- il y a 2 jours
The AI-energy Nexus: From Silicon To Sustainability
Catégorie
🤖
TechnologieTranscription
00:00Bonjour, tout le monde. Merci d'être ici avec nous.
00:03Je suis très heureux de pouvoir discuter avec vous et de cette session
00:09sur les needs de l'énergie de l'énergie, pour l'électron,
00:13et peut-être aussi sur la façon dont l'AI peut nous aider à sauver l'énergie.
00:20Donc, peut-être que nous pouvons commencer par les needs de l'AI dans le futur.
00:26Quels sont les forecasts ?
00:27Which are the countries able to provide the amount needed ?
00:31Is décarbonise l'électron a must-have ?
00:34And is it really worth developing such an energy-consuming industry
00:39at a time when resources are at stake ?
00:43So, maybe we can start by, what are the real needs of AI
00:47regarding electricity and électron ?
00:51Maybe you can start with ?
00:52Yeah, with pleasure.
00:53Hello, everyone.
00:53So, first, yes, there is a super-strong increase of demand
00:59between from 5 times 5 to times 10 of energy needed for AI
01:05between now and, let's say, 2030 or 2035.
01:09However, I always like to put things back in context.
01:13If you take the total energy consumption of the world,
01:16every year, it's 120,000 terawattres.
01:20Today, data centres is 400 terawattres, out of which 10% is AI.
01:25So, today, it's 40 terawattres out of 120,000 total energy consumption.
01:31Only?
01:31Only today.
01:32Not on electricity, on all kinds of energy.
01:34Electricity is 20,000 terawattres.
01:36So, it's 40 out of 20,000.
01:38Okay.
01:39The projection, the big one, says probably times 10.
01:42Three times more data centres, three times more workload for AI in data centres.
01:47So, 600, 700 terawattres, which is a lot of energy.
01:50But still, compared to 20,000 terawattres of electricity,
01:54or 120,000 terawattres total.
01:57Which means, for me, that one of the key questions we need to address is,
02:00what do we do with this AI if we use the 600 terawattres to optimise the 120,000?
02:08It's a good game.
02:09If we use it for something else.
02:12We will speak later about the way you can reduce energy consumption with AI, of course.
02:19Maybe, Emi, I'm sure when you design chips with IRM,
02:24sometimes for NVIDIA, your former employee,
02:28is it something you're looking at very carefully,
02:31which is what amount of energy I will need for making my chips processed?
02:38Yeah, no, that's a good question.
02:39So, yeah, nice to be joining this panel.
02:42I think if you look at,
02:44if you sort of just look at the forecast, as you mentioned before,
02:47if you look at what, by the next decade,
02:49we'll probably have 2 to 3x the amount of energy needed,
02:53and 10x the amount of infrastructure investment.
02:55And you look at sort of what's driving a lot of that.
02:58If you think about ChatGPT,
03:01it requires just a single query is 0.3 watts.
03:05To put that into perspective,
03:07if you look at charging your mobile phone,
03:10it actually requires 2 to 3 ChatGPT queries.
03:13So, a huge amount of energy consumption.
03:16And then you look at what's driving that energy consumption.
03:19The main thing that's driving that energy consumption is compute chips.
03:23So, actually, 50% of the amount of energy consumption
03:27for any of these language models
03:28is really the compute chip itself.
03:31And that's really where ARM comes in.
03:33The genesis of our company has really been in power-efficient chips.
03:37And so, we're really at the forefront
03:40to designing these power-efficient chips for data centers,
03:44and as inference moves to the edge,
03:47how are these efficient chips going to move to the edge?
03:49So, we're really looking at this problem in multiple dimensions.
03:54Olivier, you are part of a very ambitious group
03:57regarding AI and cloud and data center,
04:01which is the group 42.
04:03You are called 42, which is a part of it.
04:06Obviously, you are looking for energy
04:09because you are investing in continents
04:11and in countries like France,
04:13where you apparently think there are plenty of it,
04:18decarbonize one.
04:19So, that means that you are looking for energy.
04:21So, you are energy consuming.
04:23So, COF 42, as you said,
04:26so, we are part of group 42.
04:27We are deploying the infrastructure
04:29that AI requires to work
04:31and to provide so-called tokens
04:33that Jensen talked yesterday.
