Data Centre World 2026 Day 2: Scaling the AI Factory Model
Written by Rebecca Uffindell Fri 13 Mar 2026
Day two of the Data Centre World Keynote Theatre focused on what it means to scale AI infrastructure in operational terms. Where earlier discussions centred on discipline and growth, the emphasis shifted toward execution: power delivery, silicon readiness, commissioning gaps and lifecycle strategy.
Across sessions examining sovereign AI strategy, Rubin-class chip preparedness, grid stability and IT refresh cycles, a consistent pattern emerged. AI factories represent more than higher-density data centres. They introduce new economic, engineering and operational considerations that require coordinated execution.
Europe’s AI Factory Ambition
The morning session, Shaping Europe’s Digital Future and Exploring AI Factories, was chaired by Bertrand Delatte of Siemens.
One speaker framed the shift succinctly:
“It’s not any more a software revolution. It’s turning into an infrastructure race.”
Contributors from NVIDIA, hscale and nVent described AI factories as layered ecosystems rather than singular facilities.
“The AI Factory is really about a five-layer cake.”
Energy, silicon, physical infrastructure, models and applications must align simultaneously. Coordination across these layers determines deployment speed.
For European developers, regulatory and grid timelines were identified as structural bottlenecks.
“Megawatts on paper are different than megawatts delivered.”
Permitting cycles and grid upgrades often extend beyond hardware release timelines. As one speaker noted:
“Regions that adopt the industry and support faster regulation… all workloads will go to that region.”
Sovereign ambition, therefore, depends less on rhetoric and more on delivery capacity.
From Watts to Tokens
A notable shift in framing surfaced during the discussion. Success is increasingly measured in computational output.
“An AI Factory produces tokens.”
Infrastructure decisions – cooling strategy, power distribution, architecture -now influence token output and revenue per watt. Liquid cooling, modular reference designs and standardised architectures were positioned as accelerators of deployment and risk reduction.
“We should avoid bespoke projects, work more from reference architectures.”
Standardisation, rather than customisation, may determine how quickly Europe can scale AI capacity.
Rubin-Class Systems and Refresh Velocity
Lee Prescott of RED Engineering addressed the implications of next-generation silicon in his session, Are You Ready for Rubin?
Rack densities are moving from 50kW to 250kW, with Rubin-class systems pushing beyond previous assumptions. However, density alone is not the defining challenge; refresh velocity is.
“Power is revenue.”
If tokens are the product, then efficiency in power delivery directly influences economic return. Optimisation across cooling, DC adoption, stranded capacity reduction and infrastructure flexibility becomes financially material.
Prescott cautioned against static design assumptions in a moving silicon roadmap.
“Don’t play to the lowest common denominator.”
Flexibility in power strings, shared cooling loops and adaptive layouts were presented as mechanisms to reduce stranded capacity. Heat reuse also emerged as a strategic consideration.
“Chucking it to the atmosphere is not responsible.”
AI factories increasingly resemble industrial energy systems rather than conventional data centres.
AI Load Profiles and Grid Stability
Muhammad Naveed Saeed of Uptime Institute examined the volatility of AI power demand. Unlike traditional workloads, AI training clusters operate synchronously, creating coordinated load spikes that challenge utility thresholds and generator stability.
“We have to stabilise these. And we have to filter these power profiles before they reach the power generation equipment.”
Sub-synchronous resonance, harmonic distortion, and voltage flicker were identified as potential upstream risks.
Mitigation strategies discussed included software smoothing, GPU ramp control and energy storage buffers.
“It requires high capex.”
Energy storage introduces both capital and spatial implications, yet commissioning processes must evolve regardless. Linear load testing does not replicate AI training behaviour.
“Most of the facilities that are being commissioned today are being commissioned on the linear loads, which are not expected in these facilities.”
This gap highlights the need for commissioning models aligned with real AI load profiles.
IT Lifecycle and Structural Integration
The final session addressed IT refresh cycles and lifecycle management. Alastair Winner (Techbuyer) and Mark Acton questioned whether each new GPU generation necessitates wholesale infrastructure replacement.
“It feels like we’re sort of still having the same discussion from 20 years ago.”
The issue extends beyond hardware churn to organisational structure.
“The silos have got worse.”
As liquid cooling and direct-to-chip systems drive tighter integration between IT and facilities, separation between functions becomes increasingly difficult to sustain.
The discussion challenged assumptions around obsolescence. Equipment deemed unsuitable for AI training may still retain value in other contexts. Lifecycle intelligence, reuse strategies and circular supply chains were positioned as stabilising factors within volatile markets.
A Structural Inflection Point
Across the day, recurring pressures became evident: silicon development cycles are accelerating, infrastructure build timelines remain extended, power access and grid stability are tightening, and IT refresh intervals are compressing.
The critical variable is coordination. AI factory deployment depends on alignment across utilities, regulators, infrastructure providers and IT supply chains. Ambition alone is insufficient without execution capacity.
The AI factory model is no longer conceptual. Its sustainability now rests on governance discipline, grid readiness, commissioning adaptation and integrated lifecycle strategy.
As deployment accelerates, the differentiator will not be density alone, but the ability to engineer systems that remain resilient under evolving technical and economic constraints.
Written by Rebecca Uffindell Fri 13 Mar 2026