The AI bubble: deciphering an economic phenomenon

Arti­fi­cial intel­li­gence (AI) now occu­pies a cen­tral posi­tion in eco­nom­ic, polit­i­cal and man­age­r­i­al dis­course. It is pre­sent­ed as a dri­ver of social trans­for­ma­tion, an unprece­dent­ed lever for pro­duc­tiv­i­ty and a tool for cre­ativ­i­ty. How­ev­er, beyond the tech­no­log­i­cal enthu­si­asm, the eco­nom­ics of AI remain fraught with pro­found uncer­tain­ty because, despite the mas­sive invest­ments that demon­strate the con­fi­dence of pub­lic and pri­vate actors in its poten­tial, observed impacts on growth and pro­duc­tiv­i­ty remain unclear – some­times even disappointing.

This ten­sion between promise and real­i­ty rais­es ques­tions about the con­di­tions for val­ue cre­ation in AI-based dig­i­tal ecosys­tems. Far from being a uni­ver­sal tech­nol­o­gy, AI is deployed accord­ing to dif­fer­ent sec­toral, insti­tu­tion­al and ter­ri­to­r­i­al log­ics. Forms of AI inte­gra­tion vary based on to the avail­abil­i­ty of data, the skills of the play­ers and organ­i­sa­tion­al strate­gies. As devel­op­ments in IT have already shown in the past, pro­duc­tiv­i­ty gains depend not only on algo­rith­mic per­for­mance, but also on the trans­for­ma­tion of rou­tines, com­ple­men­tary skills and coor­di­na­tion mechanisms.

Since the spec­tac­u­lar spread of so-called ‘gen­er­a­tive’ mod­els — those that pro­duce text, images, sound or code — AI seems to have become a dig­i­tal Swiss Army knife, capa­ble of doing every­thing: writ­ing a report, gen­er­at­ing an image, com­pos­ing music, design­ing a mar­ket­ing plan or even writ­ing code. With its ‘all-in-one’ nature, AI appears to be both a uni­ver­sal tool and a lever for automa­tion and wide­spread creativity.

How­ev­er, AI plat­forms do not just pro­vide tech­ni­cal tools: they now struc­ture mar­kets, stan­dards and behav­iours, there­by con­tribut­ing to a rede­f­i­n­i­tion of pow­er rela­tions in the glob­al dig­i­tal econ­o­my. We are wit­ness­ing the rise of glob­al play­ers who, through the ver­ti­cal inte­gra­tion of infra­struc­ture, soft­ware and uses, are organ­is­ing tech­no­log­i­cal lock-in mech­a­nisms that accen­tu­ate the asym­me­tries of depen­den­cy between coun­tries, com­pa­nies and users, while rais­ing the ques­tion of dig­i­tal sovereignty.

Behind the large mul­ti­modal mod­els – which are sup­posed to be capa­ble of doing every­thing – lies an AI econ­o­my that is increas­ing­ly seg­ment­ed, con­tex­tu­al and tar­get­ed towards these usages. We are wit­ness­ing grow­ing ten­sion between two dynam­ics. On the one hand, there is a trend towards inte­gra­tion, embod­ied by large gen­er­al­ist mod­els that seek to build glob­al arti­fi­cial intel­li­gence plat­forms capa­ble of respond­ing to all types of needs. On the oth­er, there is a trend towards spe­cial­i­sa­tion, dri­ven by start-ups, lab­o­ra­to­ries and sec­tor-spe­cif­ic play­ers (health, finance, indus­try, law, ener­gy, cul­ture, etc.), that are devel­op­ing spe­cialised, con­tex­tu­al mod­els inte­grat­ed into spe­cif­ic sec­tors and val­ue chains.

In one case, AI is pre­sent­ed as a uni­ver­sal per­son­al assis­tant, an intel­li­gent inter­face for every­one. But in the oth­er case, it becomes a func­tion­al build­ing block, often invis­i­ble, inte­grat­ed into busi­ness process­es, analy­sis tools or pro­duc­tion infrastructures.

The two approach­es inter­sect and some­times con­tra­dict each oth­er. Large AI mod­els are heavy, cost­ly infra­struc­tures that require enor­mous com­put­ing and data resources; their oper­a­tion is based on eco­nom­ic prin­ci­ples sim­i­lar to those of cloud plat­forms or inte­grat­ed soft­ware ecosys­tems. Con­verse­ly, spe­cialised mod­els tend towards a more refined, often more eth­i­cal and sus­tain­able use econ­o­my, but one that is frag­ment­ed and depen­dent on tech­ni­cal stan­dards and inter­faces or APIs opened up by the major players.

