Grid to Chip: The Holistic Approach to Supporting AI in the Data Centre

Stand First: New and innovative approaches are necessary to allow data centre owners and operators to confidently deal with new AI requirements, to reduce risks and ensure effective implementation, while keeping sustainability commitments at the forefront. 

As artificial intelligence (AI) has emerged and multiplied as a workload, what has become clear about this family of technologies is that its infrastructure needs are unlike any previous generation of compute intensive workloads. 

Adoption and requirements

AI adoption has grown at an unprecedented rate. Current estimates are that the Agentic AI Market value may hit $199 billion by 2034, with forecasts that its economic contribution could reach as high as $4.4 trillion annually by 2030. 

However, all this comes at a cost, as there is growing concern that the scale of AI build out will impact energy availability and cost in many regions.  

The impact of this in the data centre is the need for new racks for heavier equipment, improved connectivity for data access and output, and higher energy density per rack with new direct current (DC) systems and liquid cooling throughout.  

While industry averages in 2024 saw around 12kW per rack, up to 100kW is not uncommon but roadmaps show dedicated AI processing systems that will reach 600kW per rack in 2027. 

It is clear that accommodating these kinds of developments requires a whole new approach. 

The Schneider Electric approach is a holistic view encompassing ‘grid to chip’ in an end-to-end strategy for powering AI infrastructure—from the utility grid all the way to the silicon chips inside data centres. 

This approach addresses all the key differentiators for AI workloads, starting with energy density and previously unseen kW per rack power draws, but also the kind of dynamic load profiles typical of AI too. An AI Large Language Model (LLM) in training mode can go from idle to maximum load almost instantaneously and repeatedly, unlike most other data centre workloads, which brings challenges for power trains.

Innovation demand

At every step of the grid to chip strategy, innovation is required. For example, Uninterruptable Power supplies (UPS) need to be able to handle the dynamic power demands of AI workloads without degrading battery life and reducing the impact on the grid at the same time. This increases reliability and supports peak loads for AI in various forms, ensuring optimal performance for data centres running GPU-accelerated AI training and inferencing tasks.

This dynamic load environment requires new and sophisticated management and orchestration systems too. 

Digital design and modelling

The AI Factory Digital Twin, co-developed with ETAP and powered by NVIDIA Omniverse is a virtual replica that simulates power needs and infrastructure performance in real time. It provides advanced electrical system design that accommodates dynamic loading and workload variability. It also provides scenario analysis for ‘what if’ situations, allowing owners and operators to see around corners and test scenarios before they arise. 

The combined power of these capabilities means that the digital twin system provides a sound basis for predictive maintenance, further reducing risks for outages or capacity restraints, and increasing overall resilience. With this deep analysis, the granular information combined with high levels of automation mean that energy efficiency optimisation becomes an ongoing process, which also supports sustainability efforts, while lowering the total cost of ownership. 

Sustainability requirements

Another key consideration is that the whole AI value proposition must be implemented sustainably. 

The grid to chip approach means that there is a high level of instrumentation at every stage, giving granular data on performance and load profiles, allowing for greater control and operational insights. The digital twin’s insights become actions for the real world.

AI-assisted systems such as Data Centre Infrastructure Management (DCIM) systems, can then move from ensuring efficiency to being able to meet sustainability obligations and ambitions, contributing to Power Usage Effectiveness (PUE) and Water Usage Effectiveness (WUE) ratings. 

Another sustainability consideration is a reliance on fossil fuel generated energy. Where energy grids are constrained, and to promote more renewable energy sources (RES), many organisations are deploying o-site renewable energy backed up by storage. This is well illustrated by the example of Adani in India, which is investing $100 billion in RES-powered AI data centres. 

Through lessons learned on a more than two decade path of sustainability, Schneider Electric has become a recognised leader. Sharing that journey with partners and customers, we have developed strong sustainability support resources and consulting services to allow us to work with data centre owners and operators in achieving sustainable operations, even in an area as demanding as AI.

AI and the future

AI Workloads are likely to increase in the coming years, as more organisations discover its utility for their own purposes. While many large companies have tended towards rationalisation of AI development efforts from often hundreds of projects down to single figures, the use cases are likely to be organization wide and so still intensive in terms of infrastructure and resources. 

The grid to chip approach therefore will evolve along the same lines. Through a process of constant optimization through the application of AI, data centres that are operated on this approach will be enabled to adaptively support evolving AI workload requirements, with efficiency and sustainability at the core.

As DeepSeek, and other examples have shown, AI development does not always rely on the cutting edge of the latest GPU-based and accelerated computing. AI models are evolving constantly and the use of blends of hardware such as previous generations of GPUs and other processors, shows that underlying infrastructure must remain flexible and adaptable. 

Holistic view

The AI revolution has demanded new solutions as previously unseen requirements are made of digital infrastructure. A holistic, end to end approach is necessary to ensure that every step of the AI adoption path can be supported across the board. Any collection of point solutions risks spikes or gaps that would be exposed by the sheer intensity of demand of this wave of AI. 

The grid to chip approach has proven the right one, supported by a broad range of partnerships, to provide data centre owners and operators with a low-risk path to future ready AI-optimized digital infrastructure that is both efficient and sustainable.