How is the acceleration of AI development shifting into a new phase? And what effect is that having on our demand for data centre infrastructure?
We’re seeing a move from experimentation to deployment at scale. AI is no longer something that sits in a lab or a discrete cluster. It’s being integrated into core business systems and running continuously, which changes what infrastructure is expected to deliver.
The key shift is intensity. Workloads are denser, more power-hungry and less predictable. This means data centres can’t rely on older assumptions around capacity, load distribution or response time. They need to be designed for higher variability, as well as for higher volume.
It feels like data centres need to deliver more power, cooling, space – everything – faster than expected using infrastructure that is either unprepared or hasn’t been built yet. How does the industry contend with these challenges?
It starts with mindset. You can’t meet today’s pace with yesterday’s approach. Operators are moving towards prefabricated modular infrastructure, shorter design-to-deploy timelines, and more integrated delivery models. Prefabrication helps and can reduce deployment time by up to 50%. So does standardising the way cooling, power and racks are designed, manufactured and assembled in a standardised factory environment, simultaneously, rather than in sequence.
Another strategy that is key to being prepared for what’s next is collaboration across the industry. For example, our strategic partnership with NVIDIA. Vertiv has worked with NVIDIA on the end-to-end power and cooling reference design for both the NVIDIA GB200 NVL72 and the GB300 NVL72 platforms. By staying one GPU generation ahead, our customers can plan for future infrastructure before the silicon lands, with deployment-ready designs that anticipate increased rack power densities and repeatable templates for AI factories at scale.
How do we deal with the discrepancy in development cycle speeds between AI and the infrastructure used to house it?
This is one of the biggest structural mismatches the industry faces. AI development is sprinting. Infrastructure is still built on marathon timelines. Speed is critical and densities are different. Therefore, a change of philosophy is needed when it comes to data centre design and build.
The new AI factories need to be ready much faster than we’ve ever seen before in the industry. By standardising everything including cooling and power distribution, critical infrastructure can be deployed at speed rather than needing to retrofit what already exists or build from scratch, which can reduce timelines significantly.
On the energy side of things, do you expect data centres to take on a new role in relation to the grid, especially as some economies work to further electrify in pursuit of net zero goals?
Yes. The old model – draw power and provide backup – is shifting. It’s no secret that data centres are prioritising energy availability challenges. Overextended grids and increasing power demands are changing how data centres consume power. Many large facilities now operate as part of the wider energy system, helping manage peak demand or stabilise frequency through intelligent battery usage or flexible loads.
Data centre operators are seeking energy solutions that enable them to minimise generator starts and reduce energy costs and reliance on the grid. Microgrids integrated with uninterruptable power supply (UPS) offer a promising solution, enabling power reliability, stabilising renewable fluctuations, and protecting critical loads. They can also provide ancillary services to the main grid, such as frequency regulation and enhance grid stability by participating in demand response and load shedding.
This is being driven partly by policy and partly by economics. As electricity becomes a more valuable and volatile resource, infrastructure that can respond dynamically will be better placed to operate cost-effectively – and in some regions, to operate at all.
On the component side of things, how is the new generation of GPUs and other internal server equipment geared towards AI changing the way data centres need to be built?
Newer GPUs and high-bandwidth interconnects are driving heat and power requirements far beyond traditional design envelopes. A rack that previously ran at 10kW might now need 50kW to 100kW or more, and forecasts indicate this may increase to 300-600kW and possible 1MW by 2030 – this changes the physical reality of the room. This means that densification is required – it’s about making sure that there is more compute in as little footprint as possible.
The newer GPUs generate far more heat, so cooling systems need to become more targeted. Airflow alone is rarely sufficient, making direct liquid cooling, cold plates or hybrid systems necessary. Cable management, power infrastructure and weight loading also shift. Even the spacing between cabinets can affect thermal performance. This could involve a redesign from the inside out or layering new kit into old frameworks.
Can you talk about Vertiv’s work with Intel and NVIDIA on cooling systems? What’s the benefit of a dual system over a pure liquid-cooled facility, for example?
Vertiv has co-developed reference architectures with both Intel and NVIDIA to address next-generation AI workload demands. For NVIDIA’s GB200 NVL72, Vertiv released a 7 MW reference architecture supporting rack densities up to 132 kW. This includes a hybrid system that combines liquid cooling for prime heat sources with air cooling for supporting infrastructure.
For Intel’s Gaudi3 platform, Vertiv validated designs capable of handling 160 kW using pumped two-phase (P2P) liquid cooling, alongside traditional air-cooled setups up to 40 kW.
Hybrid cooling systems are based on a clear set of technical and operational frameworks:
Component-level thermal targeting
Liquid cooling – direct-to-chip cold plates or rear-door exchangers – focuses precisely on AI accelerators. This means airflow systems only need to support peripheral equipment, improving overall energy use and avoiding over-engineering the facility.
Phased deployment and flexibility
Hybrid architectures allow gradual ramping up of liquid cooling infrastructure.
For smooth upgrades, it’s important to design systems that can accommodate higher liquid temperatures from the start. Operators can begin with air cooling, introduce liquid in hot zones, and expand as capacity needs grow.
Operational compatibility
These designs support mixed workloads – GPU clusters, CPUs, storage – in the same white space by delivering the cooling each requires without impacting others.
End-to-end deployment frameworks
Vertiv’s reference architectures include detailed layouts: fluid routing, rack spacing, containment strategies, plus commissioning protocols. The NVIDIA frameworks are factory-tested and SimReady via digital twins, significantly reducing onsite uncertainty.
These hybrid frameworks offer precise thermal control, deployment agility, resilience, and simplified operations. Essentially, they merge the benefits of both air and liquid cooling into a scalable and AI-ready model.
How does AI change the ways in which data centres are likely to require maintenance or even fail? What kind of adjustment will this require on the part of the industry?
The criticality definitely increases. AI systems tend to concentrate compute in fewer, more critical pieces of hardware, so if one component overheats or fails, the impact can cascade faster, disrupting the computational workload it supports. Thermal margin is tighter, fluid networks introduce new points of failure, and real-time monitoring becomes more important, not just for performance but for reliability.
This means more condition-based maintenance, more granular telemetry, and stronger alignment between IT and facilities teams. It also requires a different mindset – from reacting to faults, to proactively managing infrastructure health in real time.
- Data & AI
- Infrastructure & Cloud