A Bluep in o Mul ina ional Ad anced AI
De elopmen
by Ad ien Abecassis, Jona han Ba y, Ima Bello, Yoshua Bengio,
An onin Be geaud, Yann Bonne , Philipp Hacke , Ben Ha ack, Sophia Ha z,
Joachim Henkel, Holge H. Hoos, Ki Ki amu a, Ranji Lall, Yann Lechelle,
Cons ance de Leusse, Cha les Ma ine , Nicolas Miailhe, Julia C. Mo se,
Maximilian Negele, Kyung Ryul Pa k, Mi o Pluckebaum, Mu ielle Popa-Fab e,
Benjamin P ud’homme, Yohann Ralle, Ma k Robinson, Cha bel-Raphael
Sege ie, José-Ignacio To eblanca, Lucia Velasco, K. VijayRagha an
No embe 2025
This memo is au ho ed by:
Ad ien Abecassis — Pa is Peace Fo um
Jona han Ba y — Mila – Quebec A i icial In elligence Ins i u e
Ima Bello — Fu u e o Li e Ins i u e
Yoshua Bengio — Mila – Quebec A i icial In elligence Ins i u e
An onin Be geaud — HEC Pa is
Yann Bonne — Pa is Peace Fo um
Philipp Hacke — Eu opean Uni e si y Viad ina
Ben Ha ack — Ox o d Ma in AI Go e nance Ini ia i e
Sophia Ha z — Uppsala Uni e si y
Joachim Henkel — Technische Uni e si ä München (TUM)
Holge H. Hoos — RWTH Aachen Uni e si y, Leiden Uni e si y
Ki Ki amu a — Independen expe
Ranji Lall — Uni e si y o Ox o d
Yann Lechelle — P obabl
Cons ance de Leusse — AI & Socie y Ins i u e (ENS - PSL)
Cha les Ma ine * — Ox o d Ma in AI Go e nance Ini ia i e, Cen e pou la Sécu i é de l’IA (CeSIA)
Nicolas Miailhe — AI Sa e y Connec , PRISM E al
Julia C. Mo se — Ox o d Ma in AI Go e nance Ini ia i e, Uni e si y o Cali o nia, San a Ba ba a
Maximilian Negele — Ox o d Ma in AI Go e nance Ini ia i e
Kyung Ryul Pa k — Ko ea Ad anced Ins i u e o Science and Technology
Mi o Pluckebaum — Ox o d Ma in AI Go e nance Ini ia i e
Mu ielle Popa-Fab e — AI & Socie y Ins i u e (ENS - PSL)
Benjamin P ud’homme — Mila – Quebec A i icial In elligence Ins i u e
Yohann Ralle — The Fu u e Socie y
Ma k Robinson — Ox o d Ma in AI Go e nance Ini ia i e
Cha bel-Raphael Sege ie — Cen e pou la Sécu i é de l’IA (CeSIA)
José-Ignacio To eblanca — Eu opean Council on Fo eign Rela ions
Lucia Velasco — Ox o d Ma in AI Go e nance Ini ia i e
K. VijayRagha an — Fo me P incipal Scien i ic Ad iso o he Go e nmen o India
*Lead d a e
The au ho s a e g a e ul o Toby O d, Milo Rignell, Philip Fox, Lily S elling, and Ra ael Ande sson
Lipcsey o hei commen s and sugges ions.
A Bluep in o Mul ina ional Ad anced AI De elopmen
Thesis: an in e na ional ad anced AI esea ch & de elopmen pa ne ship o AI b idge powe s1(1) can
easibly p oduce on ie AI models; and (2) is essen ial o sa egua ding he so e eign y, democ a ic al-
ues, economic compe i i eness and g ow h, echnical inno a ion, and na ional secu i y o hese b idge
powe s a es.
Execu i e Summa y
The global ace o de elop ad anced AI has en e ed a new phase ma ked by s agge ing in es men s,
apid echnical b eak h oughs, and in ensi ying geopoli ical compe i ion. The Uni ed S a es now
con ols app oxima ely 75% o global AI compu e capaci y, China 15%, and he EU 5%.2This
concen a ion o compu e, alongside concen a ions o AI de elopmen alen , da a, and AI model
owne ship sugges s ha mid-sized economies likely ace insu moun able ba ie s o independen
on ie AI de elopmen .
