LUMI AI Fac o y Se ice Cen e
Empowe ing Eu ope’s AI Ecosys em
D7.4 Se ice Po olio
2
D7.4
Se ice Po olio
D7.4 Se ice Po olio
3
P ojec Ti le
LUMI AI Fac o y Se ice Cen e
P ojec Ac onym
LUMI-AIF
P ojec Numbe
101234208
Type o Ac ion
HORIZON-JU-RIA
Topic
HORIZON-JU-EUROHPC-2025-AI-01-IBA-01
S a ing Da e o P ojec
01.03.2025
Ending Da e o P ojec
29.02.2028
Du a ion o he P ojec
36 mon hs
Websi e
lumi-ai- ac o y.eu
Wo k Package
WP7
Task
T7.3 Se ice po olio de elopmen
Lead Au ho s
Roge K am (Sigma2)
Con ibu o s
Basi Sedighi (Sigma2), Kalle Huh ala (CSC)
Pee Re iewe s
Ma in Duda (IT4I), Aleksi Kallio (CSC)
Ve sion
1.0
Due Da e
29.08.2025
Submission Da e
29.08.2025
Dissemina ion le el
X
PU: Public
SEN: Sensi i e – limi ed unde he condi ions o he G an Ag eemen
EU-RES. Classi ied In o ma ion: RESTREINT UE (Commission Decision 2005/444/EC)
EU-CON. Classi ied In o ma ion: CONFIDENTIEL UE (Commission Decision 2005/444/EC)
EU-SEC. Classi ied In o ma ion: SECRET UE (Commission Decision 2005/444/EC)
D7.4 Se ice Po olio
4
Ve sion His o y
Re ision
Da e
Edi o s
Commen s
0.1
27.05.2025
Roge K am
Fi s e sion
0.2
28.05.2025
Roge K am
Upda ed wi h plan o de elopmen o
oadmap
0.3
23.06.2025
Roge K am
Changed s uc u e o align wi h documen
o cus ome -o ien ed ep esen a ion
0.4
24.06.2025
Basi Sedighi &
Roge K am
Added ex o s uc u e
0.5
15.08.2025
Roge K am
Rew i ing i s sec ion
0.6
24.08.2025
Roge K am
Kalle Huh ala
Ex ending he execu i e summa y, adding
i ems o he Glossa y, comple ing 3.4.1 and
eplacing chap e 6 wi h e e ence o he
Wiki. Adding he chap e on Se ice
Ca alogue.
0.7
29.8.2025
Kalle Huh ala
Final edi ing
1.0
29.8.2025
Rebekka Lampola
Quali y check pe o med by he PMO and
sen o o icial e iew
Glossa y o e ms
I em
Desc ip ion
AI
A i icial In elligence: The heo y and de elopmen o compu e sys ems able o pe o m
asks ha no mally equi e human in elligence, such as isual pe cep ion, speech
ecogni ion, decision-making, and ansla ion be ween languages.
AIF
AI Fac o y: A cen alized hub p o iding compu ing esou ces, da a, ools, and expe ise
o de elop, deploy, and scale AI applica ions.
Ai low
An open-sou ce wo k low o ches a ion pla o m o iginally de eloped by Ai bnb,
designed o au ho , schedule, and moni o da a and machine lea ning pipelines.
API
Applica ion P og amming In e ace: A se o ules and p o ocols ha allows di e en
so wa e applica ions o communica e wi h each o he .
Ba ch Job
Scheduling
A me hod o managing and execu ing asks (jobs) on a compu ing sys em wi hou di ec
use in e ac ion, ypically used in High-Pe o mance Compu ing o op imize esou ce
usage.
CSC
The Finnish IT cen e o science, which hos s and ope a es he LUMI supe compu e on
behal o he Eu oHPC JU.
D7.4 Se ice Po olio
5
CUDA
Compu e Uni ied De ice A chi ec u e: A pa allel compu ing pla o m and p og amming
model c ea ed by NVIDIA o gene al compu ing on i s g aphical p ocessing uni s
(GPUs).
CV
Compu e Vision: A ield o AI ha ains compu e s o in e p e and unde s and he
isual wo ld om digi al images o ideos.
Da ase
A s uc u ed collec ion o ela ed da a, ypically p esen ed in a abula , ex ual, isual, o
nume ical o ma , o ganized o analysis, aining o AI/ML models, o o he
compu a ional use. Da ase s may con ain aw o cu a ed in o ma ion and can a y in
size om small samples o pe aby e-scale collec ions. In he con ex o he LUMI AI
Fac o y, da ase s a e p o ided as “Da ase -as-a-Se ice,” gi ing use s s anda dized,
secu e, and o en p e-p ocessed da a esou ces o esea ch, inno a ion, and AI
de elopmen , including sensi i e and con iden ial da a whe e compliance wi h
egula ions is equi ed.
