Our Take
Enterprise AI success depends on unglamorous infrastructure work—power, cooling, network, and security—not on model capability or agentic claims.
Why it matters
Companies racing to deploy AI without addressing physical constraints, legacy system conflicts, and shadow AI risks will hit deployment walls faster than boards expect. Infrastructure is now the throttle.
Do this week
Infrastructure lead: audit your data centre power budget, cooling capacity, and network spine against your AI compute roadmap before Q1 planning closes, so you can identify constraint timelines now.
TechEx North America exposed the infrastructure reality behind AI adoption
Across four parallel tracks at TechEx North America (Edge Computing, IoT, Data Centre Congress, and Cyber Security), speakers and exhibitors focused on the operational and physical requirements that must be in place before AI reaches production.
The Data Centre Congress sessions revealed that AI economics are reshaping infrastructure needs faster than data centres can be built. Power, cooling, water, land, and permits are now the binding constraints. A recurring theme: AI strategy only becomes real when it hits the physical infrastructure stack, and that stack takes years to mature. Santa Clara, the host city, presented its own data centre capacity challenges as a case study.
The Edge Computing track, led by Ed Doran of the Edge AI Foundation, addressed how moving intelligence closer to machines changes risk profiles, latency, and dependence on cloud services. Participants included Akamai, Spectro Cloud, Scylos, and Schneider Electric. A key tension emerged: faster local decisions reduce cloud dependency but create observability and control gaps that decision-makers struggle to own.
The IoT and Digital Twins sessions flagged a broader pattern: projects work in demos but stall in deployment. Rockwell Automation and Ford presented sessions on scaling physical AI and connected asset intelligence, identifying what has been called "pilot purgatory"—deployments that work well in concept but fail when meeting real machines and legacy software. Digital twins received similar scrutiny; the sessions emphasized that operational models designed to improve maintenance and pre-test decisions matter more than visual replicas.
Cybersecurity sessions showed that shadow AI (staff using unapproved AI services in workflows) and data exfiltration are collapsing the boundary between data governance and cyber governance. Legacy systems issues echoed across both cybersecurity and IoT stages; older plant systems meeting modern intelligence create security afterthoughts in industries where that cannot happen (transport, energy).
Infrastructure constraints are now the real bottleneck, not model capability
The conference showed attendees a gap between how AI is discussed in boardrooms (agentic automation, productivity gains) and how it must be deployed in practice (networks, data centre capacity, zero-trust controls, legacy system integration). Companies that treat infrastructure as secondary to model selection will find deployment stalling at the physical layer.
Water and power constraints cut through rhetoric about AI scale. Data centre construction timelines, cooling capacity, and cybersecurity maturity now determine whether AI pilots move to production. Edge and IoT deployments depend on how carefully intelligence is applied to older machines. Shadow AI use inside enterprises, absent logging and approval, means security weaknesses don't shrink when speed becomes a priority—they expand the attack surface.
Audit your infrastructure stack before declaring AI readiness
Practitioners should map AI compute requirements against current data centre power budgets, cooling capacity, and network spine bandwidth. Identify where legacy systems will conflict with new deployments. Establish shadow AI logging and approval workflows before distributed inference pilots launch. For edge deployments, work backward from machine speed and latency requirements to define what observability and control look like at the edge, not after pilot launch.
The broader lesson from TechEx: companies that understand power grids, networks, security culture, and legacy system constraints are more likely to move AI from production demo to operational reality. Getting the bigger picture—infrastructure first, not last—is what separates projects from deployments.