A he same ime, economic, cul u al, and secu i y in as uc u es a e coming o ely e e mo e on
on ie models. S a es ha a e unable o de elop hei own on ie models o access he compu -
ing ha dwa e equi ed o ain hem will ha e o choose be ween dependency and weakness:
•Dependency: i s a es adop U.S. o Chinese AI sys ems, hese on ie AI s a es3can hen exploi
hei p i ileged posi ion in ways ha ha m dependen s a es, o example h ough da a he ,
se ice es ic ions, selec i ely wi hholding on ie capabili ies, embedding alues in ounda ion
models, and un a o able e ms o ade.
•Weakness: i , on he o he hand, s a es limi hei adop ion o on ie sys ems o a oid depen-
dency, on ie AI s a es may achie e b eak h ough capabili ies—in economic p oduc i i y, in
scien i ic disco e y, in mili a y ope a ions— ha c ea e widening gaps in economic and mili a y
capabili ies.
Ye , mid-sized economies a e also AI b idge powe s, possessing subs an ial AI de elopmen capa-
bili ies and esou ces ha , i combined, would allow hem o challenge he s a us quo. By wo king
oge he and s a egically choosing hei AI de elopmen app oaches, AI b idge powe s can de-
elop compe i i e on ie models:
•Fi s , pooled compu ing in as uc u e can suppo on ie -scale de elopmen . Coo dina ed de-
ploymen o exis ing, planned, and wi hin- each Eu opean and o he b idge powe AI compu e
capaci y is likely o p o ide su icien compu a ional esou ces o p oduce on ie AI models in
he nex ew yea s, al hough signi ican ly mo e in es men s a e p obably equi ed o keep up
wi h he mo ing on ie .
•Second, a signi ican po ion o op AI alen has ies o AI b idge powe coun ies. 87 o he 100
mos -ci ed AI esea che s o igina e om o cu en ly wo k in coun ies ou side he Uni ed S a es
and China. B idge powe s could “call home” leading esea che s i hey had an inspi ing ision
backed by su icien esou ces and an e hical de elopmen pa h.
•Thi d, while mos o he da a used o ain on ie models is al eady public, b idge powe s could
pool domain-speci ic da a and esou ces o da a cleaning and expe labeling e o s.
•Fou h, b idge powe s should make s a egic, on ie de elopmen be s, le e aging sha ed dig-
i al in as uc u es (e.g. pooled p e- aining) and R&D e o s o ocus on p omising a eas ha
do no ely on ma ching scale elsewhe e, in o de o each and hen ack o e en su pass he AI
on ie .
•Fi h, building eliable AI ep esen s an unme ma ke need whe e b idge powe s ha e s uc u al
ad an ages. High- alue indus ies equi e con ol o e AI ools and con idence in hei eliabili y
be o e deploying hem a scale. B idge powe s can ac as us ed b oke s by le e aging s ong
3
A Bluep in o Mul ina ional Ad anced AI De elopmen
da a p o ec ion egimes, obus ule o law, and esponsi e go e nance o speed up sus ainable
adop ion.
A mul ina ional pa ne ship could enable membe s o p ese e so e eign y, ha e mo e weigh in
shaping global AI go e nance, and lead h ough e hical s ewa dship. Some p eceden s o simila
mul ila e al p ojec s exis h ough CERN o Ai bus4, and he capabili ies exis h ough collec i e
ac ion. The ques ion is hen whe he b idge powe s will ac decisi ely be o e dependencies deepen
and he bipola s uc u e consolida es.
4
A Bluep in o Mul ina ional Ad anced AI De elopmen
Con en s
A. The S a egic Downside o Bipola F on ie AI 6
B. A Mul ina ional F on ie AI Pa ne ship 7
C. Feasibili y and Timeliness o a Mul ina ional Pa ne ship 10
D. Re e sing S a egic Vulne abili ies: Bene i s o Membe Coun ies 11
5
A Bluep in o Mul ina ional Ad anced AI De elopmen
In oduc ion
The s a egic implica ions o concen a ed AI in es men a e now undeniable. Wi h U.S. companies
expec ed o spend o e $300 billion and China nea ly $100 billion on AI in as uc u e in 2025,
he compu a ional di ide has become a chasm: Ame ican con ol o app oxima ely 75% o global
AI compu e capaci y, combined wi h alen , da a, and AI model owne ship concen a ions, c ea es
ba ie s ha indi idual mid-sized economies canno o e come h ough na ional e o s alone. The
ques ion is no longe whe he b idge powe s ace disad an ages, bu whe he hey will ac collec-
i ely be o e hose disad an ages ha den in o s uc u al cons ain s.