Digi al Twin
A i ual model designed o accu a ely e lec a physical objec , p ocess, o sys em. I is
used o simula ion, es ing, and moni o ing.
Eu oHPC JU
Eu opean High-Pe o mance Compu ing Join Unde aking: A legal and unding en i y
ha enables he Eu opean Union and pa icipa ing coun ies o coo dina e hei e o s
and pool hei esou ces o deploy wo ld-class supe compu e s and echnologies.
Fede a ed
Lea ning
A machine lea ning echnique ha ains an algo i hm ac oss mul iple decen alized
de ices o se e s holding local da a samples, wi hou exchanging he da a i sel .
Fine- uning
The p ocess o aking a p e- ained model (like a la ge language model) and u he
aining i on a smalle , speci ic da ase o adap i o a pa icula ask.
Founda ion
Model
A la ge-scale AI model ained on a as quan i y o b oad da a ha can be adap ed
(" ine- uned") o a wide ange o downs eam asks.
GenAI
Gene a i e A i icial In elligence: A ype o AI ha can c ea e new and o iginal con en ,
such as ex , images, music, o code.
GPU
G aphics P ocessing Uni : Specialized elec onic ci cui s designed o apidly manipula e
and al e memo y o accele a e he c ea ion o images in a ame bu e in ended o
ou pu o a display de ice; hei pa allel s uc u e makes hem ideal o complex AI and
machine lea ning asks.
HIP
He e ogeneous-compu e In e ace o Po abili y: A C++ un ime API and ke nel
language ha allows de elope s o c ea e po able applica ions o AMD and NVIDIA
GPUs om a single sou ce code.
HPC
High-Pe o mance Compu ing: The use o supe compu e s and pa allel p ocessing
echniques o sol ing complex compu a ional p oblems.
IT4I
IT4Inno a ions Na ional Supe compu ing Cen e a VŠB – Technical Uni e si y o
Os a a, Czechia. A key na ional esea ch in as uc u e and a membe o he LUMIAIF
conso ium.
KAM
Key Accoun Manage : An indi idual esponsible o managing and nu u ing
ela ionships wi h an o ganiza ion's mos impo an cus ome s.
Kube low
An open-sou ce pla o m o de eloping, o ches a ing, and deploying scalable machine
lea ning wo k lows on Kube ne es
KPI
Key Pe o mance Indica o : A quan i iable measu e o pe o mance o e ime o a
speci ic objec i e. KPIs p o ide a ge s o eams o shoo o , miles ones o gauge
p og ess, and insigh s ha help people ac oss he o ganiza ion make be e decisions.
LLM
La ge Language Model: An ad anced ype o AI model designed o unde s and,
gene a e, and p ocess human language, ained on massi e da ase s o ex and code.
Examples include Llama and GPT models.
M
Mon h, ypically ollowed by a numbe deno ing he sequence om he p ojec s a
D7.4 Se ice Po olio
6
MFA
Mul i-Fac o Au hen ica ion: A secu i y p ocess ha equi es use s o p o ide wo o
mo e e i ica ion ac o s o gain access o a esou ce, such as an applica ion o online
accoun .
MLOps
Machine Lea ning Ope a ions: A se o p ac ices ha aims o deploy and main ain
machine lea ning models in p oduc ion eliably and e icien ly. I combines machine
lea ning, da a enginee ing, and De Ops.
MLFlow
An open-sou ce pla o m o managing he end- o-end machine lea ning li ecycle.
MOOC
Massi e Open Online Cou se: A ee online cou se a ailable o many people o en ol.
MS
Miles one, ypically ollowed by a numbe iden i ying i .
Mul i enancy
A so wa e a chi ec u e whe e a single ins ance o a so wa e applica ion se es mul iple
cus ome s ( enan s), wi h each enan 's da a isola ed and emaining in isible o o he
enan s.
NLP
Na u al Language P ocessing: A sub ield o AI ocused on enabling compu e s o
unde s and, in e p e , and gene a e human language.
QPU
Quan um P ocessing Uni : The co e p ocesso o a quan um compu e , which uses he
p inciples o quan um mechanics o pe o m calcula ions.
Sandbox
In an AI/so wa e con ex , an isola ed es ing en i onmen ha enables use s o un
p og ams o open iles wi hou a ec ing he applica ion, sys em, o pla o m on which
hey un.
Se ice
Ca alogue
A cus ome - acing lis o all li e se ices o e ed along wi h ele an in o ma ion abou
hese se ices. I can be ega ded as a il e ed, use - iendly iew o he Se ice Po olio.
