Generative AI: energy consumption soars

Arti­fi­cial intel­li­gence (AI) has found its way into a wide range of sec­tors: med­ical, dig­i­tal, build­ings, mobil­i­ty, etc. Defined as “a computer’s abil­i­ty to auto­mate a task that would nor­mal­ly require human judge­ment¹”, arti­fi­cial intel­li­gence has a cost: its large-scale deploy­ment is gen­er­at­ing grow­ing ener­gy require­ments. The IT tasks need­ed to imple­ment AI require the use of user ter­mi­nals (com­put­ers, tele­phones, etc.) and above all data cen­tres. There are cur­rent­ly more than 8,000 of these around the world, 33% of which are in the Unit­ed States, 16% in Europe and almost 10% in Chi­na, accord­ing to the Inter­na­tion­al Ener­gy Agency² (IEA). Data cen­tres, cryp­tocur­ren­cies and arti­fi­cial intel­li­gence will account for almost 2% of glob­al elec­tric­i­ty con­sump­tion in 2022, rep­re­sent­ing elec­tric­i­ty con­sump­tion of 460 TWh. By com­par­i­son, French elec­tric­i­ty con­sump­tion stood at 445 TWh in 2023³.

How much of this elec­tric­i­ty con­sump­tion is actu­al­ly ded­i­cat­ed to AI? “We don’t know exact­ly” replies Alex de Vries. “In an ide­al case, we would use the data pro­vid­ed by the com­pa­nies that use AI, in par­tic­u­lar the GAFAMs, which are respon­si­ble for a large pro­por­tion of the demand.” In 2022, Google pro­vid­ed infor­ma­tion on the sub­ject for the first time⁴: “The per­cent­age [of ener­gy used] for machine learn­ing has held steady over the past three years, rep­re­sent­ing less than 15% of Google’s total ener­gy con­sump­tion.” How­ev­er, in its lat­est envi­ron­men­tal report⁵, the com­pa­ny pro­vides no pre­cise data on arti­fi­cial intel­li­gence. Only the total elec­tric­i­ty con­sump­tion of its data cen­tres is giv­en: 24 TWh in 2023 (com­pared with 18.3 TWh in 2021).

In the absence of data pro­vid­ed by com­pa­nies, the sci­en­tif­ic com­mu­ni­ty has been try­ing to esti­mate the elec­tric­i­ty con­sump­tion of AI for sev­er­al years. In 2019, an ini­tial arti­cle⁶ threw a span­ner in the works: “The devel­op­ment and train­ing of new AI mod­els are cost­ly, both finan­cial­ly […] and envi­ron­men­tal­ly, due to the car­bon foot­print asso­ci­at­ed with pow­er­ing the equip­ment.” The team esti­mates that the car­bon foot­print of the total train­ing for a giv­en task of BERT, a lan­guage mod­el devel­oped by Google, is rough­ly equiv­a­lent to that of a transat­lantic flight. A few years lat­er, Google sci­en­tists believe that these esti­mates over­es­ti­mate the real car­bon foot­print by 100 to 1,000 times. For his part, Alex de Vries has cho­sen to rely on sales of AI hard­ware⁷. NVIDIA dom­i­nates the AI serv­er mar­ket, account­ing for 95% of sales. Based on serv­er sales and con­sump­tion, Alex de Vries pro­ject­ed elec­tric­i­ty con­sump­tion of 5.7 to 8.9 TWh in 2023, a low fig­ure com­pared with glob­al data cen­tre con­sump­tion (460 TWh).

