Building Networks for Evolving AI Workloads AI workloads continue to evolve across training, inference, distributed inference and agentic AI with each workload placing a unique requirement on the infrastructure. Having a flexible infrastructure that can adapt to future workloads is critical to keeping pace with AI. As organizations build larger and more distributed AI systems, networking has become a critical determinant of overall system performance. The need for both performance and flexibility has helped make Ethernet the networking foundation of choice for modern AI data centers. Its broad ecosystem, open standards, and ability to continuously evolve have enabled widespread adoption, with Gartner forecasting that more than 80% of hyperscalers will choose Ethernet for AI workloads by 2028. However, large-scale AI training places fundamentally different demands on the network than traditional cloud and storage workloads. Keeping hundreds of thousands of GPUs tightly synchronized requires fast recovery from transient congestion and path failures, leveraging load balancing like packet spray and source routing, as well as other capabilities that traditional Ethernet transports such as RoCEv2 were not originally designed to provide. That’s why organizations across the AI ecosystem, including OpenAI, Microsoft, Meta, and AMD, are developing new Ethernet transports optimized for AI workloads. The fully programmable AI NIC architecture from AMD has helped accelerate the development, testing and validation of these transports, enabling innovation without waiting for new hardware generations. In just 16 months, three new major AI-ready Ethernet transports emerged, highlighting both the pace of AI networking innovation and the need for infrastructure that can evolve alongside it. Customer Choice at the Pace of AI The emergence of multiple AI-ready transports reflects an important reality: no single transport is likely to be the right answer for every AI deployment. Different workloads, architectures, and operational requirements can favor different approaches to congestion control, load balancing, and failure recovery. To evolve at the pace of AI, organizations need the flexibility to adopt the transport that best fits their environment rather than being limited to a single hard-coded networking approach. AMD AI NIC technology is the only fully programmable AI NIC architecture available today, giving customers the ability to implement new transport capabilities and optimizations on existing hardware without waiting for the next hardware generation. This approach also leverages existing investment to help accelerate future generations of interconnects to production. Customers can develop, validate, and deploy transports on AMD Pensando Pollara 400 AI NICs and carry that work forward to AMD Pensando Vulcano 800 AI NICs, preserving software investments and operational expertise across generations. Combined with the open ecosystem approach from AMD, organizations gain greater choice in how they build and evolve AI networks. Driving the Next Generation of Ethernet Transports The industry’s response to these challenges can already be seen across a growing set of AI-ready Ethernet transports. Developed by hyperscalers, AI leaders, and industry consortiums, these transports are helping improve congestion control, resiliency, efficiency, and scalability for large-scale AI deployments. Together, they demonstrate how an open ecosystem is accelerating innovation across AI networking. 2025: Ultra Ethernet Consortium (UEC) | Open Industry Standard The Ultra Ethernet Consortium brings together more than 100 industry members to develop an open, interoperable Ethernet architecture purpose-built for AI and HPC environments. AMD was a founding member of the consortium and authored Network-Signaled Congestion Control (NSCC), now included within the UEC congestion management specification. AMD also delivered one of the first AI NICs capable of supporting