Is AI doomed to be an energy drain?

Gen­er­a­tive AI mod­els, such as OpenAI’s GPT‑4, are pre­sent­ed as all-pur­pose tools. They include a sig­nif­i­cant num­ber of para­me­ters – now num­ber­ing in the bil­lions – that enable them to per­form any type of task. This plu­ral­i­ty of uses, which leads to com­plex­i­ty issues, makes these mod­els “in need of opti­mi­sa­tion”, accord­ing to Enzo Tartaglione, researcher and senior lec­tur­er at Télé­com Paris (IP Paris). This com­plex­i­ty also implies con­sid­er­able ener­gy con­sump­tion.

“Even for an extreme­ly sim­ple query, AI will tend to use all the resources at its dis­pos­al to respond, with­out exclud­ing those that are not use­ful. This leads to ener­gy waste, and it is real­ly some­thing we need to opti­mise.” This ener­gy con­sump­tion, esti­mat­ed at around 2% of glob­al con­sump­tion in 2024, is dri­ving research towards an alter­na­tive approach: ener­gy efficiency.

Ope­nAI has made an extreme­ly resource-inten­sive lan­guage mod­el avail­able via servers. This obser­va­tion led researchers to dis­tin­guish between the resource con­sump­tion of mod­el train­ing and that of its infer­ences, i.e. its use. Although the ener­gy con­sump­tion of train­ing is sig­nif­i­cant – approx­i­mate­ly 1,287 MWh for GPT‑3, and between 10,000 and 30,000 MWh esti­mat­ed for GPT‑4 – its impact is one-off. The impact of infer­ence, on the oth­er hand, depends on the num­ber of users, which is grow­ing con­stant­ly. A 2021 study¹ esti­mates that “between 80 and 90% of machine learn­ing work­load at NVIDIA comes from infer­ence. Ama­zon Web Ser­vices esti­mates that 90% of cloud demand for machine learn­ing is inference.”

Some researchers believe that a bal­ance needs to be found between a model’s ener­gy con­sump­tion and the task it is required to per­form. If a mod­el is used to dis­cov­er a drug or advance research – which it is capa­ble of doing – the car­bon foot­print will be eas­i­er to accept. How­ev­er, today, these mod­els can be used for all kinds of tasks, mak­ing mil­lions of infer­ences through the var­i­ous requests made of them at the same time. 

At Télé­com Paris (IP Paris), the Data Sci­ence and Arti­fi­cial Intel­li­gence for Dig­i­talised Indus­try and Ser­vices chair focus­es on sev­er­al chal­lenges: how to rec­on­cile the rise of AI and its ener­gy con­straints with­out sac­ri­fic­ing its poten­tial. “We are explor­ing issues of fru­gal­i­ty (editor’s note: seek­ing to “do more with less” and with greater respect for the envi­ron­ment), but also sus­tain­abil­i­ty (editor’s note: meet­ing the needs of present gen­er­a­tions with­out com­pro­mis­ing those of future gen­er­a­tions),” adds Enzo Tartaglione. “There is a real ques­tion in the choice of appli­ca­tions, because we can­not sim­ply point the fin­ger at AI and say it is a bad thing. Togeth­er with col­leagues, we are start­ing work on gen­er­at­ing mate­ri­als for stor­ing hydro­gen. This is also some­thing that AI can offer as a solution.”

Espe­cial­ly since the mod­els we can all use on our mobile phones require com­mu­ni­ca­tion with a serv­er. “We need to be aware of the cost of trans­port­ing infor­ma­tion, espe­cial­ly when it’s bidi­rec­tion­al,” insists the researcher. “There is there­fore a great neces­si­ty to design mod­els that can be used local­ly, lim­it­ing the need for com­mu­ni­ca­tion with an exter­nal serv­er. How­ev­er, we are talk­ing about mod­els with bil­lions of dif­fer­ent para­me­ters. This requires too much mem­o­ry for your smart­phone to do with­out the internet.”

There are there­fore sev­er­al dimen­sions to ener­gy effi­cien­cy. It is not enough to reduce the num­ber of para­me­ters required for mod­el cal­cu­la­tions, whether dur­ing train­ing or infer­ence, as the DeepSeek mod­el has done. Action must also be tak­en on the train­ing data and the data that forms the model’s knowl­edge. One solu­tion that stands out is Mis­tral, an open-source French lan­guage mod­el. There is val­ue in divid­ing the main mod­el into a mul­ti­tude of experts that can be acti­vat­ed accord­ing to the task at hand. This is one of the approach­es pro­posed for opti­mis­ing these mod­els: dis­tin­guish­ing them by spe­cial­i­ty. “The goal is to take pre-trained mod­els and imple­ment strate­gies to adapt them with as few para­me­ters as pos­si­ble to dif­fer­ent, very spe­cif­ic sub­tasks,” explains Enzo Tartaglione. “This not only avoids the impact of retrain­ing, but also great­ly improves ener­gy and performance.”

With more spe­cialised mod­els, the amount of knowl­edge required to achieve this will also require less data to be acquired. This type of mod­el could there­fore act local­ly and com­mu­ni­cate with servers in a much more stream­lined way. After all, AI is still a rel­a­tive­ly recent inno­va­tion. And, like most tech­no­log­i­cal inno­va­tions, it should log­i­cal­ly fol­low the same path of opti­mi­sa­tion as its pre­de­ces­sors. Ulti­mate­ly, fru­gal AI is more of a move­ment in AI research than a field in its own right. In fact, com­put­er sci­ence has always sought to design sys­tems that opti­mise resources and lim­it unnec­es­sary cal­cu­la­tions – fru­gal­i­ty is there­fore a nat­ur­al con­tin­u­a­tion of this log­ic of effi­cien­cy. Per­haps in the same way that com­put­ers and phones have become portable?

In any case, the appeal of fru­gal­i­ty, in addi­tion to being envi­ron­men­tal, is also eco­nom­ic for the var­i­ous play­ers devel­op­ing this type of tool. This does mean, how­ev­er, that there is still a risk of a rebound effect: more wide­spread use due to low­er devel­op­ment costs would great­ly reduce the envi­ron­men­tal ben­e­fits. How­ev­er, this approach will undoubt­ed­ly not be the only solu­tion to the ener­gy abyss that AI rep­re­sents; address­ing the issue of sus­tain­abil­i­ty will also be essential…

Pablo Andres