Europe’s AI factories, data centre development and the vital role of dispatchable power

Introduction

The European High Performance Computing Joint Undertaking (EuroHPC JU) is a pan European strategy designed to help the EU ensure it remains at the forefront of AI by, amongst other things, developing 12 high performance supercomputer data centre sites.

Europe’s AI factory strategy is a response to ensure the region does not fall behind the US, China and India in the AI arms race. At first glance it has little to tell us about how future data centres are to be operated and powered.

However, how these AI supercomputers are designed, how the data centres and halls in which they reside are being built, how they are powered and how they operate (think renewables, microgrids, waste heat reuse, efficiency of operations through power capping and a raft of other factors) – have much to tell the broader data centre market.

This transformation in how data centres are powered is driving a broader shift toward dispatchable power — energy that can be controlled, scaled, and delivered on demand. Unlike traditional standby-only systems or reliance on the grid, dispatchable power enables both independence and resilience, supporting prime and backup needs alike.

This is in part because their construction will serve AI workloads that are similar in nature to those driving billions of dollars in investment across the commercial data centre world.

This includes ML/AI model building, Gen AI and reasoning.

This publicly funded strategy is the tip of an AI factory iceberg that is coming to Europe in the form of commercial 20MW+ AI supercomputer factories that will be built across the continent.

It is clear that many of these sites will not rely solely on grid connections but instead require dispatchable power — including sustainably fuelled gensets and independent microgrid energy solutions that can operate flexibly on demand.

Already AVK is partnering with commercial AI data centre developers across Europe’s FLAP-D markets and beyond to provide resilient, sustainable dispatchabe power (prime/microgrids and standby) required by new AI architectures.

In the new generation of AI data centres the IT hardware architectures being hosted are a mix of GPU and CPU infrastructure. At a processor level they are being designed with power capping to run with maximum energy efficiency at high rack power densities that require hybrid air and liquid cooling technologies. The power resilience and back up requirements of running GPU compute densities of hundreds of kilowatts per rack mean traditional standby-only infrastructure is no longer enough — this is where AVK’s dispatchable power solutions, including integrated microgrid strategies, offer a vital bridge between standby and prime.

So, what is an AI factory supercomputer?

The Eurohpc.ju site says: “AI Factories are dynamic ecosystems that will build around AI-optimised supercomputers, offering computing resources and support services to the European industry, as well as to the European scientific users for the development of large AI models to take advantage of AI technology capabilities in the European Union, and for the development of skills and knowledge in the domain of AI.”

Often located in research institutions, they are used by academics, businesses and public bodies who seek access and time slots for their workloads [much like astronomers seek time on radio telescopes or physicists gain access to national particle accelerators.]  As part of the EU AI factory strategy there are 12 locations across Europe where AI factories are currently under construction, being developed, installed or upgraded.

AI factory supercomputers are single systems located in individual data centres. However networked access and operation across different data centres and institutions is not uncommon. The infrastructure is shared across borders. For example, Spain’s AI factory is shared among users across Spain, Romania, Portugal and Turkey. France’s AI factory strategy includes collaboration with Germany’s Jupiter AI Factory (JAIF) “to federate our EuroHPC’ two exascale systems toward a virtual supercomputer of 50,000 GPUs allowing federated learning at scale.”

AI factory applications cover disciplines such as Engineering, Mathematics, and Computer Science, Chemical Sciences, Material Sciences, Solid state physics, Computational physics, and Universal Sciences such as the fundamental constituents of matter.

Across the different locations AI factories may have key sector specialisms that reflect the economic and societal priorities of their home country.

These include Agri-food, Earth Sciences, Cybersecurity, Manufacturing (Italy’s Leonardo supercomputer).
Engineering, Manufacturing, Mobility, Automotive technologies and Academia (Germany’s JAIF).
LLMs for Bulgarian Language, Space Observation, Robotics, Consumer Goods, Agriculture and Food Production, Healthcare (Bulgaria), Energy, Aerospace, Health, Defence and security, Earth Sciences, Digital continuum, Material Sciences, Robotics, Agriculture, Finance and legal AI for Education (edtech), Mobility, Humanities (France).

What is the IT inside an AI factory? (and why this matters)

The performance of the centralized High Performance Compute (HPC) machines is measured in petaflops (a petaflop is one thousand million million floating point operations per second).

This is rising to exascale compute (1 quintillion (1 to the power of 18) operations per second), 100s of petabytes of storage and petabit speed networking. It matters because  the higher the number of flops, the more processors required the more power and cooling needed.

Separate supercomputers have different chip, networking and storage architectures.

Germany’s final Jupiter Supercomputer will run on approximately 24,000 NVIDIA GH200 Grace Hopper Superchips specifically optimized for computationally intensive simulations and the training of AI models. This will enable JUPITER to achieve more than 70 ExaFLOP/s in lower-precision 8-bit calculations, making it one of the world’s fastest systems for AI.

Others run on DGX SuperPods – DGX is an integrated server, storage and networking platform machine range made by NVIDIA.

Many supercomputers run on CPU/GPU partitions – that is a mix of GPU and CPU architectures. For example Italy’s Leonardo runs on a GPU partition delivering 240 petaflops and an x86 CPU partition provider 9 petaflops. The GPU partition is based on 12,824 “Da Vinci” Ampere GPUs (NVIDIA). It is in the process of building a new supercomputer using 20,000 GPUs.