G7 minus US
79,000
China
231,000
USA
1,079,000
Figu e 1: Global Dis ibu ion o AI Compu e in Oc obe 2025 (H100 equi alen s)5
Howe e , AI b idge powe s ha e subs an ial AI de elopmen capabili ies o hei own. By pa ne ing
s a egically and coo dina ing hei in es men s, compu e, alen , da a, and go e nance, hey can
pa icipa e in he on ie o AI de elopmen and sa egua d hei so e eign y and alues.
A. The S a egic Downside o Bipola F on ie AI
AI is poised o become he de ining asse o he 21s cen u y. F on ie AI sys ems al eady demon-
s a e apidly imp o ing abs ac hinking and easoning skills, and ma ch o exceed human expe s
ac oss wide- anging domains. They a e inc easingly enmeshed wi h economic, cul u al, scien i ic,
and secu i y p ocesses, di ec ly and h ough my iad applica ions,6bu his may jus be he ip o he
icebe g i AI ad ances con inue acco ding o cu en ends. Many expe s belie e ha supe human
gene al AI will be achie ed wi hin he nex 5–10 yea s. While such claims a e highly unce ain,
he pace o de elopmen o e he pas wo yea s sugges s ha human-le el AI is plausible in he
nea u u e.
AI ad ances a e likely o unde mine he so e eign y o non- on ie s a es. I s a es canno indepen-
den ly de elop, ain, o modi y on ie AI sys ems, hey may ind hemsel es aced wi h a choice
be ween wo di e en o ms o ulne abili y: ei he buy AI sys ems om on ie s a es and be-
come s uc u ally dependen on hem (dependency), o be le behind by on ie s a es al oge he
(weakness).7
In a dependency scena io, non- on ie s a es a e dependen on on ie s a es’ AI sys ems. This
allows on ie AI s a es o exploi he powe di e en ial be ween hem and he s a es ha depend
upon hem. Fo example, on ie s a es could copy sensi i e da a exchanged wi h hei AI models
and use i o economic o poli ical ad an age.8They could also selec i ely modi y AI access,
h ea ening se ice deg ada ion o e en a ull cu -o ;9and hei alues and design choices will be
embedded in ounda ion models, impac ing all downs eam applica ions.10 Dependency may also
make e ms o ade mo e un a o able o non- on ie s a es.
E o s o limi adop ion in o de o a oid dependency may lead ins ead o weakness: a wo ld whe e
on ie s a es use AI o achie e b eak h oughs in economic p oduc i i y, scien i ic disco e y, and
mili a y ope a ions, allowing hem o build ou asymme ic capabili ies (such as AI-enabled cybe
ope a ions ha make con en ional de enses obsole e11) o au oma e co e economic and secu i y
6
A Bluep in o Mul ina ional Ad anced AI De elopmen
unc ions. Such gaps could widen d ama ically as AI models a e used o u he accele a e hei
own de elopmen .
B idge powe s hus ace a s a egic choice be ween undamen ally di e en app oaches o access-
ing on ie AI. Each s a egy implies ha ing access o su icien compu ing powe 12, and in ol es
dis inc ade-o s be ween capabili y, so e eign y, and cos :
Table 1: S a egic Al e na i es o B idge Powe Access o F on ie AI13
S a egy F on ie -compe i i e? So e eign y p o ec ed? Financially iable?
Impo closed
models
-Yes, bu se e al mon hs
behind; access can be
es ic ed
qVulne able o se ice denial,
licensing es ic ions
¥Low up on cos ; high
ongoing dependency
Adop open
models
qBehind on ie ; eleased
open models lag mo e han
6 mon hs behind closed
models, which could g ow
due o na ional secu i y
es ic ions
-Reduces bu doesn’
elimina e o eign
dependencies
¥Low di ec cos
Na ional
champions
-Behind on ie ; agmen ed
e o s
-Pa ial so e eign y; may
equi e signi ican o eign
owne ship
qExpensi e pe coun y;
duplica es in as uc u e
Mul ina ional
pa ne ship
¥Achie able h ough pooled
esou ces and s a egic
de elopmen choices
¥Rein o ced so e eign y
h ough collec i e
go e nance, gua an eed AI
access, imp o ed domes ic
AI ecosys ems
¥Sha ed cos s; economies o
scale
As illus a ed by Table 1’s i s h ee s a egies, impo ing, adop ing, o domes ically de eloping
on ie AI each in ol e unaccep able ade-o s: a mul ina ional pa ne ship is he only iable pa h
o achie ing on ie compe i i eness while p ese ing so e eign y a manageable cos . Al hough
hey di e in many ways, AI b idge powe s ha e common in e es s in sa egua ding so e eign y and
p o ec ing hei alues and way o li e. These common in e es s imply ha he e is po en ial o
in e na ional coope a ion on on ie AI de elopmen .
