How is Vertiv Building AI-Ready Edge Infrastructure?

The acceleration of AI adoption is forcing a fundamental rethink of data centre infrastructure design, according to Andrea Ferro, VP Power and IT Systems EMEA at Vertiv. 

The company is responding to projections showing AI-ready data centre capacity rising at 33% annually between 2023 and 2030, according to McKinsey, while Goldman Sachs forecasts AI could drive a 165% increase in data centre power demand by the same date.

"It's the speed of AI adoption scaling and its impact on resource use," Andrea says when asked about the pressures AI is placing on infrastructure. 

“Today’s high-end racks already consume up to 120-132kW or more of power, but next-generation systems launching in 2027 and later, are being estimated to be up to 600kW per rack – and potentially beyond 1MW for future generations.

“This isn’t just about scaling existing infrastructure – it's about rethinking power delivery, thermal management and system integration," explains Andrea. “We’re no longer dealing with isolated server deployments but with fully integrated AI factories that require increasing amounts of power to enable computing at scale. The challenge is compounded by the evolution of Perception AI through Generative AI to Agentic AI – each generation requiring more sophisticated infrastructure integration.”

Vertiv identifies training to inference shift as a critical transformation

The movement of computational workloads from centralised training to distributed inference is reshaping infrastructure requirements across the sector. Andrea identifies 2025 as the year when computation shifts dramatically toward inference at the edge, moving away from the large-scale training clusters that have dominated recent infrastructure planning.

“From a technical perspective, training workloads can tolerate higher latency and benefit from centralised, high-throughput architectures. Inference, particularly for agentic AI applications, demands microsecond response times with consistent performance,” Andrea says. 

Applications now process hundreds of thousands of tokens in microseconds, from autonomous decision-making systems to robotics applications requiring real-time environmental analysis.

This shift drives demand for distributed infrastructure with different characteristics: lower latency tolerance, higher reliability requirements and the need to operate in diverse environmental conditions. Andrea describes it as a move from designing for consistent, predictable performance to increasingly demanding GPU power profiles.

“It’s managing the integration of multiple AI models operating in coordination, each with different infrastructure requirements,” says Andrea. 

Edge AI market growth drives data centre power and cooling requirements

The edge AI market is projected to grow from US$20.78bn in 2024 to US$66.47bn by 2030, reflecting what Andrea describes as a fundamental shift in where computation takes place.

“Enterprises need local processing for speed, control and efficiency. Edge data centres help with that,” insists Andrea. “They also support regional resilience and reduce the load on centralised systems. Perhaps more beneficial for the enterprise itself is the idea of data sovereignty and the potential benefits towards security.

“But there’s another factor: energy efficiency. Processing data at the edge reduces the energy costs of data transmission and allows for more granular power management. With data centre energy consumption set to double by decade's end, this efficiency gain is becoming crucial.”

Physical AI represents a shift in how edge computing is conceived. Traditional edge applications might process sensor data and send results to actuators, but physical AI systems need to perceive their environment, reason about it and take physical actions in real-time. 

“This creates unique infrastructure challenges,” Andrea says. “These systems often need to operate in harsh environments such as manufacturing floors, outdoor installations and mobile platforms, while maintaining the computational power of a data centre.”

The workload distribution is not about abandoning cloud infrastructure, rather about intelligent workload distribution. “AI needs a hybrid approach that leverages both centralised and distributed capacity,” says Andreas. “Training large models still happens in cloud clusters, but inference is increasingly happening at the edge. What we’re seeing is cloud rebalancing: organisations are strategically placing workloads based on performance requirements, data sovereignty regulations and cost optimisation.

Some training workloads are moving closer to data sources, while inference workloads are distributed to minimise latency. “The key is matching the right workload to the right infrastructure,” he adds. “Batch processing and model training can tolerate centralised processing, but real-time inference, especially for physical AI applications, requires edge deployment.”

