- Arrcus' Inference Network Fabric aims to solve growing latency and performance challenges
- The fabric aims to deliver programmable, policy‑driven networking
- Arrcus says its technology helps position telcos to capture value in the AI wave
As AI moves from training in data centers to inferencing at the edge, networks need to deliver performance, latency and low cost to meet AI demands. Arrcus is stepping up, with a new Inference Network Fabric.
While training is centralized, inferencing is spread out. Inferencing hardware and networking requirements are different for different applications — autonomous driving vs. oil rigs vs. retail points of service. However, the same application might have different requirements at different times.
Existing paradigms break down
"There's no way you will be able to take the existing paradigms of the cloud world and translate that over to AI inferencing," Arrcus CEO Shekar Ayyar told Fierce Network. Networking fabric needs to divert traffic based on policies for latency, throughput, time to first token and power consumption for different inferencing compute clusters and mini data centers. "The Arrcus Inferencing Network Fabric will enable our customer to do that," Ayyar said.
The software is capable of running on third-party hardware from companies Arrcus partners with, including Broadcom, Nvidia, Intel, AMD and Cisco Silicon One for networking silicon and vendors such as UfiSpace, Edgecore, Accton, Celestica, Quanta and Lanner providing hardware.
The company has a track record providing its ArcOS operating system and ACE orchestration and telemetry to customers including telcos, data center operations and large enterprises, Ayyar said.
SoftBank implemented ArcOS for its SRv6 mobile user plane, and Liberty Global for its core internet services using virtual distributed routing (VDR) architecture for scalability.
Business is booming
Arrcus tripled its business year-over-year in 2025. The company launched in 2016 but took years to hit its inflection point for growth. Telcos and data center operators are conservative, Ayyar said. "You spend a lot of time proving yourself and then once you prove yourself, the fruits of the work you have done start showing. And we're in that phase now," he said.
The vendor has raised just under $200 million in rounds through Series D, and is gearing up for Series E. Ayyar told Fierce in 2024 that it planned an IPO in 2026 or 2027. The vendor now expects the IPO late next year, he said. Previously, a company with $100 million bookings volume could successfully IPO, but now that threshold is $300 million.
Enabling new telco business
Arrcus sees itself as an enabler for telcos to build new business from AI, making up for the past, when telcos have missed out on opportunities such as the cloud and the training wave of AI, Ayyar said. Telcos are well positioned to provide inference services, because inference requires distributed architecture, and telcos have the distributed infrastructure, in base stations, edge nodes and geographic presence, to deliver on those requirements.
The company is positioning itself to meet a very real market need: the shift from centralized AI training to massively distributed inference. Inference will account for 80–90% of total AI usage according to an MIT study, and will place far greater stress on latency, routing efficiency and continuous, small‑packet traffic than training ever did. Networking, rather than compute, becomes a constraint on performance.
"The bulk of attention in AI networking over the past two years has been focused on training — scale-up fabrics, lossless scale-out Ethernet, GPU-to-GPU interconnects," AvidThink principal Roy Chua told Fierce Network. "That’s understandable; that’s where the money has been. But inferencing is rapidly becoming the dominant AI workload, and the networking requirements for inference are different from training, so there's a market need for networking fabrics to optimize inferencing workloads."
Arrcus doesn't have a direct competitor, Chua said. "Arrcus provides an end-to-end fabric that has rich policies, flexibility,and quality-of-service controls (based on standards like SRv6), and that supports a wide range of platforms and silicon," he said.
"Arrcus is betting that incumbents like Cisco, Arista, HPE/Juniper, and F5 will be slow to pivot their existing product lines toward a policy-aware inference fabric architecture. These incumbents are motivated by their existing book of business. But these are well-resourced companies, and could mount a stronger defense than expected," Chua added.