Different AI factories run on different mixes of Intel , AMD, ARM and NIVIDA chipsets (and others).

 

How much power will an AI factory data centre require?

In these days of commercial AI data centres of 500 MWs to multiple GWs the EU HPC supercomputer locations themselves could not be described as scale out data centres.

However, the EU’s AI factories are expanding rapidly through plans to grow the power requirement for the next generation of AI supercomputers.

Germany’s Jupiter supercomputer is a true exascale machine. According to estimates, Jupiter energy requirements will average around 11 MW with reports saying it will require 20MWs at peak load for running workloads such as AI modelling.

Much of the Jupiter data centre was factory constructed as the Supercomputing Centre (JSC) took a modular approach.

The data centre consists of around 50 container modules across an area of over 2,300 square metres – roughly half the size of a football pitch. The hardware for JUPITER’s booster module will occupy 125 racks, which were pre-installed at Eviden’s factory in Angers, France, and then shipped to Jülich ready for operation. The modular data centre is designed to extract the heat generated during cooling and use it to heat other buildings on the Forschungszentrum Jülich campus.

In Italy the Leonardo supercomputer is one of the three EuroHPC JU precursor exascale systems. According to Cienca – the institute operating Leonardo, the system will be housed in a the newly built data centre in the Bologna Big Data Technopole. It says: “In the first stage (2021-2025) the data centre will feature: 10 MW of IT load, 1240 m2 of computing room floor space, 900 m2 of ancillary space, a direct liquid cooling capacity of 8MW, a chilled water (18o -23o) cooling capacity of 6+2 MW and power capacity of 3+1 MW (no-break) and 9+3 MW (short-break).

The second stage (2025-2030) will see an increase to 20 MW IT load available and an additional 2600 m2 computing room space floor available. In designing the data centre, particular care was devoted to containing the PUE, estimated to be below 1.1. Simulation of the PUE was based on loads and losses calculated for all systems (IT, mechanical, and electrical). “

Elsewhere Spain’s Barcelona based 314 petaflops MareNostrum 5 “is located in BSC’s new facilities, next to the Chapel which is hosting previous systems. The datacentre has a total power capacity of 20MW, and cooling capacity of 17MW, with a PUE below 1,08.”

The new Finnish Lumi system is fully powered by carbon-free renewable energy, and its waste heat utilized in the district heating system for the town of Kajaani.

However, whether publicly funded or commercial, both types will face the challenge of sourcing, reliable clean power.

Globally the power forecast requirements for AI data centres combined with mass electrification of economies and the growing reliance on renewable energy sources are hitting the grid from all angles.

The International Energy Agency (IEA)’s special report Energy and AI, from April 2025 provides a data-driven global analysis on the growing connections between energy and AI. The report drew on new datasets and consultation with policy makers, the tech sector, the energy industry. It said: “Electricity demand from data centres worldwide is set to more than double by 2030 to around 945 terawatt-hours (TWh), slightly more than the entire electricity consumption of Japan today.”

It is hard to see how Europe’s ageing power grid, already under strain, with over half of its infrastructure exceeding 40 years in age, will be able to cope with the energy demands of AI. Expecting the grid to take the load is not only impractical but also risks alienating the local populations.  

Conclusion: Why Dispatchable Power is Vital for AI Factory Success

Clean dispatchable power — often delivered through modular microgrids using sustainable fuels — enables heat reuse, supports local grids, and delivers the flexibility AI factories need.

The characteristics of prime solutions (e.g. Microgrids) built around components that include clean fuel generators, containers, switchgear and battery storage align with design features common to AI data centres such as modularity, prefabrication, repeatability, factory quality, resilience, performance, management and speed to deployment.

Wherever suitable there is also a strong case for the integration of wind, solar and tidal renewable power generation with on-site microgrids. These are becoming more common across Europe.

As a region and across each individual European country planning, permitting and construction of microgrids for AI data centres must accelerate.

Events in the US highlight why this is vital.

In a statement to the US Senate committee hearing called ‘Winning the AI Race” in early May 2025 OpenAI’s CEO Sam Altman, equated energy, data centres and supercomputers and called for speedier power planning and permitting regimes to spur data centre building.

An example of a US microgrid powered AI factory that is already operational is the Sandia National Laboratory in the US which runs an advanced solar/battery/hydrogen microgrid as part of its investment in microgrid-based modular green data centers.

Currently the majority of LLM building and training is happening in large data centres in the US. But as high volume AI inference, agentic AI and reasoning roll out across different industries (including proximity AI for latency sensitive applications) powering the growth in the number and type of AI factories will require the latest microgrid technologies, often integrated with renewables.

AVK is already helping operators in Europe with 20MW-30MW data centers that are being planned in AI zones with the deployment of modular, low-emission microgrids at the heart of AI-ready infrastructure. From feasibility to full deployment, AVK’s unrivalled sector expertise and experience is providing the power resilience and speed these projects demand.

The JHPC AI factory strategy is hugely welcome. In the broader market the region cannot afford to wait for the grid to transform. Demand for commercial AI factories is not slowing down. The advantages of dispatchable power solution — including microgrids — as bridging and baseload power for speed to market for AI data centres are clear. A mix of microgrids and off-grid renewable energy will have a major part to play in ensuring Europe can compete in the AI arms race. 

It is a race that has already begun.