Amidst the bustling halls of International Telecoms Week in Washington, a pivotal conversation unfolded regarding the future of AI infrastructure. Senior representatives from the data centre, fibre, and wholesale connectivity sectors gathered to discuss changes driven by the surging demands of artificial intelligence. As the panelists spoke, a consensus emerged: the industry is entering a phase where practical execution takes precedence over mere ambition.
The Shift from Training to Inference
Scott Willis, CEO of DartPoints, shared insights on what 2026 may hold for AI workloads, predicting a significant shift as inference workloads become commercially viable at scale. Unlike traditional cloud tasks, AI inference is less sensitive to latency issues, allowing developers to explore locations beyond the congested tier-one data centre hubs. This shift occurs as traditional markets face power shortages, pushing demand toward regional areas with adequate energy supply.
Willis noted that many primary markets are experiencing a shortfall of half a gigawatt or more. As a result, demand is moving to sites capable of offering between 10 to 50 megawatts of power, essential for accommodating the upcoming AI workloads.
Bandwidth Demand and Network Infrastructure
Joda Schaumberg, SVP of Digital Infrastructure at Zayo, addressed the fibre aspect of this transformation, highlighting the company's efforts to densify its long-haul network. Following the acquisition of Crown Castle assets, Zayo is ramping up its metro network density to meet the expected rise in bandwidth demand driven by interconnected AI agents. Schaumberg acknowledged the uncertainty regarding the full impact of these workloads on bandwidth but projected that requirements could increase by as much as seven times, necessitating proactive development of conduit capacity.
A New Ecosystem Approach
Milad Abdelmessih, VP at KDDI Telehouse, articulated a vision where connectivity evolves from being a mere feature in data centres to a foundational requirement. The company’s initiatives include fostering shared campus environments for hyperscaler and enterprise tenants, as seen in recent developments in Toronto. This shift highlights the need for operators to act as neutral aggregators of ecosystems rather than simply providing powered space.
Optimizing Power Efficiency
From an efficiency standpoint, Nilesh Shah, an AI infrastructure advisor at FarmGPU, shed light on the challenges of power utilization in current AI data centres. He revealed that GPU clusters often operate at only 30% to 40% efficiency. Shah emphasized the potential for software-driven optimization to dynamically balance power across training and inference tasks, suggesting that such measures could yield efficiency improvements of over 20% without requiring additional physical infrastructure. This perspective positions smarter infrastructure software as a critical element in addressing the industry’s power challenges.
Looking Ahead
As the AI sector continues to grow rapidly, discussions at ITW 2026 highlighted a vital shift in focus toward power availability and execution capabilities. The insights shared by industry leaders illustrate an interconnected future where data centre operators, fibre providers, and AI infrastructure advisors must collaborate to create a resilient and efficient ecosystem capable of meeting escalating demands. The lessons learned today will shape the trajectory of AI infrastructure in the years ahead.
The stories that move AI & crypto markets — before the market reacts.
Free. 7am ET. Five stories. 62,400 readers.