Se ice Po olio
An in e nal lis ha de ails all he se ices o e ed by a se ice p o ide , including hose
in p epa a ion, li e, and discon inued. I con ains echnical speci ica ions, cos s, isks,
and alue p oposi ions.
SME
Small and Medium-sized En e p ise: A business whose pe sonnel numbe s and e enue
all below ce ain limi s.
T us wo hy AI
An app oach o a i icial in elligence ha emphasizes he impo ance o de eloping AI
sys ems ha a e law ul, e hical, and echnically obus , ensu ing hey espec human
alues and undamen al igh s.
T y-n-buy
A se ice model ha allows o ganiza ions o expe imen wi h compu ing and da a
esou ces on a limi ed scale be o e commi ing o ull adop ion.
WP
Wo k Package: A majo sub-di ision o a p ojec . Each wo k package has a se o de ined
asks, deli e ables, and objec i es.
D7.4 Se ice Po olio
7
Execu i e summa y
The LUMI AI Fac o y is a h ee-yea Eu opean ini ia i e om Ma ch 2025 o Feb ua y 2028 designed o
accele a e AI inno a ion by p o iding a comp ehensi e one-s op-shop buil on he LUMI supe compu e .
I aims o empowe SMEs and s a -ups, esea che s and public sec o ac oss Eu ope.
This documen will desc ibe he oadmap o he co e se ices o he LUMI AI Fac o y and he planned
oadmap o deli e y o he ma u i y le els o he se ices, as well as when he se ices will be deployed,
and o whom he se ices will be a ailable.
The Se ice Po olio will be di ided in o a Se ice Ca alogue o ex e nal use and a Se ice Po olio o
in e nal use. The documen includes a Se ice Ca alogue d a wi h ini ial in o ma ion on he a ailabili y
o he se ices.
Key se ices will be p og essi ely olled ou and include:
• Powe ul AI Compu ing: Scalable access o GPU-accele a ed esou ces o aining and
deploying e e y hing om con en ional ML models o massi e ounda ion models, including
secu e en i onmen o sensi i e da a
• Da a Se ices: Access o a cu a ed lib a y o da ase s ("Da ase -as-a-Se ice") and expe
suppo o da a managemen and in eg a ion
• Expe Suppo & AI Adop ion: Di ec consul a ion o echnical challenges, g an applica ions,
and business- ocused p og ams like "T y & Buy" p ojec s o ease in o la ge-scale AI
• T aining & Skill De elopmen : A ull ange o cou ses, hacka hons, summe schools, and an
accele a o p og am o AI s a -ups
• Collabo a ion Hubs: A physical co-wo king hub in O aniemi, Finland, including i ual spaces,
ma chmaking and communi y building ac oss bo de s
Th ough his oadmap, he LUMI AI Fac o y will no only deli e ad anced echnical se ices bu also
measu able impac : s eng hening Eu opean so e eign y in AI, lowe ing en y ba ie s o SMEs,
accele a ing public-sec o digi alisa ion, and equipping a new gene a ion o AI p o essionals wi h he
skills o h i e in a apidly e ol ing ecosys em.
This documen will be e ised annually wice o e lec in inc easing de ail he cu en and e ol ing s a us
o he se ice de elopmen oadmap, po olio and ca alogue.
D7.4 Se ice Po olio
8
Table o Con en s
1. In oduc ion............................................................................................................. 10
2. Implemen a ion oadmap and se ice ma u i y imeline ........................................ 11
3. Se ice Po olio oadmap ........................................................................................ 14
3.1 Co e AI compu ing se ices 14
3.1.1 Key ea u es 14
3.1.2 Ini ial se up 14
3.1.3 Roadmap 15
3.2 Access o da a 15
3.2.1 Key ea u es 15
3.2.2 Ini ial se up 15
3.2.3 Roadmap 15
3.3 Suppo o AI me hods 16
3.3.1 Key ea u es 16
3.4 AI adop ion o companies 17
3.4.1 Key accoun manage se ices in complex p ojec s 17
3.4.2 HPC/AI s a e pack 17
3.5 T aining cou ses and skills de elopmen 17
3.5.1 T aining cou ses on undamen als o AI 17
3.5.2 T aining cou ses on la ge scale AI and High-Pe o mance Compu ing (HPC) 18
3.5.3 Access o online aining ma e ials and MOOCs 18
3.5.4 En olmen in Summe school o Hacka hon 18
3.5.5 En olmen in Accele a o P og am 19
3.6 Co-wo king spaces 19
