European AI gigafactories: the true, the false and the uncertain

In the race for gen­er­at­ive arti­fi­cial intel­li­gence, we are see­ing a rush to build data centres. These facil­it­ies, which are essen­tial for train­ing and run­ning AI mod­els, are driv­ing an increas­ing num­ber of con­struc­tion pro­jects around the world. Keen to break free from Amer­ic­an dom­in­ance in data centres, coun­tries are devel­op­ing their own pro­jects, as evid­enced by the €20 bil­lion invest­ment by the European Com­mis­sion, which, in Feb­ru­ary 2025, com­mit­ted to estab­lish­ing four to five AI giga­factor­ies with­in the European Uni­on (EU). They will be equipped with around 100,000 of the most advanced pro­cessors, roughly four times more than the cur­rent gen­er­a­tion of data centres.

These “giga­factor­ies” will com­ple­ment the infra­struc­ture already being installed in Europe and will aim to sup­port industry, start-ups and the European research eco­sys­tem. The ini­ti­at­ive forms part of the AI Con­tin­ent Action Plan, which sets out a roadmap to pos­i­tion Europe as a glob­al lead­er in the field of AI through invest­ment in infra­struc­ture, as well as by improv­ing access to data, sup­port­ing the adop­tion of AI across vari­ous sec­tors, and devel­op­ing skills and reg­u­la­tions. The EU is set to launch the call for tenders for the giga­factor­ies in early 2026, as part of the Inves­tAI ini­ti­at­ive. At the same time, private play­ers such as Nvidia and Microsoft, which plans to invest $10 bil­lion in a data centre in Por­tugal, have also announced plans for giga­factor­ies across Europe, fuel­ling com­pet­i­tion with the EU-fun­ded centres.

TRUE

Between 2023 and mid-2025, the share of private ven­ture cap­it­al invest­ment in AI start-ups was 66% in the United States, com­pared with just 12% in Europe, accord­ing to the AI World Index¹, an index developed by the Centre for European Policy Stud­ies (CEPS). The share of US pat­ents in the field of AI tech­no­lo­gies stands at 32%, com­pared to 18% for Europe (and 21% for China). Fur­ther­more, in 2024, the United States pro­duced more than 50% of all sig­ni­fic­ant AI mod­els, accord­ing to the Stan­ford AI Index. Europe, mean­while, pro­duced just 6% of them.

As for data centres, accord­ing to a recent report by a team from Oxford², the United States and China alone oper­ate over 90% of those spe­cial­ising in AI. Amer­ic­an tech giants, for their part, oper­ate 87 com­put­ing “hubs” world­wide, whilst Chinese firms con­trol 39 and European com­pan­ies 6. So, it is fair to say that the EU is trail­ing behind the United States and China.

UNCLEAR

The choice of loc­a­tion for a data centre involves tech­nic­al, eco­nom­ic, envir­on­ment­al and reg­u­lat­ory factors. A bal­ance must be struck between access to elec­tri­city and the grid, energy costs, cool­ing options, prox­im­ity to eco­nom­ic hubs, and the speed of author­isa­tion pro­ced­ures. Cer­tain recog­nised tech hubs, such as Dub­lin, Frank­furt and Ams­ter­dam, are reach­ing the lim­its of their elec­tri­city capa­city, for example. 

Accord­ing to a CEPS report pub­lished in Novem­ber 2025³, the dynam­ism of the eco­sys­tem is also a key cri­terion. How­ever, the sites already selec­ted for AI facil­it­ies do not cor­res­pond to the main European centres of excel­lence in AI, the authors note. The centres are being estab­lished where energy effi­ciency and exist­ing infra­struc­ture are favour­able, rather than where sci­entif­ic and entre­pren­eur­i­al activ­ity in the field of AI is strongest.

The report thus iden­ti­fies few matches between the centres of excel­lence and the chosen sites, apart from Île-de-France, Stut­tgart and Cologne, and a few oth­er loc­a­tions: in Bologna, Cata­lonia, Sweden and Poland, not­ably. This dis­per­sion of resources may appear prob­lem­at­ic. The report emphas­ises that the European Com­mis­sion will need to cla­ri­fy wheth­er the AI hubs them­selves are inten­ded to host sci­ent­ists and start-ups, or wheth­er geo­graph­ic­al cri­ter­ia such as com­put­ing capa­city and rel­at­ively low energy costs should take precedence.

