Nexthop AI
by Nexthop Systems Inc.
Ethernet switches and network software co-designed for AI clusters, built for hyperscalers and NeoClouds
Nexthop AI builds Ethernet switching hardware, a SONiC-based network operating system and linear optics purpose-designed for AI training and inference clusters, rather than cloud networking gear adapted to the job. Founded by Arista's former COO, it sells directly to hyperscalers and NeoClouds that co-design their own network stack and are constrained by power and cost per token rather than by list features.
Nexthop AI, legally Nexthop Systems Inc., was founded in 2024 in Santa Clara, California by Anshul Sadana, who spent 17 years at Arista Networks including a long tenure as Chief Operating Officer and eight years at Cisco before that. The company designs Ethernet switching hardware and software together for AI cluster traffic patterns, on the argument that networking — not GPUs — has become the binding constraint as thousands of accelerators exchange gradients and synchronise weights at 400G moving to 800G and 1.6 terabits per second. It emerged in March 2025 with a $110 million Series A led by Lightspeed Venture Partners with Kleiner Perkins, WestBridge Capital, Battery Ventures and Emergent Ventures, and in March 2026 raised a $500 million Series B led by Lightspeed with Andreessen Horowitz and Altimeter at a $4.2 billion valuation, taking total funding to roughly $610 million. Alongside the Series B it launched three switches built on Broadcom silicon: the NH-4010 at 51.2 Tbps, claimed up to 20% more power-efficient than comparable products and worth tens of megawatts of savings at fleet scale; the NH-4220 at roughly double that capacity; and the NH-5010, a disaggregated spine design for inter-datacenter traffic that departs from conventional spine-and-leaf. The software side is Nexthop NOS, built on the Linux Foundation's open-source SONiC with custom stacks tailored per hyperscaler, plus telemetry, congestion control and load balancing, RoCEv2 and DCQCN support. The company also ships linear pluggable and linear receive optics that cut digital signal processor requirements. It sells exclusively to hyperscale cloud providers and NeoClouds.
Infrastructure and network engineering leaders at a hyperscaler or NeoCloud who operate their own AI clusters at megawatt scale, maintain an in-house network OS, and measure success in power per port and cost per token.
AI-cluster Ethernet switching co-designed with the customer's own software stack, with a claimed 20% power-efficiency gain per device that compounds into tens of megawatts across a large fleet.
At a Glance
- Category
- Infrastructure & Cloud
- Pricing
- Contact for pricing
- Target Market
- CTOs, VP Infrastructure, Network Architects, Datacenter Operators
- Deployment
- Self-hosted
- Founded
- 2024
- Headquarters
- Santa Clara, United States
Key Features
- ✓NH-4010 switch
A 51.2 Tbps switch for AI clusters that Nexthop claims is up to 20% more power-efficient than comparable products, saving tens of megawatts at fleet scale.
- ✓NH-4220 high-density switch
Roughly doubles the NH-4010's capacity to around 102.4 Tbps for denser AI fabrics, supporting the move toward 1.6 Tbps per port.
- ✓NH-5010 disaggregated spine
A spine platform for inter-datacenter and scale-across traffic that replaces conventional spine-and-leaf topology in large multi-site AI deployments.
- ✓Nexthop NOS on SONiC
A network operating system built on the Linux Foundation's open-source SONiC, with per-customer software stacks, cybersecurity hardening and patch management.
- ✓AI-specific congestion and telemetry
Supports RoCEv2 and DCQCN for automatic congestion detection, with telemetry and load balancing tuned to collective communication rather than general cloud traffic.
- ✓Linear optics (LPO/LRO)
Linear pluggable and linear receive optical modules that reduce or remove digital signal processors, cutting optical power draw and cost per link.
Use Cases
- •Large-scale AI training fabric
A cloud operator builds a scale-out fabric for thousands of GPUs exchanging gradients continuously, where switch congestion behaviour determines job completion time.
- •Power-constrained cluster expansion
An operator whose datacenter is capped on megawatts rather than floor space converts switch power savings directly into additional deployable GPU capacity.
- •Co-designed network stack
A hyperscaler running its own proprietary network OS gets hardware and a SONiC-based stack built to integrate with its existing internal libraries.
- •Multi-datacenter scale-across
A NeoCloud links AI clusters across sites using the NH-5010 disaggregated spine rather than stretching a conventional spine-and-leaf design between buildings.