3.6.1 O aniemi AI Hub (Espoo, Finland) 19
3.6.2 Loca ion and access 20
3.6.3 P emises o wo kshops 20
3.6.4 Pe sonal and eam wo king space 20
3.6.5 Vi ual co-wo king space 20
D7.4 Se ice Po olio
9
3.6.6 AI Campus se ices o s uden s 20
3.6.6.1 Campus loca ions and ac i i ies 20
3.6.6.2 LUMI-AI co-labo a o y 20
3.6.6.3 Gene al In o ma ion and access 21
3.7 E en s and collabo a ion 21
3.7.1 En olling in upcoming e en s wi hin AI Fac o y ecosys ems 21
3.7.2 Reques ing collabo a o sugges ions om he AI Fac o y ecosys ems 22
3.7.3 Reques ing LUMI AI Fac o y s a o gi e p esen a ions a you e en 22
4. Se ice Po olio de elopmen oadmap ................................................................. 23
4.1 Inpu s 23
4.2 Timelines 23
5. Se ice Ca alogue shown o end use s ...................................................................... 23
5.1 Se ice Ca alogue managemen and da a model 24
5.2 Se ice Ca alogue d a 25
6. Deli e ables and KPIs .............................................................................................. 28
D7.4 Se ice Po olio
16
• M02–M10: Iden i ica ion o ele an da ase s and license nego ia ion
• M14: A leas 100 da ase s a ailable ia he se ice
• M18+: Suppo o public-p i a e da a collabo a ion, comme cial license
handling, and au oma ic da a s aging
3.3 Suppo o AI me hods
3.3.1 Key ea u es
• HPC op imiza ion and AI de elopmen
▪ The LUMI AI Fac o y p o ides suppo o op imizing High-Pe o mance
Compu ing (HPC) o AI de elopmen . This se ice is pa o he in-dep h AI
expe guidance and suppo a ailable o acili a e he success ul
implemen a ion o AI p ojec s, co e ing ad anced opics in HPC op imiza ion o
AI de elopmen .
▪ Du a ion: 2 mon hs
• Dis ibu ed AI model aining
▪ Expe suppo is o e ed o dis ibu ed AI model aining. This alls unde he in-
dep h AI expe suppo p o ided by he AI Fac o y and is pa o he "AI expe
suppo o me hods and code op imiza ion and scalabili y imp o emen ." The
suppo co e s ad anced opics including dis ibu ed AI model aining, model
ine- uning, and hype pa ame e sea ch.
▪ Du a ion 2 mon h
• AI me hod Consul a ion Assignmen
o The LUMI AI Fac o y o e s AIF (AI Fac o y) specialis suppo o AI me hods. This is
p o ided as in-dep h AI expe suppo o help cus ome s wi h speci ic AI p ojec s.
The AI me hods suppo eam includes specialis s o a ious AI sub ields, enabling
hem o p o ide cus ome -speci ic AI consul a ion on ele an a eas such as:
▪ Na u al Language P ocessing (NLP) and La ge Language Models (LLMs)
▪ Compu e Vision (CV)
▪ Rein o cemen Lea ning (RL) and Robo ics
▪ Digi al Twins and AI o Science applica ions
o Du a ion: 1 mon h
• Reques o Cus ome -speci ic AI Consul a ion
▪ Cus ome -speci ic AI consul a ion is a se ice a ailable h ough he LUMI AI
Fac o y. This is pa o he b oad ange o AI expe guidance and suppo
o e ed o acili a e he success ul implemen a ion o AI p ojec s. Specialis
eams p o ide his cus ome -speci ic AI consul a ion ac oss a ious sub ields
o AI.
▪ Du a ion: 6 mon hs
D7.4 Se ice Po olio
17
• G an applica ion suppo consul a ion
▪ Consul a ion mee ing wi h AI Fac o y expe s, p o iding suppo o w i ing
g an applica ions, such as applica ions o na ional inno a ion unding o
Eu opean p og ams including Digi al Eu ope P og am o Ho izon Eu ope.
Consul a ion ocuses on echnical easibili y o he planned AI o HPC
app oaches, as well as hei implemen a ion on he LUMI AI Fac o y
en i onmen .
3.4 AI adop ion o companies
3.4.1 Key accoun manage se ices in complex p ojec s
• Cus ome Needs Analysis will be conduc ed o unde s and he cus ome 's expe ise and
AI eadiness
• Guidance o app op ia e se ices
• Single poin o con ac
3.4.2 HPC/AI s a e pack
Dedica ed o e ing o b ing manu ac u ing, enginee ing, li e sciences and pha maceu ical
companies close o HPC and AI solu ions and me hodologies h ough a ge ed se ices and
suppo ia HPC/AI s a e pack o companies, T y & Buy es p ojec s, o consul a ion o
T us wo hy AI.