04:36In perspective,
04:37the first element of our question,
04:40the need for AI in terms of power consumption
04:43is 600 billion of watts,
04:46which is, I mean,
04:49enormous amount of electricity we need.
04:51So, what we deploy in infrastructure,
04:52we need to be cost-effective
04:54in terms of what the energy consumption,
04:58because this is as a cost,
05:00and we need to go to bring down the cost per token.
05:03So, we need to be sure
05:05that the technology we are deploying
05:07from ARM, from NVIDIA,
05:09we use it efficiently
05:11in terms of power consumption
05:13to be cost-effective.
05:15Maybe we can elaborate on this question
05:18about France's attractivity
05:20and capabilities about producing
05:23and making energy for data centers.
05:25We know that our government
05:27is keen to attract those data centers.
05:29There are a lot of projects
05:30and AI capacity in France.
05:33Do you think, the three of us,
05:36we will have a fierce competition
05:38between electricity users,
05:41like of course, data centers,
05:43but also industries and consumers?
05:45And can we cope
05:47with those types of competition?
05:51The answer is probably yes.
05:54However, I don't think we have a choice.
05:56We don't think we have a choice
05:57because I hope everybody in this home
05:59is convinced that we need
06:01to decarbonate the energy.
06:02And the only energy
06:04that can be fully decarbonated
06:05is electricity.
06:06So, yes, the world will move
06:08towards more electricity, I hope.
06:10Otherwise, the world will move
06:12towards more global warming.
06:13So, yes, there will be a need
06:15for more and more electricity
06:16for AI, for data centers,
06:19for factories,
06:20to replace gas-fired oven
06:22by electrical-fired oven.
06:24So, yes, we need to get ready.
06:25And we need to get ready
06:26not only by having the production,
06:28but the capacity to transmit
06:30and distribute this electricity
06:31and also make sure
06:33that this electricity
06:33is as much carbon-free as possible.
06:36Because if we create electricity
06:37by burning oil,
06:39we need to save the problem.
06:40So, yes, the answer is clear.
06:41There will be more and more demand
06:42for electricity in the future
06:43for AI and mostly
06:45for decarbonation globally.
06:47Amy, how do we cope
06:48with this competition
06:49between users of electrons
06:51and electricity?
06:53Yeah, I think, you know,
06:54if you look at, you know,
06:55my belief is that you look at AI
06:57and if we use it
06:58for the right ways,
06:59I think the future
07:01will actually be better
07:02than the present.
07:03And so that will obviously
07:04require a lot more energy.
07:05We're going from, you know,
07:07megawatts that, you know,
07:08we're five years ago
07:09probably two years ago
07:10at this point
07:11to now gigawatts,
07:12obviously the first few
07:13gigawatt data centers
07:14that were built.
07:15In terms of distribution of that,
07:18as I mentioned,
07:19I think the last several years
07:20have been focused
07:21around large language models
07:22mainly in the data center.
07:24But I think that distribution
07:25is going to change.
07:26You look at by the end
07:28of this decade,
07:29at least 70% of inference
07:31will happen on the edge.
07:32And so that distribution
07:34of data center
07:35won't be exclusively
07:37in the data center.
07:38it will actually be
07:39at the consumer.
07:40You'll be inferencing
07:41on your mobile device
07:42as an example.
07:43And so that economic equation
07:46will actually change
07:47from the data center only
07:49to both your edge device,
07:52whether it's your, you know,
07:53earbuds that you put
07:54in your ears,
07:55whether it's your mobile phone
07:56that you're interacting with,
07:58to the data center.
07:59So you actually have to think
08:00about energy in a different way
08:02than just the grid.
08:04You have to think
08:04about renewables.
08:05You have to think about
08:06where you're going to get
08:07other energy sources,
08:08putting energy closer
08:10to the consumer,
08:11for example.
08:12So I think that equation
08:13is going to look very different
08:14over the next decade.
08:18Olivier, as I said,
08:19you are investing
08:20in many countries
08:21and about infrastructure,
08:24about data center,
08:25about the PRT,
08:26the AI model, etc.
08:27What kind of guarantee
08:29are you looking for
08:30when you invest somewhere
08:32about the capacity
08:34of power and grid?