B. A Mul ina ional F on ie AI Pa ne ship
A mul ina ional on ie AI pa ne ship o e s a iable and scalable s a egic op ion. A join AI b idge
powe pa ne ship has a much g ea e chance o eaching he echnological on ie han indi id-
ual b idge powe s o na ional champions ope a ing alone. Ad an ages include pooled compu ing
in as uc u e, alen , and da a.
Ad an age 1: Pooled Compu ing In as uc u e
The logic o compu e pooling is oo ed in he undamen al economics o AI de elopmen . AI in-
e ence, o use, in ol es cos s ha a e decen alized and scale p opo ionally wi h he numbe o
use s in each coun y. In con as , AI aining and de elopmen cos s a e un ela ed o he numbe o
use s in a coun y. Those cos s equi e a massi e, concen a ed expendi u e o compu e esou ces.
This economic eali y means ha by pooling esou ces o co e he high, ixed cos s o aining a
on ie model, a g oup o b idge powe s can achie e a le el o capabili y and compe i i eness ha
no single membe could each independen ly.
While indi idual b idge powe s canno ma ch he scale o ini ia i es like OpenAI’s S a ga e p ojec
(>$100B/yea announced), g oups o b idge powe s collec i ely possess signi ican da acen e ca-
paci y.14 F on ie AI de elopmen cos s may exceed se e al billions pe model by 2028 (see Fig-
7
A Bluep in o Mul ina ional Ad anced AI De elopmen
u e 2), and he equi ed in as uc u e in es men s will equi e ens o billions.
Gi en his, he cos s o main aining a on ie AI p og am unde exis ing pa adigms using only
domes ic compu e—building no jus one model, bu main aining on ie compe i i eness h ough
con inuous R&D15—a e a guably o an o de ha no single b idge powe can sus ain on i s own,
sho o shi ing in o a wa - ime ype o economy wi h poli ically un enable cos s.16
0
5B$
10B$
15B$
20B$
25B$
0.5B$ 1.1B$ 2.8B$
6.6B$ P ojec ed Cos
o Leading
T aining Run
3B$
5.7B$
10.8B$
20.6B$ P ojec ed Cos
o Leading
Compu e
Clus e
2025 2026 2027 2028
Figu e 2: P ojec ed Ha dwa e and T aining Cos s o S a e-o - he-A AI Sys ems (USD, Billions)17
Sou ces: Pilz e al., 2025;Co ie e al., 2024;Epoch AI, 2025.
The in as uc u e ounda ion is al eady aking shape: in he EU, AI Fac o ies and public supe -
compu e s like Jupi e (Ge many, 24,000 GPUs, ope a ional) and Alice Recoque (F ance, exascale,
2026) a e deploying nea - e m capaci y, while in he mid- e m, i e Giga ac o ies will deli e
100,000+ specialized AI chips each by 2027.18 Coo dina ed h ough a mul ina ional pa ne ship,
hese asse s— alued a €20+ billion in EU commi men s alone—could suppo on ie de elop-
men a scale.
In he sho e m, i is nei he easible no necessa y o b idge powe s o ma ch he scale o U.S.
o Chinese in es men . Indeed, on ie compe i i eness de i es no jus om he shee amoun o
esou ces, bu how hey a e deployed, as demons a ed by DeepSeek and Mis al, which, h ough a -
chi ec u al inno a ions and s a egic ocus, achie ed pe o mance below bu compa able o on ie
models while spending less. By coo dina ing in es men s and adop ing no el app oaches, a mul i-
na ional pa ne ship can achie e ou comes a o nea he on ie wi hou ma ching he spending
in leading coun ies.19 Fu he mo e, a a ge ed ocus can alle ia e esou ce equi emen s:20 a he
han a emp ing o elimina e all dependencies ac oss he AI s ack simul aneously, he pa ne ship
could concen a e on de eloping on ie models wi h gene al capabili ies in easoning, planning,
us wo hiness, and mul imodal unde s anding. Downs eam applica ions can le e age hese, and
hey would be aluable ca ds o b idge powe s in u u e nego ia ions a ound global AI geopoli ics.