But these fig­ures could sky­rock­et. Alex de Vries esti­mates that by 2027, if pro­duc­tion capac­i­ty match­es the com­pa­nies’ promis­es, NVIDIA servers ded­i­cat­ed to AI could con­sume 85 to 134 TWh of elec­tric­i­ty every year. The cause: the surge in the use of gen­er­a­tive AI. Chat­G­PT, Bing Chat, Dall‑E, etc. These types of arti­fi­cial intel­li­gence, which gen­er­ate text, images or even con­ver­sa­tions, have spread across the sec­tor at record speed. How­ev­er, this type of AI requires a lot of com­put­ing resources and there­fore con­sumes a lot of elec­tric­i­ty. Accord­ing to the AIE, inter­ac­tions with AIs such as Chat­G­PT could con­sume 10 times more elec­tric­i­ty than a stan­dard Google search. If all Google search­es – 9 bil­lion every day – were based on Chat­G­PT, an addi­tion­al 10 TWh of elec­tric­i­ty would be con­sumed every year. Alex De Vries esti­mates the increase at 29.3 TWh per year, as much as Ireland’s elec­tric­i­ty con­sump­tion. “The steady rise in ener­gy con­sump­tion, and there­fore in the car­bon foot­print of arti­fi­cial intel­li­gence, is a well-known phe­nom­e­non,” com­ments Anne-Lau­re Ligozat. “AI mod­els are becom­ing increas­ing­ly com­plex: the more para­me­ters they include, the longer the equip­ment runs. And as machines become more and more pow­er­ful, this leads to increas­ing­ly com­plex mod­els…”. For its part, the Inter­na­tion­al Ener­gy Agency esti­mates that in 2026, the increase in elec­tric­i­ty con­sump­tion by data cen­tres, cryp­tocur­ren­cies and AI could amount to between 160 and 590 TWh com­pared with 2022. This is equiv­a­lent to the elec­tric­i­ty con­sump­tion of Swe­den (low esti­mate) or Ger­many (high estimate).

The pro­cess­ing needs of AI can be explained by dif­fer­ent phas­es. AI devel­op­ment involves an ini­tial learn­ing phase based on data­bas­es, known as the train­ing phase. Once the mod­el is ready, it can be used on new data: this is the infer­ence phase¹⁰. The train­ing phase has long been the focus of sci­en­tif­ic atten­tion, as it is the most ener­gy-inten­sive. But new AI mod­els have changed all that, as Alex de Vries explains: “With the mas­sive adop­tion of AI mod­els like Chat­G­PT, every­thing has been reversed and the infer­ence phase has become pre­dom­i­nant.” Recent data pro­vid­ed by Meta and Google indi­cate that it accounts for 60–70% of ener­gy con­sump­tion, com­pared with 20–40% for train­ing¹¹.

While AI’s ener­gy con­sump­tion is fraught with uncer­tain­ty, esti­mat­ing its car­bon foot­print is a chal­lenge for the sci­en­tif­ic com­mu­ni­ty. “We are able to assess the foot­print linked to the dynam­ic con­sump­tion of train­ing, and that linked to the man­u­fac­ture of com­put­er equip­ment, but it remains com­pli­cat­ed to assess the total foot­print linked to use. We don’t know the pre­cise num­ber of uses, or the pro­por­tion of use ded­i­cat­ed to AI on the ter­mi­nals used by users,” stress­es Anne-Lau­re Ligozat. “How­ev­er, col­leagues have just shown that the car­bon foot­print of user ter­mi­nals is not neg­li­gi­ble: it accounts for between 25% and 45% of the total car­bon foot­print of cer­tain AI mod­els.” Anne-Lau­re Ligozat and her team esti­mate that train­ing the BLOOM AI mod­el – an open-access mod­el – emits around 50 tonnes of green­house gas­es, or 10 times more than the annu­al emis­sions of a French per­son. This makes it dif­fi­cult for the tech giants to achieve their car­bon neu­tral­i­ty tar­gets, despite the many off­set­ting mea­sures they have tak­en. Google admits in its lat­est envi­ron­men­tal report: “Our [2023] emis­sions […] have increased by 37% com­pared to 2022, despite con­sid­er­able efforts and progress in renew­able ener­gy. This is due to the elec­tric­i­ty con­sump­tion of our data cen­tres, which exceeds our capac­i­ty to devel­op renew­able ener­gy projects.”

Lim­it­ing glob­al warm­ing means dras­ti­cal­ly reduc­ing glob­al green­house gas emis­sions. Is AI at an impasse? “None of the argu­ments put for­ward by Google to reduce AI emis­sions hold water” deplores Anne-Lau­re Ligozat. “Improv­ing equip­ment requires new equip­ment to be man­u­fac­tured, which in turn emits green­house gas­es. Opti­mis­ing infra­struc­tures – such as water cool­ing for data cen­tres – shifts the prob­lem to water resources. And the relo­ca­tion of data cen­tres to coun­tries with a low-car­bon elec­tric­i­ty mix means that we need to be able to man­age the addi­tion­al elec­tric­i­ty demand…” As for the opti­mi­sa­tion of mod­els, while it does reduce their con­sump­tion, it also leads to increased use – the famous ‘rebound’ effect. “This tends to can­cel out any poten­tial ener­gy sav­ings,” con­cludes Alex de Vries. “My main argu­ment is that AI should be used sparingly.”

Anaïs Marechal