The Great Unlocking: An Open Network Alliance Rises to Power AMD's AI Push

SAN JOSE, CA – July 23, 2026 – In the high-stakes race to build the world's AI infrastructure, the network has become the critical battleground. It's a contest not just of speed, but of philosophy. A significant new alliance, showcased this week at AMD's Advancing AI Day, signals a powerful push for a more open, disaggregated future. Arrcus, a leader in distributed networking, and UfiSpace, a specialist in open networking hardware, have announced a deeply integrated solution designed to serve as the high-performance backbone for AMD-powered AI clusters. This isn't just another product launch; it's a strategic move to uncouple AI networking from the proprietary, single-vendor stacks that have long dominated the field.

The collaboration brings Arrcus's ArcOS network operating system to UfiSpace’s open switches, which are built on Broadcom's formidable Tomahawk silicon. This pre-validated networking fabric is the foundation for GIGABYTE's GIGAPOD, a modular AI cluster built around AMD's EPYC CPUs and Instinct GPUs. By uniting best-of-breed software, hardware, and compute, the partnership presents a compelling, production-ready alternative in a market hungry for choice and resilience.

“By integrating ArcOS deeply with UfiSpace's Tomahawk 6 and Tomahawk 5 Broadcom platforms, we give customers building AMD-powered AI clusters an open, high-performance Ethernet fabric they can deploy with confidence,” stated Shekar Ayyar, Chairman and CEO of Arrcus.

A Challenge to Proprietary Dominance

For years, the AI networking landscape has been dominated by NVIDIA's InfiniBand, a high-performance, low-latency technology perfectly suited for the massive data synchronization required in training large models. However, its proprietary nature creates vendor lock-in, limiting customer choice and potentially inflating costs. The industry is now witnessing a powerful counter-current: the rise of high-performance Ethernet, championed by the Ultra Ethernet Consortium (UEC) and its influential members.

This Arrcus-UfiSpace solution lands squarely in this movement. It offers a disaggregated model, allowing customers to choose their hardware and software components separately, fostering a more competitive and innovative ecosystem. This is the essence of building for permanence in a volatile tech landscape—creating systems that are not dependent on the whims or pricing power of a single entity. The use of open, OCP-compliant hardware from UfiSpace, powered by industry-standard Broadcom silicon, is a direct rebuke to monolithic design.

“From the latest Tomahawk® 6 platforms to direct liquid-cooled innovations, UfiSpace’s solutions are engineered to scale with AI,” said Vincent Ho, CEO of UfiSpace. “Coupling these platforms with a hardened, L10-validated ArcOS allows customers to accelerate their timeline from evaluation to full production.”

The strategic importance of this open approach is underscored by Broadcom's support. “We're pleased to see Arrcus and UfiSpace bring this silicon to market in an open, co-validated solution that gives AI infrastructure customers a high-performance Ethernet fabric,” noted Hasan Siraj, vice president of product management for Broadcom's Core Switching Group.

Bolstering AMD's End-to-End AI Vision

This partnership is arguably as significant for AMD as it is for Arrcus and UfiSpace. As AMD continues to build momentum with its Instinct GPUs and EPYC CPUs, it faces the challenge of competing against NVIDIA's deeply integrated, full-stack solution. A powerful compute engine is only as effective as the network that feeds it, and a networking bottleneck can leave expensive GPUs sitting idle, wasting both time and money.

By providing a pre-validated, high-performance networking layer specifically tuned for its platforms, AMD can now offer customers a more complete and coherent end-to-end AI infrastructure solution. This move strengthens AMD's ecosystem, making it a more attractive and viable alternative for enterprises and cloud providers looking to build large-scale AI factories. The GIGAPOD solution from Giga Computing, which packages AMD's compute with this new networking fabric, is a tangible example of this ecosystem strategy in action.

“Deploying AI infrastructure at scale demands incredible engineering agility,” said Daniel Hou, General Manager of Giga Computing. “With our GIGAPOD solutions, we have seamlessly paired AMD's latest platform with Arrcus and UfiSpace's open, high-bandwidth switching solutions. This pre-validated solution removes complex integration barriers, allowing our customers to deploy rapidly and lead the generative and agentic AI revolution.”

The Technical Fabric of Resilience

At the core of this announcement is a set of advanced technologies engineered to handle the unique demands of AI workloads. This is where performance meets permanence. The solution's resilience is built upon a foundation of cutting-edge hardware and intelligent software working in concert.

UfiSpace provides the physical muscle, with platforms built on Broadcom's latest AI-optimized silicon. This includes switches like the S9331-64HO, based on the Tomahawk 6 chip, which offers a staggering 102.4 Tbps of switching capacity. The availability of direct liquid-cooled and OCP ORv3-ready platforms demonstrates a forward-looking design that addresses the immense power and thermal challenges of modern AI data centers.

Arrcus's ArcOS provides the intelligence. It brings a suite of critical features that transform standard Ethernet into a high-performance AI fabric. Its implementation of lossless Ethernet using RoCEv2 (RDMA over Converged Ethernet) is key, enabling the kind of direct, low-latency communication between GPUs that was once the exclusive domain of InfiniBand. Furthermore, sophisticated congestion management algorithms prevent network traffic jams, while advanced load-balancing ensures workloads are spread evenly, maximizing the utilization of every link. Finally, deep-telemetry capabilities give operators the visibility needed to monitor, troubleshoot, and fine-tune the network for peak performance, ensuring the entire cluster operates with predictable efficiency.