The conversation at ITW 2026’s Digital Infrastructure Stage keynote panel on AI infrastructure was blunt in places. Panellists representing fibre operators, data centre providers and AI infrastructure advisors were broadly aligned on one point: the industry is moving into a new phase in which execution matters more than ambition, and where the geography of AI compute is being dictated not by latency requirements or enterprise clusters but by where power can be found.

Scott Willis, CEO of DartPoints, argued that 2026 is shaping up as the year inference workloads become commercially real at scale. The shift from training to inference, he said, changes the calculus around location. AI inference does not carry the same latency sensitivity as traditional cloud workloads, which means developers are free to follow power into secondary and regional markets rather than clustering in tier-one data centre hubs where capacity is constrained.

Willis noted that half a gigawatt or more of power is simply unavailable in many primary markets, pushing demand into the 10 to 50 megawatt range in locations where supply exists.

Joda Schaumberg, SVP of digital infrastructure at Zayo, addressed the fibre dimension directly. The operator has been densifying its long-haul network, putting, in his words, significant fibre miles in the ground, and the acquisition of Crown Castle assets gave Zayo a means to accelerate its metro network density without having to build from scratch.

He acknowledged that what Agent AI workloads will ultimately do to bandwidth demand remains uncertain, but cited projections that interconnected AI agents could drive bandwidth requirements up by a factor of seven, meaning operators must be building conduit capacity now for expansion that cannot yet be fully modelled.

Milad Abdelmessih, VP at KDDI Telehouse, framed the data centre operator’s position around the ecosystem model: the idea that connectivity is not a feature of a data centre but a precondition for it.

The company has been working to bring hyperscaler and enterprise tenants into shared campus environments, with recent activity in Toronto cited as an example of the direction of travel. His central point was that operators in the data centre and colocation space need to function as neutral aggregators of ecosystems rather than simply providers of powered space.

Nilesh Shah, AI infrastructure advisor at FarmGPU, brought a compute efficiency perspective to the power discussion. GPU clusters in current AI data centres are running at utilisation rates of 30$ to 40% in many cases, he argued, and software-driven optimisation, dynamically allocating power between training and inference workloads, managing GPU utilisation in real time, could unlock more than 20% efficiency gains without additional physical infrastructure. The implication is that the industry’s power problem is partly a utilisation problem, and that smarter infrastructure software is as important as raw capacity expansion.

The panel’s closing exchanges focused on what the next four months require from each business represented. The answer from multiple speakers converged on execution, the ability to close deals, deploy capital, and deliver on commitments in a market where demand is real, and competition is intensifying.

Willis was direct: the opportunity is there, but so is the risk of being outpaced. Schaumberg framed it as an obligation to show up as a flexible commercial partner, willing to take on different ownership structures and deal constructs rather than defaulting to standard arrangements.

The panel was chaired by Sam Evans, senior managing director at Teneo. Praveen Agarwal, chief sales officer at CONSTL, also participated.

RELATED STORIES

Lightpath boosts NYC fibre network across 2,400 cell tower sites

Sparkle CEO Enrico Bagnasco re-elected as Global Leaders’ Forum chair