08:37What are you looking at
08:38and what kind of guarantee
08:40do you get from the government,
08:42for instance,
08:42or from the big power company?
08:45So we are looking at countries
08:47where there is a possibility
08:49to have access
08:50to a lot of energy
08:51because when you saw
08:53some announcement
08:54which has been made,
08:55like one gigawatt is available
08:56and we will deploy
08:57one gigawatt.
08:58and a few weeks ago
09:00when President Trump
09:01came to UAE,
09:02we said five gigawatt
09:03to be deployed.
09:04But it's a perspective.
09:06It's up to.
09:07And we have to go by stage
09:08because it's impossible
09:09for your information,
09:11one gigawatt
09:12equal 500,000 GPUs
09:14to be deployed.
09:15It cannot be deployed at once
09:17and there is no workloads
09:18able to use 500,000 GPUs today.
09:21It's impossible.
09:22It doesn't work.
09:23It will not fly.
09:24Okay?
09:24And to come back
09:25to your initial question,
09:26there is no competition
09:27between AI
09:28and the traditional industry.
09:30Why?
09:30Because the traditional industry
09:32will have to use AI
09:34so the power consumption
09:35will not be against,
09:37it will be within
09:38those energy
09:39and those industries
09:41because every single industry
09:44has to use AI today
09:46or will use AI in the future.
09:49That's a good transition
09:50because we talk about
09:51AI as a problem for energy
09:53but we can talk now
09:54about AI as a solution
09:56for the consumption of energy.
09:59So maybe let's give
10:00some very practical examples
10:02because in, for instance,
10:03in Scheller Electric,
10:04Philippe,
10:05how do you use AI
10:06in all your instruments
10:07and apparels
10:11and solutions
10:13to save this energy?
10:15So clearly, as I said
10:16and as you say,
10:17we need to move
10:18to electricity
10:18and the demand
10:19will be in the industries,
10:21in the buildings,
10:21in the factories
10:22and the energy
10:24which impacts less the planet
10:25is the energy
10:26that you don't use
10:27for obvious reasons.
10:28And we can use AI
10:29to use less energy
10:30for the same outcome.
10:31And I'm sorry,
10:32I'm going to do
10:33two minutes
10:33of hardcore science
10:34but if you want
10:36to optimize the process,
10:37if you want
10:38to use less energy
10:39to heat or cool a building,
10:41heat or cool a room,
10:42you need to have a model
10:43to simulate what happens
10:45in your factory,
10:46in your building.
10:47That model,
10:48the best one,
10:48is usually built
10:49using physics laws
10:50but very often
10:51you can't do that
10:52using machine learning.
10:53And traditional AI technique
10:54that, by the way,
10:55use much less energy
10:57to generative AI,
10:58you can build these models
10:59and use them
11:00to optimize.
11:01A room like this one
11:02using a controller
11:04that has AI in it
11:05can save 20% energy
11:07on how you heat
11:08and cool the room
11:09and that's 30%
11:10of the carbon emission
11:12of buildings
11:12heating and cooling
11:13simply because
11:14by AI
11:15you are able
11:15to machine learn
11:16the law
11:17that how the temperature change
11:21based on how much energy
11:23you bring in
11:23or out of the room.
11:24So that's a very
11:25special example,
11:26a basic room controller
11:27with a simple AI in it
11:29and you save 20% energy.
11:31The other thing
11:32which is very important
11:32in the world of tomorrow
11:33with a lot of data centers,
11:35a lot of electrified factories,
11:38lots of renewables,
11:39what is important
11:40in the world of electricity
11:41is a big average demand
11:43but what is extremely important
11:45is a peak demand
11:46because when everybody
11:47wants energy
11:48you need always
11:49on the grid
11:50to balance
11:50electrical supply
11:51with electrical demand.
11:53And in a world
11:55with a lot of renewable,
11:56the only thing
11:56that you can start
11:58or stop easily
11:59is not a solar panel,
12:00it's a gas fire power plant,
12:03a fuel fire power plant.
12:04so at peak demand
12:05the electricity
12:06is expensive
12:07and carbonated.
12:09With AI
12:10you can help with that again.
12:11How do you do that?
12:12Again,
12:13take this building
12:13and those solutions exist.
