How big is AI's carbon footprint? | BBC News

you're watching the context it's time

for our weekly regular segment AI

decoded welcome to Aid decoded every

week in this program we take you deep

into the world of artificial

intelligence and what we've tried to do

is focus these programs on one

particular theme so last week we looked

at the advance of AI on the battlefield

tonight we're going to consider

something a number of you have raised

and that is AI and energy or more

specifically the energy that AI consumes

there are climate experts who are

warning that the advance of artificial

intelligence could lead to an 80%

increase in our Global carbon emissions

well let's start with sustainability

because that cloud that AI models live

on is actually made out of metal plastic

and powered by vast amounts of energy

and each time you query an AI model it

comes with a cost to the planet that is

Sasha luch lead climate researcher at

the AI company hugging face she's going

to be joining us from Montreal in just a

moment how much energy are we using

while according to this report from Vox

AI is already consuming as much energy

as a small country and we are only at

the beginning the next web says this is

where our two existential crises collide

with one another climate crisis and the

exponential growth of AI can one help

solve the other or will it exac acate

the problem here in the studio our

regular AI contributor PRI Lani CEO of

the AI powered education company Century

T welcome good to see you right look

when it comes to digital everybody knows

there is a cost there's the wiring

there's the chips there's the precious

metals there's the water that cools the

data processing centers what we often

don't don't talk about is the energy

that goes into producing Ai and

specifically the training of AI language

models like chat GPT yeah why because

they're far more hungry than what we've

seen before so you have these AI models

that you and I have talked about we've

shown we've played with right these air

models are trained and they're deployed

in data centers and the data centers

consume vast amounts of electricity so

if you are powering the data center with

non-renewable sources then essentially

you have potentially huge carbon

emissions and those particular models

that we're talking about these

generative AI models not just language

models right not just llm in terms of

chat models but also the multimodal

models images videos consume an enormous

amount so if you remember in about 2019

we used to say if you stream an hour of

video right it was 36 grams of CO2 just

to put that into context for everyone

because that's why we're here okay

that's driving a car typical kind of

petrol car about 160 m now meta they

just released llama 3 on hugging face

one of my favorite platforms okay and

that model they said um emitted about 2,

290 metric tons of CO2 and so you're

going to say put that into context

that's about 500 average cars and what

they Adit in an entire year so you start

to get what we had with digital which is

what you asked all the way to these

really huge AI models and I just want to

show you uh Christian what I looked at a

little bit earlier before I came in so I

just spent I just looked at half an hour

of the images produced on Mid Journey so

this is mid Journey that you're seeing

now okay so you've got about 10 unique

images that have been produced by this

particular platform and that was about

half an hour about 4:30 p.m. UK time

today okay just those images and that's

just on one channel of where you can

find mid Journey okay one there are

there are lots hundreds and hundreds but

just those 10 alone would take about two

and a half times the battery charge of a

phone so this is a significant problem

and if anyone's interested there are

some great interviews actually with Mark

Zuckerberg and others that say that this

is there is going to be an energy Bott

neck right and obviously for activists

this is a significant problem so I'm

really looking forward to having Sasha

on to discuss this with her well let's

bring her in Sasha lney uh she is the

lead climate researcher at hugging face

uh which priet really likes and she

spent nearly a decade looking at data

storage and machine learning and how

this all contributes to our energy

consumption welcome to the

program thank you for having me right

before we before we get started maybe I

could just frame our conversation with

an image that you sent us actually that

that for me really underlines just how

expensive our digital usage has become

so here is Google's annual energy use

18.3 trillion Watts that's 10 or 15% of

that is going towards Ai and here's what

the Republic of Ireland uses in any one

year

29.3 trillion wat so one company Sasha

is now consuming about 2third of what a

small country uses each year and yet I

don't hear governments talking about

that problem

I think it's because uh AI is really a a

horizontal it's not like your typical

vertical like agriculture transportation

and it actually affects all Industries

anything that uses AI anything from

navigation to web search and so I think

governments don't really know what

bucket to put it in and when you don't

know what bucket it goes in you tend to

kind of let it slip through the cracks

and Sasha can you describe this in terms

of the scale of large language models so

what their usage is like and actually

tell us what therefore that means in

terms of energy consumption and the

impact on the

environment definitely language models

have become one of the most popular

usages of AI and they're being deployed

in everything nowadays you can talk to

to your stove or your fridge and uh in a

recent study we did we found that so

training a large language model is

definitely very uh energy intensive and

that's the the numbers that you gave but

actually each query also uses energy and

depending on the size of the model 200

to 500 million queries will equal the

amount of energy used for training so it

might seem like a lot but uh for chat

GPT it it it averages around 10 million

users a day so within a couple of weeks

you have this vast amount of energy that

you know is equivalent to all these cars

over a year but just with people using

the tool did you see this story this

week um that microsof Microsoft are

going to plow in about a hundred billion

dollars into this super computer called

Stargate um and it's going to be powered

by not one they say not one but several

nuclear power stations and and that got

me thinking because I I have heard Sam

Alman at at chat gbt talk about this

open AI he's talked about this and he he

says yeah that that's how we're going to

have to work we're going to have to

create our own Energy Systems is that is

that perhaps where new energy comes from

these the biggest companies in the world

driving the

investment well I mean it is it is a

problem because to what extent do you

want big tech companies to be building

their own nuclear reactors and um maybe

that energy can be better used used uh

for other things right because we should

be decarbonizing our energy globally and

currently if we're going to funnel all

that investment into the energy use for

AI maybe other sectors will get

overlooked and we should be focusing on

those if we really want to decarbonize

and Sasha what I'm quite excited about

is look when we talk about AI models and

I was on hugging face today so for those

who don't know what it is it's an AI

model repository and playing with llama

3 I like the fact that models now have a

description potentially of you know how

much energy they have used you've

produced something pretty novel at

hugging face haven't you so um when

you're when you're building these AI

models we're focused on the latency the

speed of the model and how performant

the model is but you're potentially

creating a like a little trip advisor of

how efficient it is so tell us about

that yeah currently when people go on

hugging phas they tend to shop around a

little bit for models that work for

their for the task that they want to do

it could be language it could be it

could be audio it could be image

generation now we're even looking at

video and typically they'll look at

things like performance or latency but

I'm calculating the energy usage across

all different tasks and models on on the

Hub on the hugging hugging face website

and I want to provide that information

to people so that they can Factor it in

so maybe this model uh is not only

faster but it's more efficient maybe

this model is slightly less performant

but it's vastly more efficient so I'm

I'm develop en I'm developing energy

star ratings for AI models well listen

since we're talking about how we

mitigate the problem let me introduce

you to Chris starky because he is the

CEO of the London Bay startup next Jen

cloud and they been in business since

2020 and they Source data centers that

are entirely powered by renewable energy

welcome Chris to the program tell us

what you do and how your clients would

typically

work well I mean so I mean we're on the

other side of the fence where our

business is all about building you know

large scale um GPU clusters basically H

we've got a core focus on building high

density accelerate accelerated compute

basically this is typically what

companies um maybe that use hugging face

or companies that are using uh or

building their their own foundational

models this is the type of

infrastructure that they would be uh

consuming um and our mission effectively

is to deliver this at scale 100%

renewably powered and Chris how do we

achieve this sort of optimization of

models because there's all sorts of

things that we could be looking at so

we're looking at infrastructure

efficiency uh that I know you look at in

terms of cooling Technologies I'd love

to hear more about that and then also

the optimized um Hardware the say for

example Nvidia brought out the black

wellpath platform and what they say with

Blackwell is that it reduced the costs

and energy consumption by about 25x for

tech companies so can you describe these

sorts of methodologies you know how

Reliant we are on these and when we talk

about cooling methodologies can you

explain that to us you know we know that

data centers and we know that these

racks need cooling but it would be

really great if you can walk the

audience of the process how long have we

got um I'm not too sure how how long

we've got but uh not long

so so I a short high level I mean you

know only a few years ago you know we

were building um high performance

environments that you know maybe 10 to

to 20 kilowatts um was was was deemed as

kind of high dense um and uh you know

now quite commonly we're building out um

you know environments that are 50 to 60

kilowatt per rack um and the new

iteration the next generation of chips

um and some of the infrastructure you

know that that we're bringing into play

for next year 2025 you know we're going

to be north of 120 kilowatts per act so

you know we're seeing a a clear you know

uh increase in exponential growth in

density um and you know that that's

great we can fit more power into Data

Centers but you know the the the data

centers will eventually be drawing a

huge amount of resource a huge amount of

energy so it's it's just unsustained

able to to have traditional techniques

of calling like air coing which is quite

common now um so you know we've got a

keen focus on on you know triing and

testing new ways to cool the chips um

one of the new ways um is something

called you know liquid cool direct to

chip um this brings a huge amount of

efficiency but at the same time you know

as chips get more powerful you know

we're we're obviously drawing a huge

amount more power you know per square

feet or if you like or per square foot

in each dat Center so you know yeah St

no I'm just going to say because we're

really pressed for time but I just want

to get a really quick final answer from

both of you maybe you could chip in on

this um pardon the P um what about

sovereignty because everybody wants

control of their own computers and and

obviously some of that's going to come

down to where the cloud is what energy

they have what energy they can generate

if we're going to make this available to

everybody how concerned are you both by

that Sasha let me start with you first I

feel that AI is really uh slipping

through the cracks when it comes to

accounting for energy and carbon because

it's often companies in one country

using cloud comput in another country

and often the the for example every time

I talk to Cloud providers they like we

don't know what's running on our centers

we it could be streaming it could be AI

so it's really hard for them to count uh

to to account for this energy usage so

every time I'm like okay give me a

number they're like we don't have any

numbers so I've seen that uh it's

currently not being accounted for let's

say yeah

Chris yeah I mean well I mean if they if

they're trying to do it sustainably I

think a lot of you know countries will

will struggle you know they they they

absolutely will there's just not enough

infrastructure you know locally to to to

provide sustainable infrastructure not

at the scale of the demand that we're

seeing currently every country is going

to want a sovereign Cloud they're all

absolutely going for it right now

everyone wants their own Sovereign gbt

for example um they're not going to be

able to do it currently certainly not

here in the UK I do not think no I'm PR

was just saying everybody wants a super

computer every wants a super computer

also as well I mean you know I just

thinking back to G20 you know when when

you you know to to to the the climate

conferences and and and you know when

people talk about carbon Footprints and

and what belongs to that carbon

footprint we talk about emissions but we

never talk about Cloud power or or

generating comp you know computer

generating it's the length of time that

will take for Supply to to create these

renewable energy you know data centers

100% to the demand think about how quick

it was that chat GPT exploded right

amazing Sasha Lon Chris starky amazing

stuff thank you for being with us here

on AI deoda come back soon after the

break Pria will guide us through some of

the big stories of the week anyone seen

the ad for the new Apple iPad not

everyone's cup of tea and I'll show you

a biopic on President Putin made

entirely back by Ai and how long before

we can speak to the animals we'll be

right

back you're watching AI decoded now the

new ad for the AI powered iPad has spot

quite the backlash among Hollywood

creatives uh why we will find out

because it's uh probably something to do

with this the fact that you can create

an entire movie now without auditioning

someone who looks like the Russian

president this is the AI generated

President Putin a biopic the guardian

says will be out for release in

September uh if you can reproduce

Vladimir Putin then it stands to reason

you can recreate anyone there's a story

here from the times that reports there's

been a surge in dead Bots or grief Bots

families using AI to bring their loved

ones back to life but does it require

tighter regulation and how long before

we can understand what the animals are

saying scientists have spoken to Sky

News and they say they think AI will one

day help them communicate with a sperm

well uh Dr doitt I presume let's start

with uh this huge sensitivity forer that

there is within the creative arts

industry um you know about about what AI

is doing replacing Talent destroying The

Human Experience of playing an

instrument or writing a song and so with

that in mind let me just play you the

new ad from apple and we'll talk about

it off the

B when I'm down and all

alone All I Ever Need is

[Music]

You

wi

they and we watch the M SN

oh yeah it's called Crush um imagine all

the things that it'll uh be used to uh

to create Apple CEO Tim Cook said let me

just uh that it'll uh obviously

revolutionize uh the Arts industry but

let me just read you this from Asif

capadia who is a a writer and director

she said I don't know why anyone thought

this AB was a good idea it's the most

honest metaphor she says for what tech

companies do to the Arts to musicians

creators writers filmmakers squeeze them

use them not pay well take everything

and that's the point right we've talked

about this on the program before yeah a

bit insensitive isn't it it's

insensitive I think obviously there's a

lot of anger because of two reasons

there the threat that AI poses but then

also it's just this idea that Tech

really misunderstands the art and the

fact that you know what's Technology's

role going to be in diminishing artistic

and cultural mediums right but what's

what's really interesting and I think

there's a really funny video do you want

to play and just it would take literally

30 seconds on actually what one director

created in response to Apple's um

adverts and I don't know if we can get

that up right now but it was a filmmaker

who reversed it and essentially um

described that you know was it was it

was a very very good quality video and

it was essentially all of the Arts all

the instruments crushing the iPad and it

was the reverse of that but you know why

do I not think that this is an actual

threat okay because instruments are not

just tools right your guitar your piano

that's a tactile that's an immersive

experience I've got a nickon DSR

um DSLR it enables me to have that sort

of artistic creativity when I'm taking a

picture right which you know an iPad as

great as it is and I love my iPad but

it's not able to do and then when it

comes to painting and drawing for

example you've got these artists that

use you know different physical mediums

like oils and acrylics it's not just all

about that digital photo but it was

incredibly insensitive I think an apple

people are wondering what what the

movies of the future might look like uh

and what it might do to the Cinematic

Arts in particular um maybe you should

have a quick look at this this is the

new AI biopic it give you an idea

entirely created by artificial

intelligence the president has found

time for you after all this will

culminate Nuclear

Strike only I can save you and your

family from prison Vladimir Putin has

won with

5344 of the votes it's an embarrassing

result what would have been the

satisfactory result for you Mr President

100 you interesting thing about that and

this is what panics people in Hollywood

is that you don't need auditions for

people that look like Putin you can

create Putin they've got a Bill Clinton

in there looks well it's Bill Clinton I

mean then there might be an issue over

that I mean absolutely in terms of

obviously your image rights right and I

think that's obviously we talked about

the legal issues before but I mean you

can do this at low cost with this sort

of Technology now right you don't have

to deal with the egos necessarily right

that you might otherwise have to but

there are huge ethical concerns with

this not just because it's easy to talk

about that being Putin I think a lot of

the world would say oh that's fine but

actually you're starting to potentially

blur the lines between between reality

and fiction you can misrepresent in many

ways I mean you do question why they

didn't just have an actor in The biopic

this is clearly just about using that

technology but I know there's so much

more to come in the future when we talk

about deep FES um and whether they're a

good use or a bad use but you know

clearly if this was if you if was of

rishy sunak kissed armor then it it

would be people would be up in arms

about it yeah it takes on a much greater

effect doesn't it when it looks

specifically like them exactly um look

the best story of the week is this story

uh about sperm whales we've known for

some time of course that their calls and

their clicks are highly

sophisticated to coordinate and

communicate with one another but what if

there is actually a language and these

scientists that have been studying the

Welles in the Eastern Caribbean found

that they use mors code with a a rhythm

or sort of tempo that suggests that

there is some sort of meaning behind it

and that's where potentially AI comes in

yeah how how all right okay so when

you're looking at language okay we map

language we map the words like imagine

creating a map of all the words okay and

where they are on this map and then you

calculate essentially the statistical

difference between words okay and this

is how you end up with these large

language models and you look at

similarities between words okay and so

here what they're looking at is when

they've got these clicks they're looking

simplistically for patterns between

those clicks and then they're actually

recording what the whales are doing and

then finding how those patterns might

relate to diving really deep right or

coming up or feeding and then they're

attaching essentially those patterns in

and creating this map and this language

and then there will be an exercise where

they'll try and essentially overlay that

with the English language the test is

right when they hear these clicks

because they're really loud by the way

so if you watch videos on on these

whales they're they're super loud right

um but when they hear those clicks can

the AO model that they then create

predict what the whale is going to do

and then that's how you constantly then

test and if that works then you can

extend that a bit further and you can

then start to create the language and

talk back to it exactly what people want

to do is potentially talk to their pets

the this does raise huge ethical issues

in terms of how we are interacting with

animals and whether we should should you

I don't want to be asked for a biscuit

every five

minutes absolutely not but I've talk to

the whales I don't want my dog to tell

me that it prefers my husband to me

that's

I the worst one yeah exactly but the

thing is so so obviously I think there's

a lot we can learn as well um but

there's going to be huge questions about

this but lots of people working on bats

whales elephants and the Ft has done a

fantastic podcast on it and I suggest

you listen to it oh go have listen to

that it's a great story uh

congratulations to Sky News who actually

picked that up with the scientists uh

it's a really good one though um that's

it from us we're out of time we'll do

this again same time next week just

before we go let me remind you that all

the programs that we do are on uh the

BBC's YouTube channel uh all the big

interviews lots of good people coming on

this program now so if you want to look

back at previous episodes go and have a

look at that