The most expensive machine in an AI data center may spend part of its day waiting. A rack of GPUs can cost millions, yet its usefulness depends on an unglamorous neighbor: the network carrying data between processors, storage and applications. When packets arrive late, accelerators idle. When an inference request travels to the wrong place, users wait and operators pay. When sensitive data crosses the wrong border, the problem is no longer merely technical.
Arrcus has made a business out of that connective tissue. Founded in San Jose in 2016, the company writes the operating system and control software for routers and switches. It does not begin with a proprietary metal box. Instead, its software is designed to work across supported merchant-silicon hardware, virtual machines and public clouds. The proposition resembles a lesson the server industry learned years ago: separate the intelligence from the machine beneath it, and customers gain room to choose.
That sounds tidy. Networking rarely is. A carrier core and a leaf switch inside a data center have different jobs, failure patterns and economics. Arrcus's central claim is that one modular foundation, ArcOS, can stretch across those places without forcing the operator into a different command center at every turn. Around it sit analytics, route reflection, cloud connectivity and orchestration. More recently, the company has packaged the same foundation for AI training and inference.
The box is no longer the product
Traditional networking vendors sell an integrated system: chassis, chips, operating system and support. Integration can be comforting, especially when downtime is measured in lost customers. It can also make a network expensive to change. A buyer who wants a new chipset, port speed or form factor may discover that the software decision and the hardware decision are inseparable.
Arrcus reverses the arrangement. ArcOS is a Linux-based, microservices operating system built for routing and switching. Its data-plane adaptation layer lets the software address different forwarding hardware, while standards-based interfaces such as OpenConfig and YANG let automation tools configure and observe it. The practical attraction is not philosophical openness. It is the possibility of selecting hardware for the job, automating it with familiar systems and avoiding a forklift replacement when requirements change.
The sealed appliance
Hardware + software + lifecycle from one vendor. Operationally familiar, but choice travels as a bundle.
The Arrcus model
Independent software + qualified merchant hardware + open APIs. Choice moves to the operator.
The family names make the stack easier to decode. ArcOS runs the network. ArcIQ watches it, ingesting telemetry for performance, fault and security analysis. ArcRR handles the specialized job of reflecting BGP routes at very large scale. ArcEdge and ArcOrchestrator power FlexMCN, which connects data centers and colocations to AWS, Azure and Google Cloud with encryption and policy. The Virtual Distributed Router replaces a giant chassis with a scale-out fabric and a virtual control plane.
“The network is becoming the control plane for performance, sovereignty and economics.”Shekar Ayyar, chairman and CEO
An inference request has an address - and a budget
AI gave Arrcus a new way to explain an old expertise. Training a model requires enormous east-west flows as accelerators synchronize. Inference is different: requests arrive from users and machines scattered across the world, and the right model may live at an edge site, a regional data center or a public cloud. The best destination can change with congestion, power use, data-residency rules, model availability and price.
Arrcus splits its AI offer along that seam. ACE-AI provides Ethernet fabrics for training clusters and distributed compute, including lossless traffic, telemetry and load balancing. The newer Arrcus Inference Network Fabric, or AINF, treats inference placement as a policy-aware routing problem. An operator can express constraints such as latency target, service tier, geographic boundary and preferred model. AINF then chooses the site and path, while local systems choose the exact model replica.
This is where Arrcus's carrier history matters. Telecom operators already think in slices, service levels, distributed edges and geography. The company co-developed an SRv6 Mobile User Plane approach with SoftBank, connecting 5G traffic to cloud-native routing without preserving every legacy protocol. In 2026, TELUS began a proof of concept using AINF as a foundation for sovereign inference across Canadian public safety, government and enterprise services. The ambition is to keep sensitive workloads inside national boundaries while steering them quickly enough for mission-critical use.