12:15They can be bought
12:16in Schneider
12:16and in competitors
12:17but it's not a dream,
12:18it exists.
12:19You have solar panels
12:20on the roof
12:20because you want
12:21to be a green building.
12:22If you use
12:23what your panels produce,
12:24it's great,
12:25you contribute to the planet.
12:26It's probably not optimum
12:28because in reality
12:29when the sun is shining
12:30like today,
12:31it's shining all over Paris
12:32and the electricity is green
12:35and abundant.
12:36Or use AI
12:37and with AI
12:38you forecast every 15 minutes
12:40the consumption
12:41of your building
12:41for the next 48 hours.
12:43You forecast the production
12:45of your panel
12:45for the next 48 hours
12:47and then you optimize
12:48and you decide
12:49if right now
12:50it's better to use
12:51what you produce,
12:52sell it to the grid
12:53because the grid
12:53is very carbonated,
12:55store it in your batteries
12:56for tomorrow
12:56because tomorrow
12:57the grid will be carbonated,
12:58etc.
12:59and that of course
13:00cannot be done
13:01with a human being.
13:01nobody can stand
13:02in front of a computer,
13:03look at all of that
13:04and decide.
13:04AI can do it
13:05and again
13:06we can help by that
13:07decarbonating the energy.
13:08So yes,
13:09we need a lot of energy
13:10for AI
13:10but AI can also
13:12help solve
13:13a lot of energy.
13:16And just to give
13:17one final number
13:18you've seen
13:18from my first discussion
13:19that I love facts and data
13:21I'm an engineer.
13:22Today we have checked
13:23that what we do
13:24for energy saving works
13:26which means that
13:27the carbon emitted
13:28to run the model
13:29and again
13:30those are simple models
13:31not generative AI
13:32for one ton of carbon
13:34emitted by running the model
13:36we save between 300 to 600 tons
13:38of carbon
13:39by not heating,
13:40not cooling,
13:41buying energy
13:42when it's green,
13:42etc.
13:43So yes,
13:44AI will also
13:45tremendously help
13:46on energy savings.
13:47So the proportion
13:48you calculate
13:49between energy
13:51used by AI
13:53and energy saved
13:54with AI.
13:55I don't know if you were there
13:56when Christelle
13:57was speaking from Orange
13:58she was speaking
13:58of frugal AI.
13:59That is extremely important.
14:01We need to choose
14:02the most energy efficient
14:05way of doing things,
14:06the most simple AI
14:07for the task
14:07we want to do.
14:08What I described to you
14:10we could do it
14:10with generative AI
14:11but we can do it
14:12with much more simple AI
14:13that uses
14:14100 times less energy.
14:16Amy,
14:17you're working
14:17for one of your customers
14:19is of course NVIDIA
14:21because you're designing
14:22one part of their chips.
14:25This is probably
14:26the company
14:27and the chips
14:28who are the more
14:29energy consuming
14:30and we know
14:31that when you make
14:32a quest to LLM
14:35it's like
14:36100 times more
14:37consuming
14:38than Google search.
14:39So what do you do
14:40to be able
14:41to reduce
14:42this amount
14:43and how do you design
14:45your chips
14:47to make it
14:48the lowest possible?