The illu­sion of a promise of uni­ver­sal or gen­er­al AI is based on two impor­tant fac­tors. The first is the fas­ci­na­tion with gen­er­al­ist mod­els. The ‘big mod­els’ that make up Chat­G­PT, Gem­i­ni, Claude, Lla­ma and Mis­tral are pre­sent­ed as uni­ver­sal plat­forms thanks to their ver­sa­til­i­ty and mul­ti­modal­i­ty, as well as the ‘all-in-one’ image they project.

The dom­i­nant play­ers in AI — main­ly large Amer­i­can and Chi­nese com­pa­nies — are char­ac­terised by a ver­ti­cal inte­gra­tion strat­e­gy: they simul­ta­ne­ous­ly con­trol the infra­struc­ture lay­ers (cloud, proces­sors, net­works), the appli­ca­tion lay­ers (AI tools, inter­faces) and end uses. This mod­el of ‘sys­temic plat­formi­sa­tion’ allows them to cap­ture not only the val­ue gen­er­at­ed by inno­va­tion, but also the exter­nal­i­ties pro­duced by inter­ac­tions between users, devel­op­ers and data pro­duc­ers. These com­pa­nies oper­ate as ecosys­tem archi­tects, set­ting the rules for data access and shar­ing, inter­op­er­abil­i­ty stan­dards and eco­nom­ic con­di­tions for participation.

How­ev­er, AI is also marked by two phe­nom­e­na already encoun­tered with IT and dig­i­tal tech­nol­o­gy. The first is the Solow para­dox: AI is every­where – except in pro­duc­tiv­i­ty sta­tis­tics. Any gains only involve very spe­cif­ic seg­ments and are quick­ly absorbed by invest­ments and the deploy­ment of new ser­vices and activ­i­ties. AI only gen­er­ates sus­tain­able gains if it is accom­pa­nied by a recon­fig­u­ra­tion of work and pro­duc­tion, upskilling and col­lec­tive learn­ing. How­ev­er, it takes time between acquir­ing tech­nolo­gies and being able to use them effectively.

A sec­ond phe­nom­e­non is a bub­ble effect. On the one hand, it stems from mas­sive invest­ments across the board: Open AI is plan­ning to invest $10 tril­lion by 2033, while announc­ing only $13 bil­lion in annu­alised rev­enue this year and no prof­itabil­i­ty until 2030. The scale of these invest­ments is also due to finan­cial manoeu­vres that go round in cir­cles with endog­a­mous agree­ments (see Fig­ure 1) and a lim­it­ed num­ber of play­ers and com­pa­nies involved in AI, feed­ing off each other.

Faced with the risk of the bub­ble burst­ing, major dig­i­tal play­ers are turn­ing their atten­tion to the cor­po­rate world. B2B finan­cial mod­els are health­i­er because they do not require invest­ment in com­put­ing pow­er for infer­ence. The enter­prise mar­kets and the digi­ti­sa­tion of indus­try are prov­ing to be a more impor­tant issue for the econ­o­my than the mass mar­ket and indi­vid­u­als, which are more often discussed.

Fur­ther­more, the devel­op­ment of AI in busi­ness­es faces the more gen­er­al dif­fi­cul­ties of dig­i­tal­i­sa­tion of the com­pa­nies them­selves. Tech­nolo­gies are organ­ised and inter­twined in ‘sys­tems’ with­out it being pos­si­ble to iso­late them from one anoth­er. They com­bine ‘tech­ni­cal build­ing blocks + organ­i­sa­tion­al ele­ments + pro­ce­dur­al rules and imple­men­ta­tion process­es’ in a way that is absolute­ly insep­a­ra­ble from one anoth­er. In their prac­tices, agents and work col­lec­tives simul­ta­ne­ous­ly mobilise dif­fer­ent build­ing blocks with­out being able to iden­ti­fy the spe­cif­ic con­tri­bu­tions of each tech­ni­cal component.

Val­ue cre­ation in the dig­i­tal sec­tor is based on spe­cif­ic com­bi­na­tions of tech­ni­cal resources and eco­nom­ic appli­ca­tions (see Fig­ure 2). Sev­er­al main inter­con­nect­ed links can be iden­ti­fied: the hard­ware that con­sti­tutes the phys­i­cal infra­struc­ture (com­po­nents, infra­struc­ture, equip­ment), the soft­ware that enables this infra­struc­ture to be used (oper­at­ing sys­tems, inter­faces, devel­op­ment tools), the data that rep­re­sents the flow of infor­ma­tion (col­lec­tion, stor­age, pro­cess­ing), the mod­els that struc­ture the trans­for­ma­tion of this data into knowl­edge or deci­sions (algo­rithms, machine learn­ing, process rep­re­sen­ta­tion), and final­ly the appli­ca­tions, which trans­late these capa­bil­i­ties into con­crete eco­nom­ic and social uses (ser­vices, plat­forms, end uses), with­out even men­tion­ing direc­tives that estab­lish the reg­u­la­to­ry framework.