Customers buy freedom, then demand proof
The likely buyer is not a small office replacing a Wi-Fi router. Arrcus aims at communication service providers, cloud operators, data-center companies and large enterprises whose networks are large enough for flexibility to have economic weight. CoreSite uses Arrcus virtual routers in its Open Cloud Exchange for cloud-to-cloud connections. The company's public references also include NTT, LinkedIn, Liberty Global and SoftBank. These are demanding environments, but public customer counts and pricing are not disclosed.
The business model follows enterprise infrastructure convention: software licenses, support and services, sold directly and increasingly through partners. Hardware makers supply qualified boxes; chipmakers supply forwarding silicon and DPUs; integrators help design and operate the complete system. That ecosystem is not decorative. A software-only vendor must answer a blunt question when something breaks: who owns the whole outcome?
Partnerships provide the answer. UfiSpace combines its switches with ArcOS, and a 2026 design places that network beside GIGABYTE systems using AMD processors and accelerators. NVIDIA integrates Dynamo, BlueField DPUs and Spectrum Ethernet with AINF, while its investment arm participated in Arrcus's 2024 financing. Fujitsu and 1Finity signed a strategic agreement in 2025 to sell, integrate and support Arrcus technology, particularly in Japan, then expanded collaboration around the energy-efficient FUJITSU-MONAKA processor. Lightstorm is taking the inference fabric toward Asia-Pacific customers.
Where Arrcus earns its right to compete
How to read the bars: They are an editorial positioning map, not measured market share. Arrcus's strongest argument is architectural flexibility; its hardest contest is distribution and trust against incumbents with decades of installed equipment.
A narrow lane between giants and open source
Arrcus competes on several fronts at once. Cisco and Arista sell deeply integrated data-center systems. Juniper and Nokia carry long histories in service-provider routing. DriveNets also disaggregates carrier routing. SONiC gives operators an open-source network OS, supported through vendors and commercial distributions. Aviatrix and Alkira approach networking from the cloud side. Each alternative draws the market boundary differently.
Arrcus's answer is breadth without returning to the bundle: the same ArcOS lineage from physical switch to virtual router, joined to analytics and orchestration, and qualified across multiple silicon families. That can reduce operational variation and preserve negotiating leverage. It also asks buyers to accept a younger vendor in infrastructure where boring reliability is a compliment. Technical range must be matched by support depth, validated designs and partners willing to stand behind them.
Funding has bought time to build that credibility. Arrcus emerged from stealth with a $15 million Series A in 2018, added $30 million in 2019, raised $28 million with strategic investors in 2021 and announced $50 million in capital in 2023. Another $30 million financing followed in 2024. A 2025 strategic investment associated with Fujitsu was reported at $67 million. Private-company totals vary because databases treat extensions, debt and strategic rounds differently, but the direction is clear: a specialist networking company has assembled a patient syndicate of venture, telecom, silicon and industrial backers.
The useful lesson is architectural
Arrcus's evolution looks like a pivot only from a distance. First came an independent network OS. Then analytics, multi-cloud and distributed routing extended the places it could run. The 5G work added programmable paths across dispersed infrastructure. AI changed the payload and sharpened the economics, but the underlying job stayed recognizable: move traffic at scale, apply policy and keep the operator free to choose the machinery.
For customers, the immediate uses are practical. A data-center team can build an IP Clos fabric on qualified hardware. A carrier can modernize edge and core routing or add SRv6-based services. A cloud team can connect regions and steer egress toward less expensive paths. An AI operator can join training clusters or route inference according to location and performance. ArcIQ gives each group a view into what the network is doing, which is particularly useful when the network is no longer one giant box.
The larger bet is that AI makes networking visible again. Cloud computing taught buyers to think of compute as fluid. Distributed inference exposes the limits of that metaphor: distance, borders and congestion return with a bill attached. Arrcus does not eliminate those constraints. It gives operators a software layer for deciding what to do about them. In infrastructure, that is often where the durable business lives - not in pretending the hard edges disappeared, but in making them programmable.