3.5 T aining cou ses and skills de elopmen
The LUMI AI Fac o y places a signi ican emphasis on aining cou ses and skills de elopmen o
empowe Eu ope’s AI ecosys em, add ess he alen gap, and accele a e he adop ion and
de elopmen o AI solu ions. This comp ehensi e app oach is designed o bene i a di e se ange
o s akeholde s ac oss Eu ope.
3.5.1 T aining cou ses on undamen als o AI
The LUMI AI Fac o y p o ides ounda ional lea ning oppo uni ies h ough a modula
and s uc u ed aining p og amme, pa icula ly o use s wi h limi ed o no AI and/o
HPC backg ound.
• ”Fundamen als o machine lea ning”: Co e s essen ial concep s o machine lea ning,
including s a is ical, p obabilis ic, and compu a ional p inciples.
• ”In oduc ion o using AI models and ools”: P o ides basic skills o using exis ing
AI models, such as la ge language models, as pa o esea ch and de elopmen
ac i i ies.
D7.4 Se ice Po olio
18
• ”P ac ical deep lea ning”: Del es in o deep lea ning and common neu al ne wo k
a chi ec u es, GPU compu ing, and ools o aining and applying deep neu al
ne wo ks in a eas like NLP and compu e ision.
These co e cou ses a e scheduled a leas wice a yea and can be deli e ed online o on-
si e.
3.5.2 T aining cou ses on la ge scale AI and High-Pe o mance Compu ing
(HPC)
Specialized aining is o e ed o use s wi h exis ing cloud o AI expe ise o acili a e he
con inuous upda ing o hei skills.
• ”AI o HPC”: Co e s AI amewo ks (PyTo ch), con aine s, isualiza ion, dis ibu ing
and scaling AI wo kloads ac oss mul iple GPUs/nodes, and op imiza ion
echniques.
• ”AI wo k lows and MLOps”: Focuses on AI enginee ing echniques o building
model aining wo k lows and MLOps o deploy and main ain models, ensu ing
scalabili y, ep oducibili y, and eliabili y.
• ”GPU p og amming”: Co e s low-le el GPU p og amming using echnologies like
CUDA/HIP, OpenMP/OpenACC, SYCL, and Kokkos o op imal esou ce u iliza ion.
• ”Applica ion pe o mance analysis and op imiza ion”: In ol es aining on p o ile s,
message passing analysis, ha dwa e coun e ools, and op imiza ion echniques.
3.5.3 Access o online aining ma e ials and MOOCs
High-quali y, comp ehensi e sel -lea ning ma e ials a e p o ided o enable use s o
acqui e skills a hei own pace.
• MOOCs:
– ”LUMI AI Fac o y compu ing en i onmen ”: Co e s he basics o he LUMI
and LUMI-AI supe compu e en i onmen s.
– ”Elemen s o supe compu ing”: In oduces he basic p inciples o high-
pe o mance compu ing.
• Online Documen a ion: Comp ehensi e use guides and documen a ion o
so wa e packages and ools.
• AI Assis an : A cha bo will be a ailable o answe ques ions, gene a e sc ip s, and
analyse da a, buil on an open-sou ce LLM ained on LUMI-AI.
• In eg a ed se ice po al: A single poin o en y o all in o ma ion and se ices,
in eg a ing web pages, documen a ion, and use po als o esou ce managemen .
3.5.4 En olmen in Summe school o Hacka hon
D7.4 Se ice Po olio
19
The LUMI AI Fac o y ac i ely o ganizes wo kshops, hacka hons, and summe schools as
pa o i s s uc u ed aining p og am.
• Hacka hons:
– “Mo ing and op imizing you AI jobs o LUMI-AI supe compu e ”: Helps
use s po , modi y, and op imize hei AI so wa e o he supe compu e .
– “AI s uden s hacka hon”: A collabo a i e e en based on indus y-p o ided
use cases and da ase s.
• Summe Schools:
– “T aining and using AI on he LUMI-AI supe compu e ”: Co e s he
LUMI/LUMI-AI en i onmen , AI amewo ks, dis ibu ed AI, and
op imiza ion.
– “AI o science summe school”: In eg a es compu a ionally in ensi e
simula ions wi h AI and co e s he de elopmen o digi al wins.
Pa icipan s ecei e a ce i ica e wi h a ecommenda ion on co esponding ECTS c edi s.
3.5.5 En olmen in Accele a o P og am
An accele a o p og am o AI s a ups is planned o p omo e he excellence o LUMI-AI
and suppo he de elopmen o ad anced models.
• Launch: The AI s a up incuba o p og amme is expec ed o be launched by Mon h
10 (app ox. Decembe 2025).
• Scope: The p og am will suppo one ba ch o 10–20 s a ups om Eu oHPC membe
s a es.