The mul ina ional pa ne ship should aim o minimize i s use o compu e loca ed in non-membe
s a es and compu e owned by o eign en i ies. Al hough i could begin by en ing compu e, his
ein o ces and c ea es subs an ial ulne abili ies o e ime.21 B idge powe in es men s in AI in-
as uc u e mus he e o e con inue and inc ease signi ican ly. I is impo an o highligh he key
ole wi hin he AI s ack o on ie AI algo i hms o capabili y ad ances: i capabili ies app oach o
su pass human-le el as cu en ends sugges , on ie AIs hemsel es will enable apid inno a ion
and imp o emen s o he o he componen s o he s ack.
Ad an age 2: Pooled AI Talen
AI b idge powe s ha e good p ospec s o ec ui ing and e aining signi ican AI alen h ough
p eexis ing ies o AI esea che s. E idence sugges s ha a subs an ial pool o eli e AI alen is
needed o each he on ie . Despi e an abundance o compu e and da a esou ces, companies like
8
A Bluep in o Mul ina ional Ad anced AI De elopmen
Me a appea o ha e s uggled o each he on ie in pa because hey lack a c i ical mass o
leading esea che s. A he same ime, o he 100 mos ci ed AI esea che s in he wo ld, 87 come
o iginally om coun ies o he han he U.S. and China, o a e cu en ly wo king in hem.22 Thus, a
g oup o b idge powe s migh “call home” a la ge p opo ion o e en a majo i y o leading esea che s,
d awn by subs an ial sala ies and he chance o wo k on ci iliza ion-de ining ques ions wi hin
collec i e go e nance s uc u es. The b ain d ain is e e sible when in as uc u e mee s alues
and a gal anizing, isible, and easible p ojec .
Ad an age 3: Pooled Da a
Da a pooling o e s a signi ican ad an age o pa ne ship membe s. While mos o he da a needed
o on ie AI R&D is publicly a ailable on he in e ne , he mos expensi e and ha d- o-ge
da a, which comes om expe labeling and anno a ion, is usually p i a ely gene a ed and owned.
B idge powe s can hus pool he cos o da a cleaning and labeling e o s, which companies in
on ie AI s a es und independen ly a massi e scale. Fu he mo e, some companies may be will-
ing o sha e hei da a wi h he pa ne ship h ough p e e en ial access o licensing a angemen s.
Combined wi h he p op ie a y da ase s o membe na ions, and in some cases hose sha ed by
hei domes ic indus ies, hese pooled da a esou ces would p o ide a scale and di e en ia ion
ad an age.
To a ac a di e se and du able pool o AI esea che s, a pa ne ship would equi e an inspi ing
ision, c edible and su icien capaci y and esou ces, and exci ing, leading-edge asks. B idge
powe s could join ly o e his combina ion by pooling hei alen and compu ing in as uc u e.
They could also o e access o subs an ial po ions o he wo ld’s AI da a. Toge he , hese h ee
esou ces—compu e, alen , and da a—comp ise he co e inpu s o on ie de elopmen .
Adop ing o eplica ing exis ing AI models a e no sil e bulle s: One migh hink ha b idge
powe s need no de elop on ie models hemsel es, bu could ins ead adop open-weigh
eleases (like Llama) o apidly eplica e on ie ad ances om behind (“ as - ollowing”).
Fo many comme cial applica ions, his may p o e adequa e, since some open models now
nea ly ma ch closed on ie models in ce ain capabili ies, wi h only a 3-mon h lag. Ye ,
his s a egy aces undamen al limi a ions:
1. Licensing and access es ic ions impose limi s on use. Fo eign closed-sou ce models
o e no gua an eed access, wi h p o ide s able o es ic , deg ade, o e mina e se -
ice.23 Bu e en “open sou ce” models ca y licensing es ic ions: Me a’s Llama p ohibi s
mos mili a y applica ions, while o he s may impose comme cial-use limi a ions. De-
ense planning and c i ical in as uc u e should no ely on sys ems whe e legal access
may be e oked o denied in he i s place.
2. The mos ad anced closed-sou ce models ha e epo edly al eady exceeded na ional se-
cu i y isk h esholds ha igge s ong secu i y measu es o p o ec hei weigh s.24
When open-weigh models app oach ha poin , he e will likely be s ong p essu e om
na ional secu i y agencies in on ie s a es o p e en hei elease o mi iga e he isks
ha hese models be weaponized in dange ous ways. This would allow he gap wi h
leading closed-sou ce models o g ow. His o ical cases o o he dual-use echnologies
(e.g., nuclea echnology,c yp og aphy) show how logics o na ional secu i y and indus-
ial ad an age jus i ied s ic con ols on in e na ional ans e . Bu AI di e s in scope:
in addi ion o secu i y, on ie models may unde pin a subs an ial po ion o economic
ac i i y, agg a a ing he consequences o dependence o exclusion. Coun ies con olling
on ie AI will ha e limi ed s uc u al incen i es o di use hese capabili ies globally.