Vertiv CoolChip CDU 100 addresses power density challenge

The infrastructure challenges facing teams are multifaceted. Power density remains the primary concern, with the move from traditional 10-20kW racks to systems consuming up to 120-132kW or more today, and next-generation architectures targeting 600kW per rack by 2027.

“This creates a cascade of technical challenges. Power delivery requires 33kW DC power shelves and 1400A busbars – infrastructure that was not required or imagined for pre-AI traditional data centres,” Andrea says. 

“Thermal management demands advanced liquid cooling systems like our Vertiv CoolChip CDU 100, capable of handling heat loads that would overwhelm conventional air cooling.”

But, Andrea notes, there is an additional complexity to consider: network sprawl.

“Infrastructure teams now manage distributed AI workloads across potentially hundreds of edge locations, each requiring remote monitoring and management capabilities,” he says. “Unlike centralised data centres with on-site technical teams, edge deployments often operate with minimal local support.

“The integration challenge is equally significant. Modern AI infrastructure requires seamless coordination between compute racks, power distribution, cooling systems and monitoring. A failure in any single component can cascade through the entire system, making holistic design critical.”

Andrea provides the example of a manufacturing facility deploying AI-powered robotics for quality control, where the system needs to process high-resolution visual data in real-time, make quality decisions in milliseconds and coordinate with production line systems.

“The technical implementation might include GPU clusters operating at 100+ kW per rack, requiring liquid cooling systems with precise temperature control,” he explains. “The power infrastructure needs 33kW DC power shelves integrated with 1400A busbar systems to handle the load distribution.

“But here’s where integration becomes critical: if the cabling infrastructure isn't designed for high-density deployment, it can obstruct cooling airflow, creating thermal hotspots that degrade performance. If the liquid cooling system doesn't integrate properly with the rack architecture, you end up with inefficient heat transfer and potential reliability issues.

Vertiv 360AI reference architectures target retrofit market

Organisations are pursuing both retrofit and new-build approaches, but the technical decision matrix has become more sophisticated. 

Retrofitting remains attractive for reducing disruption and leveraging existing investments, but legacy power distribution systems were not designed for current loads, let alone the 600 kW per rack targeted by 2027.

“Upgrading from traditional power delivery to 33kW DC power shelves and 1400A busbar systems often requires electrical infrastructure changes,” Andrea says. “Cooling presents similar challenges. Retrofitting liquid cooling systems into facilities designed for air cooling requires significant modifications to both the mechanical infrastructure and the control systems.”

For edge deployments, Vertiv is seeing success with modular approaches. 

“Our Vertiv 360AI reference architectures, ranging from 88kW to 115kW configurations, can be deployed as integrated systems that work within existing facility constraints while providing the performance density needed for AI workloads. This approach reduces deployment time by up to 50% compared to custom integrations,” says Andrea.

Sustainability requirements are evolving alongside AI demands. AI workloads increase power and resource consumption, but also create opportunities for energy efficiency optimisation. 

Regulators in many regions are monitoring developments, and there is increasing pressure from enterprise customers to demonstrate measurable energy usage and resource improvements.

“This goes beyond traditional metrics like Power Usage Effectiveness (PUE) to include circular economy practices like heat reuse, water recycling and end-of-life equipment management,” Andrea says. 

The design principle that Andrea emphasises in light of this evolution is adaptability. “The main principle is designing for adaptability and scalability, not just current requirements. AI is evolving rapidly, and infrastructure needs to accommodate today's applications as well as tomorrow's breakthroughs,” he says.

According to Andrea, standardisation and early integration remain key factors in deployment success. 

“Using reference architectures where possible, engaging partners early in projects, and planning for modular growth reduces deployment risks and speeds time-to-value,” he says.

"Build with change in mind as it is the only constant. AI isn't a fixed target – it's a rapidly evolving field that will continue to surprise us with new requirements and capabilities,” concludes Andrea. "The decisions we make today about edge infrastructure will enable or constrain the next decade of AI innovation."