• Fas -T ack Model: A “T y & Buy es p ojec ” model allows o ee-o -cha ge es ing
wi h a small esou ce alloca ion and expe suppo be o e egula use.
3.6 Co-wo king spaces
The LUMI AI Fac o y places a s ong emphasis on os e ing a collabo a i e and inno a i e AI
ecosys em h ough i s di e se ange o co-wo king spaces and se ices. This includes bo h
physical and i ual en i onmen s designed o suppo s uden s, s a ups, SMEs, esea che s,
and AI p o essionals ac oss Eu ope.
3.6.1 O aniemi AI Hub (Espoo, Finland)
The O aniemi AI Hub se es as he main physical co-wo king space o he LUMI AI
Fac o y, s a egically loca ed o concen a e AI alen and expe ise.
D7.4 Se ice Po olio
20
3.6.2 Loca ion and access
The hub is si ua ed on he O aniemi campus o Aal o Uni e si y a Tie o ie 2, Espoo,
Finland. I is co-loca ed wi h he upcoming ELLIS Ins i u e Finland and su ounded by key
na ional AI compe ence cen es, s a up campuses, and incuba o s. The a ea has
excellen connec i i y o Helsinki in e na ional ai po and is easily accessible ia public
anspo a ion. Onboa ding ac i i ies and ou s will be a ailable h ough he in eg a ed
se ice po al.
3.6.3 P emises o wo kshops
The hub is equipped wi h ad anced physical aining acili ies well-sui ed o wo kshops.
These include a ully equipped compu e class oom, wi eless ne wo k access, and a
nea by audi o ium o la ge e en s. Addi ional acili ies a e p o ided by CSC in
Keila an a and collabo a o s a Aal o Uni e si y and he Uni e si y o Helsinki, allowing
o a high numbe o pa allel aining sessions.
3.6.4 Pe sonal and eam wo king space
The main co-wo king space ea u es an open o ice wi h desks o lexible wo king. Fo
con iden ial con e sa ions o ocused wo k, p i a e phone boo hs a e a ailable. The
space is adminis a ed by a local coo dina o and has pe manen on-si e s a o assis
isi o s.
3.6.5 Vi ual co-wo king space
The LUMI AI Fac o y complemen s i s physical spaces wi h a obus i ual wo king space
o enable e ec i e collabo a ion ac oss pa icipa ing coun ies. This digi al en i onmen
p o ides a sui e o ools o eam collabo a ion, including:
• Sha ed documen managemen
• Co-c ea ion ools
• Wo k managemen ools
• Discussion o ums
• Videocon e encing ools
These esou ces a e con inuously a ailable o suppo collabo a ion and inno a ion
wi hin he AI ecosys em o he du a ion o he p ojec .
3.6.6 AI Campus se ices o s uden s
3.6.6.1 Campus loca ions and ac i i ies
The AI Fac o y Hub in O aniemi will include a dedica ed s uden hos ing acili y. Smalle
AI Campuses a e also planned o Os a a (hos ed by IT4I) and o he pa ne acili ies like
he Uni e si y o Helsinki’s Kumpula campus. These campuses o e ac i i ies such as
s udy g oups, pee - u o ing, ma chmaking e en s, ne wo king, isi s, and lec u es.
S uden s will also ecei e use accoun s wi h small compu ing en i onmen s o p ac ical
HPC amilia iza ion.
3.6.6.2 LUMI-AI co-labo a o y
D7.4 Se ice Po olio
21
This key se ice o e s use - iendly web-based access o sho - ime GPU compu ing
esou ces, ideal o cou se exe cises, sel -s udy, and hesis wo k. I p o ides Jupy e
No ebooks and o he web-based in e aces, allowing access o aining en i onmen s
and ma e ials any ime wi hou needing local so wa e ins alla ions.
3.6.6.3 Gene al In o ma ion and access
The LUMI AI Fac o y o e s an in eg a ed se ice po al which se es as a single poin o
en y o all in o ma ion and se ices. This po al will in eg a e public webpages, use
documen a ion, and use po als o managing esou ces and di ec ly using se ices ia
Open OnDemand. Open se ices will be a ailable o all pa icipan s om Eu oHPC
membe s a es ia his po al. Fo ailo ed se ices, a Cus ome Needs Analysis may be
conduc ed o guide use s o sui able o e ings.
3.7 E en s and collabo a ion
The LUMI AI Fac o y is designed o be a comp ehensi e hub o AI inno a ion, o e ing a ious
se ices o os e collabo a ion, skill de elopmen , and he widesp ead adop ion o AI ac oss
Eu ope. This includes bo h oppo uni ies o pa icipa e in e en s and o le e age he expe ise
and ne wo ks o he AI Fac o y.