The high­est val­ue cap­ture occurs at the lev­el of mod­els and their appli­ca­tions. High mar­gins can be achieved through API licences, sub­scrip­tions and user lock-in. To a less­er extent, cloud and plat­forms also ben­e­fit from sig­nif­i­cant recur­ring rev­enues due to the lock-in they can achieve through their inte­gra­tion. Final­ly, semi­con­duc­tor foundries and man­u­fac­tur­ers (high­ly cap­i­tal-inten­sive, high bar­ri­ers to entry) and spe­cialised data providers (local or niche monop­o­lies) may cap­ture only a small pro­por­tion of the val­ue of AI, but their posi­tion remains the most strategic.

Gen­er­a­tive AI mod­els rely on the mas­sive col­lec­tion of het­ero­ge­neous data — often from open ecosys­tems — and its large-scale reuse. This process gives a major com­pet­i­tive advan­tage to play­ers capa­ble of accu­mu­lat­ing and pro­cess­ing con­sid­er­able vol­umes of infor­ma­tion. How­ev­er, the val­ue derived from this data is not sole­ly a func­tion of its quan­ti­ty, but also of its qual­i­ty and con­tex­tu­al­i­sa­tion. The most effec­tive mod­els are those that man­age to com­bine large scale with local rel­e­vance, inte­grat­ing data spe­cif­ic to a par­tic­u­lar use, sec­tor or language.

Beyond con­cen­tra­tion, the AI econ­o­my is also see­ing a diver­si­ty of sit­u­a­tions emerge. Infra­struc­ture mod­els are based on the pro­vi­sion of com­put­ing and stor­age capac­i­ty (cloud, GPU). This mod­el, dom­i­nat­ed by a few play­ers, relies on economies of scale and net­work effects. Appli­ca­tion mod­els, on the oth­er hand, focus on cre­at­ing sec­tor-spe­cif­ic solu­tions (health, finance, ener­gy, edu­ca­tion) tai­lored to spe­cif­ic needs. This mod­el pri­ori­tis­es val­ue in use and cus­tomer prox­im­i­ty. Final­ly, Euro­pean pub­lic and indus­tri­al ecosys­tems favour con­sor­tium or part­ner­ship mod­els in which com­pa­nies coop­er­ate to devel­op con­tex­tu­alised AI solu­tions, pool­ing data and risks.

These tra­jec­to­ries con­firm that val­ue does not lie in the tech­nol­o­gy itself, but in the abil­i­ty to orches­trate an ecosys­tem of het­ero­ge­neous play­ers around an archi­tec­ture. AI thus becomes an instru­ment of strate­gic struc­tur­ing, rather than a sim­ple pro­duc­tion tool: it rede­fines rela­tion­ships of depen­den­cy, sources of legit­i­ma­cy and levers of com­pet­i­tive dif­fer­en­ti­a­tion. More specif­i­cal­ly, we can dis­tin­guish between uses of AI that relate to pro­duc­tion process­es, deci­sion-mak­ing process­es, coop­er­a­tion mech­a­nisms, or sup­port for inno­va­tion (see Fig­ure 3).

Thus, although pub­lic dis­course on arti­fi­cial intel­li­gence (AI) often focus­es on dis­rup­tion and the replace­ment of human labour, its most tan­gi­ble effect in most con­tem­po­rary busi­ness­es is incre­men­tal improve­ments in pro­duc­tiv­i­ty. In prac­tice, AI acts as a lever for stream­lin­ing process­es, accel­er­at­ing deci­sion-mak­ing, improv­ing coor­di­na­tion and enhanc­ing inno­va­tion capa­bil­i­ties. These trans­for­ma­tions do not nec­es­sar­i­ly dis­rupt the struc­ture of the com­pa­ny, but they do pro­found­ly recon­fig­ure its inter­nal effi­cien­cy and val­ue cre­ation. This can be seen in the vari­ety of exist­ing use cas­es, which reflect the rise of ver­ti­cal and sec­tor-spe­cif­ic AI in health­care, finance, law, indus­try, defence, ener­gy, cul­ture, and more. This vari­ety is due to the spe­cial­i­sa­tion of mod­els and their adap­ta­tion to data sets and busi­ness constraints.