9
A Bluep in o Mul ina ional Ad anced AI De elopmen
Bibliog aphy
•Adamson, G., & Allen, G. C. (2025). Oppo uni ies o S eng hen U.S. Biosecu i y om AI-Enabled
Bio e o ism: Wha Policymake s Should Know.h ps://www.csis.o g/analysis/oppo uni ies-s en
g hen-us-biosecu i y-ai-enabled-bio e o ism-wha -policymake s-should
•An h opic. (2024). Powe ing he nex gene a ion o AI de elopmen wi h AWS.h ps://www.an h o
pic.com/news/an h opic-amazon- ainium
•An h opic. (2025a). Ac i a ing AI Sa e y Le el 3 p o ec ions.h ps://www.an h opic.com/news/ac
i a ing-asl3-p o ec ions
•An h opic. (2025). Dis up ing he i s epo ed AI-o ches a ed cybe espionage campaign.h ps:
//asse s.an h opic.com/m/ec212e6566a0d47/o iginal/Dis up ing- he- i s - epo ed-AI-o ches
a ed-cybe -espionage-campaign.pd
•ARIA. (n.d.). Sa egua ded AI.h ps://a ia.o g.uk/oppo uni y-spaces/ma hema ics- o -sa e-ai/sa
egua ded-ai
•Bel ield, H. (2025). Domes ic on ie AI egula ion, an IAEA o AI, an NPT o AI, and a US-led
Allied Public-P i a e Pa ne ship o AI: Fou ins i u ions o go e ning and de eloping on ie AI
(No. a Xi :2507.06379; Ve sion 1). a Xi . h ps://doi.o g/10.48550/a Xi .2507.06379
•Bengio, Y. (2025). In oducing LawZe o. Yoshua Bengio.h ps://yoshuabengio.o g/2025/06/03/
in oducing-lawze o/
•Cha ez, P. (2024). So e eign AI in a Hyb id Wo ld: Na ional S a egies and Policy Responses.
Law a e.h ps://www.law a emedia.o g/a icle/so e eign-ai-in-a-hyb id-wo ld--na ional-s a
egies-and-policy- esponses
•Clyme , J., Duan, I., Cundy, C., Duan, Y., Heide, F., Lu, C., Minde mann, S., McGu k, C., Pan, X.,
Siddiqui, S., Wang, J., Yang, M., & Zhan, X. (2025). Ba e Minimum Mi iga ions o Au onomous AI
De elopmen (No. a Xi :2504.15416). a Xi . h ps://doi.o g/10.48550/a Xi .2504.15416
•Cong essional Resea ch Se ice. (2001). Enc yp ion Expo Con ols.h ps://web.a chi e.o g/we
b/20190228095041/h p://www.au.a .mil/au/awc/awcga e/c s/ l30273.pd
•Co ie , B. (2024). How a behind a e open models? Epoch AI. h ps://epoch.ai/blog/open-model
s- epo
•Co ie , B., Rahman, R., Fa o ini, L., Maslej, N., Besi oglu, T., & Owen, D. (2025). The ising cos s
o aining on ie AI models (No. a Xi :2405.21015). a Xi . h ps://doi.o g/10.48550/a Xi .2405.