3.7.1 En olling in upcoming e en s wi hin AI Fac o y ecosys ems
• Gene al access o e en in o ma ion: The p ima y en y poin o all in o ma ion
and se ices, including e en schedules and egis a ion, is he in eg a ed se ice
po al o he LUMI AI Fac o y.
• Types o e en s: A ange o ac i i ies a e designed o a ious skill le els and a ge
g oups, including:
– S uc u ed aining p og ams
– Wo kshops on speci ic domains
– Hacka hons based on indus y use cases
– Summe schools
– Webina s and lec u es
– Ma chmaking and ne wo king e en s
• Eu opean-wide e en s: Join aining e en s, wo kshops, and hacka hons a e planned
wi h o he Eu oHPC AI Fac o ies (expec ed wo pe yea ). The AI Fac o y also
suppo s an Annual Eu opean AI Fac o ies e en o connec he en i e AI
communi y.
• Na ional ecosys em e en s: The AI Fac o y ac i ely engages wi h and suppo s
na ional AI ecosys ems in conso ium coun ies (Czechia, Denma k, Es onia,
D7.4 Se ice Po olio
22
Finland, No way, Poland). Na ional pa ne s will con ibu e o he aining
po olio, wi h e en s accessible h ough he common po al.
• En olmen p ocess: The in eg a ed se ice po al will acili a e cou se and e en
egis a ion. Mos open se ices a e a ailable o all pa icipan s om Eu oHPC
membe s a es.
3.7.2 Reques ing collabo a o sugges ions om he AI Fac o y ecosys ems
The LUMI AI Fac o y is a ca alys o collabo a ion and inno a ion wi hin he AI
ecosys em.
• Ma chmaking se ices: A co e componen is ma chmaking be ween AI de elope s,
da a p o ide s, and use s o acili a e new pa ne ships, especially o public-p i a e
collabo a ions and g an p oposals.
• Access o ecosys em insigh s: The AI Fac o y’s ex ensi e ne wo k p o ides deep
insigh in o ele an AI ecosys ems, helping use s iden i y and connec wi h sui able
collabo a o s.
• S akeholde engagemen : Key Accoun Manage s assis in de ining p ojec scope
and encou age companies o build collabo a i e p ojec s wi h uni e si ies, esea ch
ins i u es, and public bodies.
• AI Campus and incuba o p og ams: The AI Campus acili ies hos ma chmaking
e en s. The AI s a up incuba o p og am (launching app ox. Dec 2025) p o ides
u he a enues o connec ing wi h p omising s a ups.
3.7.3 Reques ing LUMI AI Fac o y s a o gi e p esen a ions a you e en
The LUMI AI Fac o y places a s ong emphasis on communica ion, ou each, and
knowledge dissemina ion.
• Expe p esen a ions: The AI Fac o y’s di e se pool o AI, HPC, and domain-speci ic
expe s a e a ailable o gi e p esen a ions on a a ie y o opics.
• Communica ion and ou each s a egy: A dedica ed wo k package (WP8) is asked
wi h aising awa eness o he AI Fac o y’s esou ces and suppo ing hei up ake
h ough e icien communica ion and dissemina ion.
• Knowledge sha ing and aining: The AI Fac o y’s expe s a e ac i ely in ol ed in
educa ing he communi y h ough aining e en s, webina s, wo kshops, and
hacka hons.
• How o eques : To ini ia e a eques o collabo a o sugges ions o o a
p esen a ion, use s can con ac he ele an eams (e.g., Cus ome Engagemen ,
AI Expe Suppo , o Communica ion) h ough he in eg a ed se ice po al.
D7.4 Se ice Po olio
23
4. Se ice Po olio de elopmen oadmap
4.1 Inpu s
• Ini ial cus ome su ey de eloped and pe o med h ough ask 7.1 and Deli e able 7.1
o Needs o indus y
o Needs o SMEs and s a ups
o Needs o public adminis a ion
o Needs o academia and esea ch
• Inpu om con inuous sc eening in ask 7.1 o po en ial:
o Cus ome s and Use s
o Technologies and la es inno a ion
o Compe i o s and o he ma ke da a
• Business model planning, cos e sus ewa d analy ics in ask 7.2
• Cus ome p ocess
o Cus ome jou ney analy ics om onboa ding p ocesses in ask 2.1 and ask 2.2 (D2.2)
o Cus ome needs analy ics in ask 2.3
▪ In-dep h discussions wi h cus ome s
▪ Well-s uc u ed su eys
▪ Ma u i y su eys p io o onboa ding new cus ome s (D2.3)
o HPC and AI p ojec implemen a ion expe ience
o Dialog wi h KAM
4.2 Timelines
• Re iew oadmap in M6, M12, M18, M24, M30
5. Se ice Ca alogue shown o end use s
The Se ice Ca alogue is de ined as ollows: “Cus ome - acing lis o all li e se ices o e ed along wi h
ele an in o ma ion abou hese se ices. No e 1: A Se ice Ca alogue can be ega ded as a il e ed
e sion o and cus ome s’ iew on he Se ice Po olio.”
