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.