21015
•Donnelly, J. M. (2022). Cos o ebuild US semiconduc o manu ac u ing will keep g owing.h ps:
// echxplo e.com/news/2022-02- ebuild-semiconduc o .h ml
•Embe son, L., & Somala, V. (2025). Build imes o gigawa -scale da a cen e s can be 2 yea s o
less. Epoch AI.h ps://epochai.subs ack.com/p/build- imes- o -gigawa -scale-da a
•Epoch AI. (2024). Machine Lea ning T ends. A chi e.Ph. h ps://a chi e.ph/Nzogc
•Eu oHPC JU. (2024). New Call o P ocu e he Eu opean Exascale Supe compu e , Alice Recoque.h ps:
//www.eu ohpc-ju.eu opa.eu/new-call-p ocu e-eu opean-exascale-supe compu e -alice- ecoq
ue-2024-09-09_en
•Eu opean Commission. (2025). AI Fac o ies.h ps://digi al-s a egy.ec.eu opa.eu/en/policies/a
i- ac o ies
•Fa ell, H., & Newman, A. L. (2019). Weaponized In e dependence: How Global Economic Ne -
wo ks Shape S a e Coe cion. In e na ional Secu i y,44(1), 42–79. h ps://doi.o g/10.1162/isec_a
_00351
•Goodin, D. (2025). New hack uses p omp injec ion o co up Gemini’s long- e m memo y. A s
Technica. h ps://a s echnica.com/secu i y/2025/02/new-hack-uses-p omp -injec ion- o-co u
p -geminis-long- e m-memo y/
•Gooding, M. (2025). Ci ies ac oss Eu ope line up o hos EU’s AI “giga ac o ies.” h ps://www.da ace
n e dynamics.com/en/news/ci ies-ac oss-eu ope-line-up- o-hos -eus-ai-giga ac o ies/
•Google Schola . (n.d.). Resea che s agged “Machine Lea ning.” h ps://schola .google.com/ci a io
ns? iew_op=sea ch_au ho s&hl=en&mau ho s=label:machine_lea ning&be o e_au ho = sT-_
16
A Bluep in o Mul ina ional Ad anced AI De elopmen
_H1DgAJ&as a =0
•IEEE Spec um. (2025). Eu ope’s Fi s Exascale Supe compu e JUPITER Powe s Science.h ps:
//spec um.ieee.o g/jupi e -exascale-supe compu e -eu ope
•In e na ional AI Sa e y Repo . (2025). Fi s Key Upda e: Capabili ies and Risk Implica ions.h ps://
in e na ionalaisa e y epo .o g/publica ion/ i s -key-upda e-capabili ies-and- isk-implica ions
•Kleinman, Z. (2025). Wha caused he AWS ou age—And why did i make he in e ne all apa ?
h ps://www.bbc.com/news/a icles/ce 1en9077 o
•Knowledge Cen e Da a & Socie y. (2025). In es AI ini ia i e. EN. h ps://da a-en-maa schappij.
ai/en/publica ions/eu opese-commissie-in es -ai
•Lambe , N. (2023). DeepSeek V3 and he cos o on ie AI models.h ps://www.in e connec s.ai
/p/deepseek- 3-and- he-ac ual-cos -o
•Lee, S. (2025). Secu ing he Backbone o A i icial In elligence: P o ec ing Da a Cen e s. New Ame -
ica. h p://newame ica.o g/ u u e-secu i y/ epo s/secu ing- he-backbone-o -ai/
•Male a, G. (2025). Non-p oli e a ion, Nuclea Technology and Peace ul Uses: Examining he Role
and Impac o Expo Con ols.
•McGui e, S., & Islam, N. (2015). Indigenous echnological capabili ies, eme ging ma ke i ms
and he ae ospace indus y. Technology Analysis & S a egic Managemen ,27(7), 739–758. h ps:
//doi.o g/10.1080/09537325.2014.959482
•Mo e i, E. (2021). The E ec o High-Tech Clus e s on he P oduc i i y o Top In en o s. Ame ican
Economic Re iew,111(10), 3328–3375. h ps://doi.o g/10.1257/ae .20191277
•OpenAI. (2025b). Announcing The S a ga e P ojec .h ps://openai.com/index/announcing- he-s
a ga e-p ojec /
•Pance ski, B. (2020). O icials Say Huawei Can Co e ly Access Telecom Ne wo ks. Wall S ee
Jou nal.h ps://www.wsj.com/a icles/u-s-o icials-say-huawei-can-co e ly-access- elecom-n
e wo ks-11581452256
•Pang, J., & Pom e , J. (2024). Chinese esea che s de elop AI model o mili a y use on back o
Me a’s Llama. Reu e s. h ps://www.cnbc.com/2024/11/01/chinese- esea che s-build-ai-model
- o -mili a y-use-on-back-o -me as-llama.h ml
•Pilz, K. (2025). Da a on GPU clus e s. Epoch AI. h ps://epoch.ai/da a/gpu-clus e s
•Pilz, K. F. (2025). The US hos s he majo i y o GPU clus e pe o mance, ollowed by China. Epoch AI.