- •400G to 800G and 1.6T migration
An operator upgrades port speeds ahead of next-generation accelerators, using linear optics to hold power and cost per link down through the transition.
Ideal For
Best For
- ✓Scale-out AI training fabrics where GPU-to-GPU collective traffic, not north-south traffic, sets the design
- ✓Operators whose binding constraint is datacenter power budget rather than switch list price
- ✓Hyperscalers and NeoClouds that already run their own network operating system and want hardware co-designed around it
- ✓Inter-datacenter and scale-across connectivity using a disaggregated spine instead of conventional spine-and-leaf
- ✓Migrations from 400G toward 800G and 1.6 Tbps per port on Broadcom-based merchant silicon
Not Ideal For
- ✗Regular enterprises and mid-market buyers — Nexthop sells exclusively to hyperscale cloud providers and NeoClouds and has no described enterprise channel, partner program or self-serve motion
- ✗Organisations that need published list pricing, a reseller quote or a trial, since every engagement is a custom co-design deal
- ✗Teams standardising on a turnkey vendor NOS with a mature support ecosystem, rather than a SONiC-derived stack customised per customer
- ✗Buyers who require multi-vendor silicon options, as the announced switch line is built on Broadcom networking chips
Deployment
Market Analysis
Pros
- ✓Founder-market fit is unusually strong — Anshul Sadana spent 17 years at Arista including as COO, and a16z cites his existing hyperscaler trust as the reason customers engaged early
- ✓Power efficiency is the metric that matters to these buyers, and a claimed 20% per-device gain compounds into tens of megawatts across a large fleet
- ✓$610 million raised at a $4.2 billion valuation gives it the balance sheet to fund multi-year hardware cycles, which is the usual killer for networking startups
- ✓Building on open-source SONiC lets hyperscalers keep their own operating system and libraries instead of adopting a proprietary vendor NOS
Cons
- ✗Sells exclusively to hyperscalers and NeoClouds, so there is no enterprise or mid-market path — Network World reports the focus is on hyperscale cloud providers only
- ✗Very young for infrastructure procurement — founded 2024 and around 100 people at Series A — competing against Arista, Cisco, Juniper and NVIDIA on support, lifecycle and RMA track record it has not yet built
- ✗The switches use Broadcom merchant silicon, so differentiation sits at the system, power and software level where white-box vendors on free open-source NOS and commoditised hardware compete directly
- ✗Power-efficiency and capacity claims are vendor-stated with no published independent benchmarks, and no named customers or deployment references have been disclosed
- ✗NVIDIA's Spectrum-X and proprietary NVLink are pulling scale-up AI networking toward a vertically integrated stack, a structural headwind for any merchant-silicon Ethernet challenger
Pricing
Custom hyperscaler engagement
Contact for pricing
- ✓Co-designed switch hardware
- ✓Nexthop NOS built on SONiC
- ✓Linear optics and cables
- ✓Direct engineering engagement
- ✓No self-serve or reseller purchase path
No pricing is published anywhere and there is no trial, quote tool or reseller path — the site routes entirely to contact and support channels. Deals are custom co-design engagements with hyperscalers and NeoClouds covering hardware, a per-customer software stack and optics together, so total cost is a negotiated system price rather than a per-unit switch price. The economic argument the company makes is cost per token and power per port at fleet scale, not list price per device, which means evaluation requires a power and workload model rather than a price comparison.
Security & Compliance
Connect
Sources
This page was written from 7 sources, 6 on domains other than nexthop.ai.
- 1.nexthop.ai — nexthop.aivendor
- 2.siliconangle.com — ai networking startup nexthop ai raises 500m launches new sw
- 3.networkworld.com — former arista coo launches nexthop ai for customized network
- 4.a16z.com — investing in nexthop
- 5.datacenterdynamics.com — nexthop ai launches with 110m funding round
- 6.invezz.com — nexthop ai raises 500m at 4 2b valuation
- 7.fierce-network.com — arista rides ai past cisco now nvidia looms
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
InsightFinder
Predictive reliability for AI agents and IT estates — unsupervised anomaly detection that flags incidents before they land
Tensormesh
KV-cache-native LLM inference that bills cached input tokens at $0
OCI Enterprise AI
OpenAI-compatible agents, tools and memory on Oracle's own cloud, with the data staying put
Empirik
Change observability for infrastructure — compute the blast radius before the change lands