TRUE

Gold­man Sachs Research fore­casts that glob­al elec­tri­city demand from data centres will rise by 50% by 2027 and by 165% by the end of the dec­ade, com­pared with 2023. Each giga­fact­ory will con­sume the equi­val­ent amount of elec­tri­city as a medi­um-sized town. In France, a new report by Ademe estim­ates that elec­tri­city con­sump­tion by data centres could increase 3.7‑fold by 2035 (37 TWh)⁴.

To recon­cile this with green trans­ition tar­gets, giga­factor­ies must be loc­ated in regions that com­bine an abund­ance of low-car­bon energy with cool­ing capa­city. They must also rely on the addi­tion of renew­able sources rather than divert­ing exist­ing capa­city. One con­cern is that these facil­it­ies may still con­sume too many resources and have a neg­at­ive impact on the loc­al envir­on­ment­al and social eco­sys­tem. They could con­flict with oth­er needs, such as elec­tric vehicles, indus­tri­al elec­tri­fic­a­tion or the decar­bon­isa­tion of build­ings, or com­pete for water resources with agri­cul­tur­al activ­it­ies, for example.

UNCLEAR

How­ever, based on elec­tri­city prices and the share of renew­able energy, the Scand­inavi­an coun­tries rank first, fol­lowed by Aus­tria, Por­tugal and Spain. Only AI plants in Sweden and Fin­land can bene­fit from sim­il­ar prices (€/MWh) to those of Amer­ic­an and Chinese hubs, accord­ing to the CEPS report. Dur­ing an ini­tial call for tender for its giga­factor­ies in June 2025, the European Com­mis­sion received 76 pro­pos­als from 16 Mem­ber States, far more than it had anti­cip­ated. Dis­cus­sions are under­way to merge con­sor­tia and pri­or­it­ise the best-posi­tioned coun­tries: those where bat­tery factor­ies already exist and where access to energy is cost-effect­ive. The CEPS ana­lys­is high­lights the import­ance of geo­graph­ic­ally con­cen­trat­ing giga­factor­ies in areas with favour­able energy effi­ciency and low-car­bon infrastructure.

FALSE

Brus­sels claims that these massive invest­ments will strengthen European tech­no­lo­gic­al sov­er­eignty by provid­ing Europe with the in-house capa­city to train state-of-the-art mod­els. The aim is to demo­crat­ise access to this infra­struc­ture to boost research and devel­op­ment among SMEs and start-ups. Accord­ing to what has been announced so far, how­ever, all giga­factor­ies will use Nvidia chips. Accord­ing to the report, this could under­mine the EU’s tech­no­lo­gic­al sov­er­eignty. Moreover, diver­si­fy­ing sup­pli­ers would not be enough to guar­an­tee sov­er­eignty. To train AI mod­els on Nvidia GPUs, the CUDA pro­gram­ming and com­put­ing mod­el is used – which is also owned by Nvidia. The com­pany there­fore also con­trols which chips are com­pat­ible with the soft­ware, thereby lim­it­ing altern­at­ives. Fur­ther­more, infra­struc­ture alone is not enough to make the EU inde­pend­ent. Invest­ment in train­ing, open data, inter­op­er­ab­il­ity stand­ards and eth­ic­al frame­works is also essential.

UNCLEAR

The tender require­ment stip­u­lates that the AI factor­ies must have plans to col­lab­or­ate with each oth­er. At present, pub­lic inform­a­tion on this sub­ject is lim­ited. The only excep­tion is the announce­ment of a Franco-Ger­man col­lab­or­a­tion, which will enable joint train­ing on com­put­ing across both facil­it­ies. In terms of co-author­ing and co-pub­lish­ing sci­entif­ic papers, there have been very few part­ner­ships between the regions where the AI facil­it­ies are loc­ated, mean­ing it will be essen­tial to build new bridges.

From a tech­nic­al per­spect­ive, the facil­it­ies are not yet fed­er­ated, mean­ing there is still no single point of access to com­put­ing resources and no fibre-optic con­nec­tions between the facil­it­ies. Fur­ther­more, it is unclear wheth­er the AI facil­it­ies’ infra­struc­ture sup­ports inter­op­er­ab­il­ity. Over­all, there is a clear inten­tion to col­lab­or­ate between the plants, both tech­nic­ally and organ­isa­tion­ally, but sig­ni­fic­ant effort will be required to make this work.

Interview by Célia Chaboud