h ps://epoch.ai/da a-insigh s/ai-supe compu e s-pe o mance-sha e-by-coun y
•Pilz, K. F., Sande s, J., Rahman, R., & Heim, L. (2025). T ends in AI Supe compu e s
(No. a Xi :2504.16026). a Xi . h ps://doi.o g/10.48550/a Xi .2504.16026
•Rosenbe g, S. (2025). Ma k Zucke be g uels AI’s alen lo e y wi h high pay o e s. Axios. h ps:
//a chi e.ph/2FYgO
•Se , G. (2024). How AI Can Au oma e AI Resea ch and De elopmen .h ps://www. and.o g/pubs/
commen a y/2024/10/how-ai-can-au oma e-ai- esea ch-and-de elopmen .h ml
•Shen, X. (2025). China’s AI capi al spending se o each up o US$98 billion in 2025: Bo A. Sou h
China Mo ning Pos . h ps://www.scmp.com/ ech/ ech-wa /a icle/3315805/chinas-ai-capi al-s
pending-se - each-us98-billion-2025-amid- i al y-us
•Smi h, T. (2023). ‘The economics o ading equi y o compu e a e no g ea ’—Mis al eleases i s i s
model. Si ed. h ps://si ed.eu/a icles/mis al- eleases- i s -ai-model/
•S an o d HAI. (2025). The 2025 AI Index Repo .h ps://hai.s an o d.edu/ai-index/2025-ai-index
- epo
•S e ens, R., Ma suoka, S., Vale o, M., Ca le , C., Fos e , I., & Gaglia di, F. (2025). The T illion
Pa ame e Conso ium (TPC).
•Subin, S. (2025). Tech megacaps plan o spend mo e han $300 billion in 2025 as AI ace in ensi ies.
CNBC. h ps://www.cnbc.com/2025/02/08/ ech-megacaps- o-spend-mo e- han-300-billion-i
n-2025- o-win-in-ai.h ml
•Suominen, A., Kauppinen, H., & Hyy inen, K. (2021). ‘Gold’, ‘Ribbon’ o ‘Puzzle’: Wha mo i a es
esea che s o wo k in Resea ch and Technology O ganiza ions. Technological Fo ecas ing and
17
A Bluep in o Mul ina ional Ad anced AI De elopmen
Social Change,170, 120882. h ps://doi.o g/10.1016/j. ech o e.2021.120882
•Tan, J., Jackson, B., Be jon, R., & Coyle, D. (n.d.). Ai bus o AI: A global s a egy o public alue
c ea ion. Benne School o Public Policy, Uni e si y o Camb idge.h ps://publicai.co/ai bus- o
-ai.pd
•Temple-Ras on, D. (2021). A “Wo s Nigh ma e” Cybe a ack: The Un old S o y O The Sola Winds
Hack. NPR.h ps://www.np .o g/2021/04/16/985439655/a-wo s -nigh ma e-cybe a ack- he-u
n old-s o y-o - he-sola winds-hack
•The Ez a Klein Show. (2024). Wha i Da io Amodei Is Righ Abou A.I.? The New Yo k Times.
h ps://www.ny imes.com/2024/04/12/opinion/ez a-klein-podcas -da io-amodei.h ml
•Tiku, N. (2024). Top AI esea che s say OpenAI, Me a and mo e hinde independen e alua ions.
The Washing on Pos .h ps://www.washing onpos .com/ echnology/2024/03/05/ai- esea ch-l
e e -openai-me a-midjou ney/
•Todd, B. (2025). When do expe s expec AGI o a i e? 80,000 Hou s.h ps://80000hou s.o g/2
025/03/when-do-expe s-expec -agi- o-a i e/
•Uni e si y o B is ol. (2025). UK-F ance Summi : Uni e si y o B is ol o lead a supe compu ing
pa ne ship wi h F ance.h ps://www.b is ol.ac.uk/news/2025/july/ ance-ai-pa ne ship.h ml
•Ve i ica ion o In e na ional AI Go e nance. (n.d.). Ox o d Ma in AIGI.h ps://aigi.ox.ac.uk/publ
ica ions/ e i ica ion- o -in e na ional-ai-go e nance/
•Vip a, J., & Ko inek, A. (2023). Ma ke Concen a ion Implica ions o Founda ion Models
(No. a Xi :2311.01550). a Xi . h ps://doi.o g/10.48550/a Xi .2311.01550
•Wang, Z., Shi, T., He, J., Cai, M., Zhang, J., & Song, D. (2025). Cybe Gym: E alua ing AI Agen s’
Real-Wo ld Cybe secu i y Capabili ies a Scale.h ps:// di.be keley.edu/blog/cybe gym/
18
A Bluep in o Mul ina ional Ad anced AI De elopmen
19
