIC maps to a dedicated rail of the fabric, non-blocking or lightly oversubscribed spine tiers, and why AI traffic behaves nothing like web traffic: synchronized bursts, a handful of very large flows, and tail-latency sensitivity because the slowest flow gates the whole training step. Then the interviewer pushes on failure. What happens to the job when a spine dies, how fast can you route around it, and how do you keep a PFC storm from taking down more than the link that actually failed.
Depth round by discipline
For a SONiC or NOS role, be ready to trace a route end to end. A BGP update arrives, FRR computes the best path, it lands in the kernel and in SONiC’s Redis-based databases, orchagent turns it into a SAI call, and syncd programs the ASIC. If you can name those hops and explain what SAI abstracts away from the vendor chip, you have shown you know the platform they build on. Expect questions on how config and state live in APPL_DB and ASIC_DB, and what has to happen during a warm reboot so the datapath keeps forwarding.
For silicon and hardware, the conversation moves to the chip. Shared versus dedicated packet buffers on a Broadcom Tomahawk-class device, why buffer headroom matters on a lossless fabric, cut-through versus store-and-forward and the latency they trade, and how you would debug a link that trains but throws symbol errors. These are experience questions. They are checking whether you have done this, not whether you can recite a block diagram.
Behavioral, and what a startup at this stage screens for
A senior engineer or a founder-level interviewer usually runs the values round. Nexthop is small and moving fast, so it screens for ownership and comfort with ambiguity. Expect “tell me about a time you owned something end to end when it was not clear whose job it was” and “why leave a stable role at a big company for a startup.” A candid answer lands better than a rehearsed one. If you are leaving because you want to build the whole stack instead of maintaining one component, say exactly that.
Comp and how to prepare
Nexthop does not publish salary bands, and I will not invent them. What you can reason about: it is a well-funded startup at a $4.2 billion valuation competing for the same networking talent as Arista, Nvidia, and the hyperscalers, so cash tends to track that market and the equity grant is where both the upside and the risk sit. Ask the recruiter for the band, the preferred price, and how much of the offer is equity, then check levels.fyi and recent Arista or Nvidia networking offers as a reference. Pre-IPO equity at this valuation is a bet on the number climbing from here, so size it as a bet.
To prepare, get concrete about what the product touches. Read the SONiC architecture docs until you can trace a route to the ASIC without notes. Be able to talk through BGP and EVPN behavior, RoCEv2, PFC, ECN, and DCQCN, and why lossless Ethernet is both necessary and dangerous. Know the leaf-spine well enough to reason about oversubscription and reconvergence on a whiteboard. And be ready to write forwarding-path code in C or C++ that a reviewer who has shipped it would respect. The through-line across every round is the same: they want people who understand the network down at the layer where the abstractions leak.
Practice the behavioral round:
