1
Su ey
o Cus ome Needs
and HPC/AI Ma u i y
Sep embe 2025
Conduc ed by: LUMI AI Fac o y, 2025
License: CC BY-SA 4.0, e e o licenso “LUMI AI Fac o y”
Disclaime : This epo is o in o ma i e pu poses only.
Con ac : con ac @lumi-ai- ac o y.eu
2
Glossa y o Te ms
I em
Desc ip ion
AI
A i icial In elligence
De Ops
De elopmen Ope a ions
DL
Deep Lea ning
GDPR
Gene al Da a P o ec ion Regula ion
GPU
G aphic P ocessing Uni
HEI
Highe Educa ion Ins i u es
HW
Ha dwa e
HPC
High-Pe o mance Compu ing
Hugging Face
AI model hub & lib a y
IoT
In e ne o Things
IP
In ellec ual P ope y
LUMI AIF
LUMI AI Fac o y
LLM
La ge Language Models
ML
Machine Lea ning
NLP
Na u al Language P ocessing
PyTo ch
Flexible deep lea ning amewo k
RL
Rein o cemen Lea ning
ROI
Re u n on In es men
SME
Small and Medium-Sized En e p ises
Tenso Flow
Scalable machine lea ning pla o m
Type o m
Tool o conduc ion online su eys
VR
Visual Recogni ion
WP
Wo k Package
3
Table o Con en s
1. Execu i e Summa y ...................................................................................................... 4
2. Responses .................................................................................................................... 5
2.1 SECTION I: O ganisa ion P o ile 5
2.2 SECTION II. Cu en Use and Ma u i y 8
2.3 SECTION III. Technical Needs and Da a Managemen 14
2.4 SECTION IV. Skills and Suppo Needs 19
2.5 SECTION V. Collabo a ion and Ou look 22
3. Conclusion ................................................................................................................. 28
3.1 S a -Ups and SMEs 28
3.2 La ge En e p ises and Resea ch Ins i u ions 29
3.3 C oss-Cu ing Needs Highligh ed in Open Responses 29
4
1. Execu i e Summa y
A cus ome needs and ma u i y su ey was designed and launched o collec da a on he needs o cu en ,
p e ious, and po en ial cus ome s o he LUMI AI Fac o y (LUMI AIF) se ices. The da a o be collec ed
in ol ed cu en s a e o High-Pe o mance Compu ing (HPC) and A i icial In elligence (AI) adop ion,
echnical needs, ba ie s, and collabo a ion ou look. F om mid-June o mid-Augus 2025, he su ey was
dis ibu ed among he LUMI AIF conso ium coun ies Finland, No way, Denma k, Es onia, Poland, and
he Czech Republic, esul ing in 40 esponses, incl. esponses om hese and o he Eu opean coun ies.
The esponden s ep esen ed a b oad mix o s a -ups (30.8%), SMEs (30.8%), highe educa ion
ins i u ions (HEI) (15.4%), la ge en e p ises (15.4%), and esea ch ins i u es (7.7%), p o iding a snapsho
o cu en p ac ice and immedia e ope a ional equi emen s in he apidly e ol ing HPC/AI landscape.
The con en is s uc u ed a ound su ey design, esponden s’ o ganiza ional and echnical p o iles, and
hei day- o-day expe ience wi h AI and HPC sys ems. Key ou comes include s ong demand o scalable
GPU access, uni ied in as uc u e suppo ing hyb id and cloud wo kloads, echnical onboa ding, and
specialis suppo o in eg a ing new ha dwa e o pla o ms. Access o cu a ed Eu opean da a esou ces,
s aigh o wa d unding pa hways, and ongoing wo k o ce aining eme ged as dis inc needs,
pa icula ly om o ganiza ions scaling beyond pilo p ojec s.
Key indings e eal ha AI adop ion among he pa icipan s is highly ma u e, wi h 92.3% o o ganiza ions
al eady using AI o machine lea ning (ML) echnologies, compa ed o 53.8% using HPC. Mo eo e , 28.2%
o o ganiza ions plan o in oduce HPC soon, demons a ing a g owing in e es . O ganiza ions sel -
assess hei ma u i y le els as p edominan ly ad anced (50%) o in e media e (35%), while only 15%
iden i y hemsel es as beginne s. The main d i e s o adop ion include e iciency gains (72.5%),
inno a ion (45%), and cos educ ion. Se e al ba ie s cons ain b oade in eg a ion, mos no ably
budge cons ain s (65%), ollowed by da a p i acy issues (25%) and unclea business alue (20%).
In e ms o echnical needs, GPU compu ing s ands ou as he mos c i ical equi emen (79.5%),
alongside AI in e ence se ices, da a cleaning, and as s o age. Compu a ional asks ocus on AI model
aining (85%), simula ions and digi al wins (47.5%), and la ge-scale da a analysis (37.5%). Scalabili y,
as esul s, and eliabili y a e he leading pe o mance equi emen s. Skills and wo k o ce eadiness
show p omising ends: 69.2% o o ganiza ions al eady employ dedica ed AI/HPC s a , and 61.5% a e
ac i ely in es ing in wo k o ce upskilling. None heless, gaps emain in HPC ope a ions (57.1%), AI/ML
expe ise (40%), cloud in as uc u e (40%), and da a science (37.1%). Collabo a ion is eme ging bu
une en. While 53.8% o o ganiza ions pa ne wi h ex e nal ac o s, awa eness and use o compe ence
cen es and EU- unded HPC p og ammes emain limi ed. Only 33.3% o esponden s epo ed
pa icipa ion in publicly unded p og ammes, while 41% had no engaged in any.
O ganiza ions highligh op p io i ies in in as uc u e access (cloud, GPUs, scalable HPC), la ge model
aining and ine- uning, comme cializa ion o AI-based p oduc s, and domain-speci ic applica ions (e.g.,
ag icul u e, manu ac u ing, ansla ion). S a egic goals include wo k o ce de elopmen , demons a ing
ROI, and making socie al con ibu ions, such as add essing labou ma ke challenges. The e idence
highligh s he impo ance o a ge ed in as uc u e in es men s, communi y- ocused aining, and
lexible, pa ne ship-o ien ed app oaches o ensu e b oad and sus ainable AI/HPC adop ion ac oss
Eu ope’s esea ch and indus ial sec o s.
5
2. Responses
The ollowing sec ion p o ides he esponses o he su ey pa icipan s.
2.1 SECTION I: O ganisa ion P o ile
a. Name o O ganisa ion
Con iden ial
b. Wha is he ype o you o ganiza ion?
The 40 o ganiza ions ep esen ed in he su ey we e p ima ily s a -ups (30.8%) and SMEs (30.8%),
ollowed by highe educa ion ins i u es (15.4%) and la ge en e p ises (15.4%). A smalle sha e was
ep esen ed by esea ch ins i u es (7.7%); no esponden s om public adminis a ion ook pa .
6
c. Wha is you o ganiza ion’s main business sec o ?
In e ms o he o ganiza ion’s main business sec o , he domain o AI p e ailed (42.1%), ollowed by
esea ch and de elopmen (13.2%) and manu ac u ing (10.5%). O he domains, such as so wa e
de elopmen , enginee ing, manu ac u ing, ene gy, li e sciences, and space applica ions, we e
ep esen ed by ei he ewe o single esponden s; o he domains by none.
O he :
• Quan um compu ing
• Compe i i e & ma ke in elligence
• S ainless s eel hose p oduc ion
• Languages and audio isual ansla ion
• Heal hca e
7
d. How many employees does you o ganiza ion ha e?
As o he pa icipan s’ size and s uc u e, jus o e hal o he pa icipan s ep esen ed small
o ganiza ions wi h ewe han 50 employees (55%), while la ge o ganiza ions wi h mo e han 1,000
employees accoun ed o 22.5%. Ca ego ies o o ganiza ions wi h 50-250 employees (12.5%) and 251-
1,000 employees (10%) achie ed a lowe pa icipa ion.
e. In which coun y is you o ganiza ion loca ed?
The esponse a e pe coun y is laid ou in he ollowing able. Finnish pa icipan s a e ep esen ed wi h
he highes esponse a e (35.0%), ollowed by Czechia (22.5%), Denma k (17.5%), and Poland 12.5%.
One esponse came om No way, and no esponse came om Es onia. Answe s om non-conso ium
coun ies we e ecei ed om Belgium, Tu key, Sweden, and Ge many.
Coun y
Numbe o Pa icipan s
Pe cen age
Finland
14
35.0%
Czechia
9
22.5%
Denma k
7
17.5%
Poland
5
12.5%
No way
1
2.5%
Sweden
1
2.5%
Ge many
1
2.5%
Tu key
1
2.5%
Belgium
1
2.5%
Es onia
0
0.0%
To al
40
. In which o he coun ies does you o ganiza ion ope a e?
In 72.5 % o cases, pa icipa ing o ganiza ions no ed ha hey had ope a ions ab oad, wi h Ge many,
F ance, Spain, Poland, Slo akia, Hunga y, Czechia, Swi ze land, Sweden, No way, Denma k, and he UK
being he mos common ones. Responden s o en g ouped hese unde EU o Eu ope, some imes lis ing
speci ic Cen al & Eas e n Eu opean coun ies alongside Wes e n Eu ope. Apa om ha , No h
Ame ica and he Middle Eas we e also men ioned, indica ing a global each o he companies’ ac i i ies.
8
2.2 SECTION II. Cu en Use and Ma u i y
a. Does you o ganiza ion cu en ly use HPC esou ces?
This ques ion e ealed ha 53.8% o pa icipa ing o ganiza ions ha e al eady implemen ed HPC in o
hei ope a ions, wi h an addi ional 28.2% planning o do so. Only 17.9% o esponden s s a ed ha hey
do no use HPC a he cu en s a e. These s a is ics show a ce ain le el o expe ience among he HPC
communi y membe s and willingness o explo e his ield.
I you o ganiza ion cu en ly uses o plans o use HPC esou ces, please speci y he use case.
• Time se ies o ecas using suppo ec o eg ession
• Pushing he limi s o quan um sup emacy
• Medical image p ocessing
• We mainly u ilizes High-Pe o mance Compu ing (HPC) esou ces o suppo he de elopmen and
alida ion o ad anced echnologies o ehicles (bu no only), pa icula ly in he con ex o so wa e-
de ined ehicles like: Ad anced D i e -Assis ance Sys ems (ADAS) de elopmen , So wa e-de ined
ehicle de elopmen , AI and Machine Lea ning in eg a ion, Tes ing and Valida ion, Senso Fusion,
Scalabili y, and In e changeabili y.
• LLM in e encing
• De elopmen o ma ke ing echnology using AI
• Building p edic i e models, using DL
• We use HPC clus e s (IT4Inno a ions, Eu oHPC) o GPU-accele a ed AI c op pheno yping, CFD o
e ical a ms, massi e pa ame ic LED/wa e ing op imiza ion, and LCA simula ions; plans genome-
scale analysis nex .
• Planning in p oduc ion
• T aining ounda ion models
• hea y analysis and model building
• LLM, AI agen , da a analy ics
• Can' ell abou i ; i 's a sec e
• La ge Language Models
• Va ious esea ch p ojec s mainly ocused on LLMs and physical simula ions
9
• T aining AI & unning AI models
• Inno a ion o demanding compu a ions o clien cases and in-house s udy.
• We ope a e HPC esou ces o p o ide ou se ices o ou cus ome s.
• AI aining, e alua ion, and de elopmen
• We' e looking o ways o suppo ansla o s' expe ise
• We ha e ou own IT depa men , and we un mul iple applica ions o in e nal pu poses
• AI/ML model aining
• T aining LLMs, Fine- uning GenAI, GRPO aining
• Model uning, model aining, ein o cemen lea ning
• Cu en ly, we a e mos ly ine uning GRPO, LLMs on a smalle numbe o ou own GPU ca ds, as we keep
scaling, ou own GPUs will no su ice
• Specialized model aining, unning open sou ce LLM
• CFD
• HPC o esea ch and educa ion
• Fo esea ch
• Compu a ional chemis y; Ma e ial Science; Bioin o ma ics, e c.
b. Does you o ganiza ion use AI o machine lea ning echnologies?
The u iliza ion o AI and machine lea ning epo ed by he pa icipan s e ealed a whopping sha e o
92.3% o companies ha al eady ha e engaged wi h AI o machine lea ning in hei ope a ions, and only
7.7 % ha ha e no . These s a is ics con i m he ma u i y o AI as compa ed o HPC.
I you o ganiza ion cu en ly uses o plans o use AI o machine lea ning echnologies, please speci y
he use case.
• Time se ies o ecas using suppo ec o eg ession
• Image in e p e a ion
• Al eady answe ed he ques ions be o e, mainly o ADAS
• Pe sonaliza ion and Lea ning om Use Da a. The app “lea ns” use p e e ences: -which pic og ams
a e chosen mos o en, - he ypical ocabula y o he amily/ca egi e , - he child’s speci ic
communica ion habi s.
16
c. Wha compu a ional asks a e ele an o you o ganiza ion? (selec up o h ee)
The la ge majo i y o o ganiza ions ega ded he aining o AI models as a ele an compu a ional ask
(85.0%). Tasks, such as a ious simula ions and digi al wins (47.5%) and la ge-scale business da a
analysis (37.5%) o IoT and senso da a (35.0%) ollowed.
O he :
• Business p oblem op imiza ion; sea ch o classical limi s on quan um sup emacy
• ma ke ing use cases, eal- ime g aphic gene a ion based on consume da a
• Voice- o- oice
• Running LLM o achie e as much independence as on big LLMs
17
d. Wha a e you pe o mance equi emen s? (Selec up o h ee)
Pe o mance equi emen s, acco ding o he su ey, a e anked in e ms o impo ance as ollows:
scalabili y (65.0%), as esul s (62.5%), eliabili y (52.5%), high h oughpu (42.5%), and da a secu i y
(37.5%).
O he :
• Speed - esponse ime
• Use - iendliness
• Suppo o mode n AI so wa e s ack (Py o ch 2.7, lash-a en ion2, i on) 2. ease/uni ied
deploymen (k8s jobs ins ead o slu m) 3. GPUs wi h 80+GB o VRAM o i 32k con ex size o la ge
du ing model aining ( he e is a lowe bound o sha ding he model laye s). 4 Ha dwa e FP8 suppo
18
e. How o en do you expec o use HPC/AI esou ces?
The su ey concludes ha he majo i y o o ganisa ions engaged wi h HPC and AI expec o use hese
esou ces and ools daily (66.7%), and a u he 12.8% o o ganiza ions in end o do so weekly.
. Wha a e you main conce ns ega ding da a secu i y? (Selec up o h ee)
The mos common conce n ega ding da a secu i y was s a ed o be sensi i e da a p o ec ion (62.5%).
Secu e da a s o age and ans e was men ioned by 50 % o esponden s and mus be conside ed a
ele an conce n. Compliance isk (35.0%) and cloud se ice anspa ency (32.5%) can be anked
seconda y conce ns, while IP con iden iali y is a e ia y conce n. No speci ic conce ns we e s a ed by 5%
o esponden s.
19
2.4 SECTION IV. Skills and Suppo Needs
a. Do you ha e dedica ed AI/ML o HPC pe sonnel?
Almos h ee-qua e s (69.2%) o he o ganiza ions al eady ha e a dedica ed AI/ML o HPC pe sonnel,
and ano he 17.9% o o ganiza ions in end o in eg a e such pe sonnel. Tha shows a su p isingly good
le el o eadiness and ma u i y. Mo e de ailed analysis shows ha all esea ch ins i u ions ha e dedica ed
pe sonnel. All s a -ups ei he ha e o a e planning o in eg a e dedica ed pe sonnel. While SMEs show
a s ong dedica ion o in eg a ing pe sonnel, la ge en e p ises show a lowe commi men .
20
b. Wha skills a e lacking mos in you o ganiza ion?
O ganiza ions ha e epo ed ha se e al skills a e missing, which may limi u he HPC and AI adop ion.
The mos se ious gap is seen in he ield o HPC ope a ions (57.1%), ollowed by AI and ML expe ise
(40.0%), cloud in as uc u e 40.0%, and da a science (37.1%).
O he :
• When wo king wi h esea ch ins i u ions o co po a e pa ne s, he bigges skills gap we see is in
deploying ou ha dwa e: pa ne IT eams o en lack he ne wo k and secu i y know-how o open
he igh po s o con igu e VPNs, he de -ops expe ience o egis e applica ions and manage API
keys o ce i ica es, and he da abase skills. F om expe ience, many o he IoT SMEs ha e big
ouble he e.
• De Ops enginee ing, web de elopmen , o b ing a wide a ay o cloud se ices o ou cus ome s.
• High-quali y sales
c. A e you in es ing in wo k o ce upskilling o AI/HPC?
Mos o ganiza ions epo ed hey we e in es ing in wo k o ce upskilling o AI and HPC (61.5%),
illus a ing a dedica ion o engage e en u he wi h AI and HPC. Only 17.9% o o ganiza ions s a ed hey
made no such in es men s, while he emaining 20.5% o o ganiza ions in end o in es .
21
d. Wha ype o ex e nal suppo would be mos aluable o you o ganisa ion? (Selec up o
h ee)
Responden s claimed ha unding suppo would be mos alued (55.0%), ollowed by HPC
in as uc u e access (52.5%) and echnical consul ing (42.5%). Funding suppo was pa icula ly
men ioned by SMEs, s a -ups, and HEIs, bu no by la ge en e p ises and esea ch o ganisa ions. La ge
en e p ises a e pa icula ly in need o echnical consul ing. Fo Resea ch Ins i u es, access o HPC
in as uc u e is mos aluable. Skills aining and egula o y guidance we e men ioned by jus unde
one- hi d o he o ganiza ions.
22
2.5 SECTION V. Collabo a ion and Ou look
a. A e you collabo a ing wi h ex e nal pa ne s on AI/HPC?
O e hal o he o ganiza ions (53.8%) epo ed collabo a ing wi h ex e nal pa ne s on AI and HPC,
indica ing ha awa eness o compe ence cen es and o he ex e nal ac o s suppo ing he in eg a ion o
hese echnologies emains limi ed and should be u he s eng hened.
I you selec ed 'Yes' please speci y he mode o collabo a ion wi h ex e nal pa ne s on AI/HPC.
• Deucalion
• Collabo a ion wi h uni e si ies
• MS Azu e
• Ge ing esou ces (inluding GPU/AI) wi h L2/L3 suppo
• Cy one AGH
• IT De , building solu ions o ou needs
• AI s a ups
• HPC access
• In ol emen o a ious pa ies du ing esea ch p ojec s
• Google o S a ups Pa ne .
• We igh ly coope a e wi h ou la ges cus ome s o help hem be success ul in aining and unning
AI models a scale.
• P ojec s
• Open esea ch
• We a e looking in o his wi h a b oadcas ing company. We a e hei p e e ed ansla ion p o ide .
• Consul ing, human esou ces
• We' e collabo a ing wi h Czech uni e si ies and hei supe compu e cen e s (e.g., CEITEC)
• Eu opean p ojec s
• N idia (pa ne , GPUs,...), CVUT (access o Os a a clus e )
• Resea ch
• Resea ch conso ia
23
b. Wha a e you goals o adop ing AI/HPC? (Selec up o 3)
The mo i a ion o he AI and HPC adop ion is clea ; he majo i y o o ganiza ions seek an inc ease in
e iciency (72.5%). Inc easing e iciency is he mos impo an goal o HEIs, SME’s and s a -ups. The
second mos ecu en goal s a ed is inno a ion (45.0%), pa icula ly d i en by HEIs and Resea ch
o ganisa ions. A leas one qua e o companies also highligh ed cos educ ion, he launch o new
se ices, and cus ome se ice imp o emen . La ge en e p ises pa icula ly s ess he aim o cos
educ ion and an inc ease in cus ome se ice.
O he :
• Inc ease he quali y o he ansla ions
• be e condi ions han buying GPUs o p o isioning GPU IaaS on big ech clouds
24
c. Wha is you ime ho izon o you nex AI/HPC p ojec ?
The p e ailing long- e m ho izon (75.0%) o he o ganiza ions’ nex AI o HPC p ojec signals ha hey
adop mo e o a s a egic decision-making app oach. Only 5.0 % s a e AI/HPC is a one-o p ojec .
d. Ha e you pa icipa ed in any publicly unded AI/HPC p og ams?
The sha e o o ganiza ions ha ha e no pa icipa ed in any publicly unded AI/HPC p og ammes (41.0%)
is highe han ha o he ones ha ha e (33.3%). Tha con i ms some lack o awa eness o such
p og ammes among he o ganiza ions po en ially in e es ed in AI and HPC deploymen . The high
p opo ion o esponden s s a ing “no su e” may be e lec ed in di icul ies unde s anding he ques ion.
I you selec ed 'Yes' please speci y which publicly unded AI/HPC p og ams you ha e pa icipa ed in.
• Eu oHPC Deucalion
• Cy one g an o s a ups
• PLG id
• EDIH
25
• LUMI es pe iod
• Danish na ional unded
• No way's Oli ia pilo es ing, NTNU's IDUN e en s
• Eu oHPC g an o Lumi-G
• HORIZON
• EDIH Os a a
• Resea ch Council o Finland, Eu oHPC
• PRACE, Eu oHPC, LUMI....
e. Wha a e you op p io i ies o AI/HPC in he nex 12–24 mon hs?
The collec ed s a emen s clus e in o six main hemes: in as uc u e and compu e needs (cloud, GPUs,
HPC in eg a ion, debugging issues), LLM and model aining (scaling, ine- uning, in e encing, p ojec
comple ion), and p oduc iza ion and comme cializa ion (AI-based p oduc s, domain-speci ic sys ems,
s a ups, dis ibu o oles). Addi ional hemes include esea ch and wo k o ce de elopmen ( ained
pe sonnel, esea che in ol emen , scalable a chi ec u es), domain-speci ic applica ions (ag icul u e,
manu ac u ing, ansla ion, op imiza ion), and s a egic goals (planning, ROI, pilo ing, socie al impac ,
such as labou ma ke challenges). Toge he , hese codes e eal a balance be ween echnical
equi emen s, applied inno a ion, and b oade economic o socie al objec i es.
• Launching new AI-based p oduc s & se ices, enhancing exis ing AI applica ions, wo k o ce
upskilling
• In ol emen o esea che s and le e aging he ull po en ial o HPC
• Mo e mo e wo kloads o GPU esou ces, inc ease he adop ing AI wo kloads o HPC
• Model aining, model ine- uning, model specializa ion
• T ain LLM model, launch an online se ice based on you cus om own LLM
• To ge an o e all unde s anding o he possible bene i s s. cos s s. esou ces needed
• Building ou new ma ech p oduc s based on AI
• In as uc u e compe ence
• Scaling compu a ions, model aining/ ine uning, ML ope a ions
• Help companies in Eu ope wi h ou so e eign cloud se ices o AI
• In eg a ion o AI/HPC in public cloud
• Mo e GPUs
• Building a scalable and cos -e icien pipeline. Due o using mul iple pipelines o mul iple use-cases,
he o e a ching a chi ec u e is he mos impo an opic.
• T ain and in e he high- h oughpu la ge ision model o my b and-new ision ounda ion model,
which I am cu en ly de eloping.
• Use la ge esou ces o demanding compu a ion and op imiza ion asks. This way, we can p o ide
he bes solu ions o clien s and en ich he inno a ion o he company's own p oduc s, which will
b ing long- e m alue. (Howe e , we canno eally in es a lo in his o he han ime, and Business
Finland only unds la ge companies doing andom useless "EU p ojec s" ha hey, o nobody,
needs, e c. Jus my expe ience.)
• Any hing how o u ilize MI250 machines bes way possible. The AI lib a ies de elop oo as o keep
up wi h such old ha dwa e. How o keep he la es lib a ies and p og ammes lis ed. How o debug
uns in LUMI, while using sba ch sc ip s and Py hon, many e o messages a e no clea a all.
• Sol e Finland’s labou ma ke p oblems
