Digital twins: Solving the capacity crunch for the ‘AI factories’

Executive Summary

• Digital twins are becoming decision systems for environments that are too dense, dynamic, or too interconnected for static models and siloed software – such as AI data centres and electrical grids
• Using AI-based tools, data centres can be mapped and specified virtually, via digital twins, with digital infrastructure deployed as units of compute, says Vertiv – an important approach to achieve future gigawatt-scale buildouts
• Schneider Electric and ETAP have led the way in building the world’s first digital twin to simulate AI factory power requirements from grid to chip level

The data centre industry is responding to the myriad pressures AI workloads are placing upon it. As Nick Marlow, head of project management at BCS Consultancy wrote for this publication in March, JLL forecasts that AI will represent half of all data centre workloads by 2030, while BCS’s Data Centre Truths report found more than three quarters of those polled had seen a notable uplift in demand linked directly to AI over the past 12 months. The future data centres are, increasingly, becoming AI factories.

When the pressures of AI and high-performance compute (HPC) hit, the capacity ramifications are evident. Power density, cooling and rack layout are all enhanced to affect change. Yet the bottleneck remains; as Verdantix reports, in London alone, around 400GW of grid connection requests are waiting in the queue. There is one way of combating the issue – and going back to the drawing board.

Digital twins, once solely software for simulations or future transformation initiatives, today have a more practical purpose. “In 2026, the more interesting story is operational,” Sai Teja, writing for DataM Intelligence, noted earlier this month. “Digital twins are becoming decision systems for environments that are too dense, too dynamic, or too interconnected to manage with static models and siloed software.”

AI data centres and electrical grids, alongside brownfield factories and oil and gas production networks among others, are feeling that change. Vertiv’s Vertiv Frontiers report at the start of this year helps outline the rationale. “With increasingly dense AI workloads and more powerful GPUs also come a demand to deploy these complex AI factories with speed,” Vertiv noted.

“Using AI-based tools, data centres can be mapped and specified virtually, via digital twins, and the IT and critical digital infrastructure can be integrated, often as prefabricated modular designs, and deployed as units of compute, reducing time-to-token by up to 50%. This approach will be important to efficiently achieving the gigawatt-scale buildouts required for future AI advancements.”

One of the most interesting developments in this field came in March last year, when Schneider Electric, alongside ETAP, unveiled the world’s first digital twin to simulate AI factory power requirements from grid to chip level. The offering aimed to go beyond basic visualisation of electrical systems. ETAP’s modelling technology created a virtual replica of a data centre’s electrical infrastructure and combine it with real-time power system data, advanced analytics, and insights.

Writing for CIO.com in March, Timothy W. Coleman, senior director at BasisPath, advocates for a semantic digital twin to solve the puzzle and be the missing layer ‘so decisions become computable, verifiable and governable, rather than negotiated and guessed.’

In a coupled system, Coleman notes, you could theoretically have free rack units, but have nowhere to put a workload because they sit behind the wrong power path or under the wrong redundancy state, for instance, leading to stranded capacity. The semantic twin forces these disagreements into explicit, resolvable definitions.

“As AI continues to drive densification and energy becomes the growth limiter, the advantage won’t necessarily come from procuring GPUs or negotiating better colocation terms,” Coleman added. “It will depend on whether the enterprise can defined the data centre in a machine-readable way, connect it to workloads and business commitments and govern it reliably rather than on gut feel.”

Picture credit: Magnific.com