Some se ices a e al eady a ailable wi h exis ing LUMI supe compu e . O he s can be made a ailable o
cus ome s soon, while some se ices need he ac ual LUMI-AI and inally LUMI-IQ machines o be
a ailable.
D7.4 Se ice Po olio
24
5.1 Se ice Ca alogue managemen and da a model
The in o ma ion p esen ed in he Se ice Ca alogue mus be use - iendly and in o ma i e, conside ing
he expec ed spec um o use s anging om s a ups and adi ional en e p ises o academia and
esea ch p o essionals. P oducing sui able se ice desc ip ions calls o co-ope a ion be ween Se ice
Owne s and Communica ions.
A eposi o y ( able, da abase, exis ing se ice, o be de ined) o s o ing se ice in o ma ion is needed,
along wi h he p ocedu es o managing his in o ma ion.
A da a model / empla e o desc ibing he se ice de ails is needed. Such a empla e will be de ined
ASAP o he pu pose o opening he LUMI AIF websi e in ea ly Sep embe 2025.
The se ice desc ip ion empla e should include e.g. he ollowing opics ha a e use ul o end use s. Fo
managing se ices in e nally, many mo e de ails would be equi ed (e.g. se ice desc ip ion ields in he
Se ice Now -da abase used by CSC o main ain i s se ice o e ing):
• Se ice Name: [En e AI Se ice Name He e]
• Se ice Summa y: [B ie o e iew o he AI se ice]
• Se ice Ca ego y: [e.g., Machine Lea ning, NLP, Compu e Vision, T aining, Consul a ion…]
• Se ice Owne : [Team o indi idual esponsible]
• Ta ge Use s: [Who can use his se ice]
• Se ice ea u e lis
• [Fea u e 1]
• [Fea u e 2]
• Se ice Le els (SLAs): [A ailabili y, esponse ime, e c.]
• Reques P ocess: [How o eques he se ice]
• Suppo In o ma ion: [Con ac de ails, suppo hou s]
• Cos and Billing: [P icing model i applicable]
• Dependencies: [O he se ices o sys ems equi ed]
• Secu i y and Compliance: [Da a p o ec ion and egula ions]
• Change and Main enance Schedule: [Planned upda es]
• Se ice S a us: [Cu en ope a ional s a us]
• e c.
D7.4 Se ice Po olio
25
5.2 Se ice Ca alogue d a
The ollowing lis is he i s d a o LUMI AI Fac o y Se ice Ca alogue. The se ices lis ed he e will be
shown wi h app op ia e in o ma ion on he LUMI AIF websi e, wi h he se ices ma ked ei he as
a ailable o wi h a no e o es ima ed a ailabili y. The eigh ca ego ies o se ices co espond o he
Se ice Po olio. An es ima e o he s a o se ice a ailabili y is ma ked in pa en heses a e he
se ice name. E.g. “2026” means he se ice will be a ailable in ha yea , no necessa ily beginning ha
yea .
1. AI compu ing capaci y
1.1. Small compu ing package (A ailabili y: 2025+)
• Sui able o aining con en ional machine lea ning models
1.2. Medium compu ing package (2025 wi h limi a ions. Fully in 2026)
• Sui able o ine- uning la ge language models
1.3. La ge compu ing package (2025 wi h limi a ions. Fully in 2026)
• Sui able o aining la ge language models o up o 10B pa ame e s
1.4. G and challenge package (2025 wi h limi a ions. Fully in 2027)
• Sui able o aining la ge language models o 10-500B million pa ame e s o o he e y la ge
ounda ion models
1.5. Quan um compu ing package (No ye )
• Sui able o compu ing asks equi ing he combining o quan um compu ing and AI o HPC
compu ing
2. Access o da a
• Access o Da ase s-as-a-Se ice (DaaS) (2025: Pa ially a ailable. Fully 2026+)
• Da a ca alogue (DaaS componen ) (2025: Pa ially a ailable. Fully 2026+)
• Use suppo (DaaS componen ) (2025)
• Da a pe mi managemen (DaaS componen ) (2026+)
• Da a ans e (DaaS componen ) (2026+)
• LUMI AIF da a en i onmen use guide (2025 pa ially a ailable, 2026+)
• S o age con ac (2025 pilo ing, 2026+)
• Da a space in eg a ion (2026+)
• Da a s eaming (2025 pilo ing, 2026+)
