en.Wedoany.com Reported - Clayton Wagar, Nokia's Head of AI and High-Performance Networking, said during a webinar hosted by BizClik that the real challenge in AI infrastructure lies in ensuring data movement is fast enough to keep pace. He believes AI infrastructure is entering a new phase where networking is becoming as critical as compute power.
For data center operators and telecom service providers, the opportunities from this shift feel familiar. Clayton explained that cloud providers took decades to master hyperscale operations, while AI now merges the demands of traditional hyperscale infrastructure with the performance requirements of supercomputing, prompting operators to rethink how to connect AI clusters distributed across buildings, campuses, and even different regions. As AI systems grow increasingly complex with agentic AI, queries rely on distributed storage, multiple services, and geographically separated infrastructure, making high-performance networking the foundation of user experience.
The five key takeaways from the webinar include: AI performance bottlenecks are shifting, with the key factors determining AI performance and scalability moving from compute alone to data movement; AI deployments are scaling toward millions of GPUs, creating unprecedented demands on land, power, and networking; physical limitations are accelerating the transition from electrical to optical connectivity, with optical networks delivering greater bandwidth over longer distances at higher efficiency; AI infrastructure is increasingly resembling service provider networks, making long-distance connectivity, resilient recovery, and granular management more essential; and traditional data centers must adapt to AI's rising compute density through new cooling technologies and redesigned layouts.
Clayton believes the biggest challenge may be preparing for infrastructure growth that far exceeds current expectations. AI clusters have expanded from tens of thousands of GPUs to hundreds of thousands, and some facilities are now being designed around one million or more GPUs. At this scale, networks begin to look more like telecom infrastructure than traditional enterprise networks. He said, "Every time we hit a limit in scale, we see there's another limit, and it's usually 10 or 100 times what we needed before." As enterprises connect AI campuses spanning hundreds or thousands of kilometers, long-distance connectivity, resilience, and visibility all become critical.
Clayton also noted that a technological transformation from electrical to optical interconnect is underway. Electrical technology is approaching its physical limits, while optical networks can deliver higher speeds over longer distances with lower power consumption, though this requires significant redesign of communications equipment. Legacy data centers present another challenge, as they were not designed for AI's power density or heat output, making liquid cooling and hybrid cooling methods especially important for operators looking to reuse existing sites. He believes that as AI infrastructure expands from individual facilities to interconnected AI campuses, automation and high-capacity optical networks will become essential, and the industry's accumulated networking expertise will provide a competitive advantage.