14:50Yeah, you know
14:50so if you sort of
14:51look at
14:52you know
14:53if you look at
14:54some of the data center
14:55a place that we
14:56we play in
14:57we partner with folks
14:58like NVIDIA
14:58with Grace Blackwell
15:00actually Grace
15:00is all based on ARM
15:02and paired with
15:02NVIDIA's GPU
15:03then you look at
15:05Amazon, Google, Microsoft
15:06they're all basing
15:07their chips on ARM
15:09and the reason for that
15:10is if you look at
15:10in totality
15:11most data centers
15:12we've been able
15:13to decrease
15:15energy efficiency
15:16by
15:16or decrease
15:18energy by 40%
15:19by using ARM
15:20so it's a dramatic
15:22increase
15:23in terms of
15:24energy efficiency
15:25and so
15:26most data centers
15:27are actually now
15:27building with ARM first
15:29and that's primarily
15:30due to
15:31chip design
15:32and chip design
15:33being more efficient
15:34and so the way
15:35we do that
15:36is actually
15:36the layout
15:37of the chip itself
15:38we're able to
15:39lay out the chip
15:40in such a way
15:41that you're able
15:41to get maximum
15:42performance
15:43and perf per watt
15:44with that design
15:46and so that's one
15:47of the things
15:47that we're doing
15:48whether it's
15:48custom silicon
15:49with our
15:50hyperscale partners
15:51whether it's
15:52at the edge
15:52where we talk
15:53about mobile phones
15:54if you look
15:54at the origination
15:55of the company
15:56we actually
15:57started with
15:59mobile phones
16:00and powering
16:01mobile phones
16:02off of batteries
16:02so we're taking
16:03that same
16:04architectural concept
16:05and designing
16:06those same chips
16:07in the data center
16:08so now you can
16:09actually put
16:10our energy
16:12efficient chips
16:13in the data center
16:14and pack more
16:15compute in that
16:16same data center
16:17with being more
16:18energy efficient
16:19so it's a lot
16:20about how we design
16:21the chip
16:22with our ecosystem
16:23partners
16:24Olivier, how would you
16:26convince us
16:27that your solution
16:28and your data center
16:29and your infrastructure
16:29will help us
16:31to be more
16:32energy efficient
16:36so an infrastructure
16:37an AI factory
16:38is a supercomputer
16:39meaning that
16:40a workload
16:42is using multiple
16:43nodes inside
16:45to work
16:46one node
16:46so if one component
16:47of the node
16:48is failing
16:48then you lose
16:49your job
16:49you need to restart
16:50so what we have
16:51to do
16:52is to do
16:52like
16:53kind of
16:54put an AI agent
16:55inside the infrastructure
16:56to do preventive
16:58maintenance
16:58to understand
16:59that if a node
17:01has a chance
17:02to fail
17:02not to use
17:03to run
17:04that job
17:04so try to be
17:06efficient
17:07in the workload
17:08usage
17:09of the infrastructure
17:11maybe we can
17:12sorry
17:13yeah
17:13no try to be
17:14shorter
17:14than my colleagues
17:15okay
17:17maybe we can
17:18take another example
17:19we saw
17:20and the other panel
17:22talked about it
17:22a few minutes ago
17:24but the
17:26Spain blackout
17:28of course
17:29it's a very
17:29interesting example
17:31I don't think
17:31we know everything
17:32about it yet
17:33but
17:34it's very
17:35instructive
17:36and very
17:36interesting
17:38to think
17:38about it
17:38what happened
17:39and how AI
17:41can help us
17:42to manage
17:42all those
17:43decentralized
17:44ways of
17:44making electricity
17:45which are
17:46solar
17:46and wind
17:47electricity
17:48are we able
17:49to manage
17:50such a
17:51decentralized
17:53organization
17:53of the grid
17:55I'm sure
17:56I'm sure
17:56in Schneider Electric
17:57you
17:58talked a lot
17:58about that
18:00I think
18:01the first
18:02lesson
18:02we can learn
18:03from the
18:03Spain blackout
18:04is
18:05we really
18:06need electricity
18:07when electricity
18:08is not there
18:09life becomes
18:10suddenly
18:10very very
18:11complicated
18:11so maybe
18:13that's the first
18:13learning
18:14the second
18:15learning
18:15and honestly
18:16I don't know
18:17the detail
18:17of what happened
18:18but the second
18:18learning is
18:19and to some
18:20extent
18:20we knew it
18:21we have moved
18:22from an
18:23electrical world
18:23where you had
18:24big production
18:25plant
18:25transmission
18:26networks
18:27distribution
18:28network
18:28consumers
18:29that was
18:30first
18:31easier to
18:32operate
18:32and I cannot
18:33explain too
18:34much
18:34technically
18:35but
18:35that was
18:36almost
18:36self-equilibrated
18:37to some
18:38extent
18:38when an
18:39issue was
18:39happening
18:40we have
18:40moved to
18:40a situation
18:41now
18:41where we
18:42have
18:42production
18:43everywhere
18:44solar panels
18:44wind
18:45generation
18:45this production
18:47is not
18:48controllable
18:48you cannot
18:49decide
18:49to start
18:50a solar panel
18:51you cannot
18:51decide
18:51to start
18:52a wind
18:52generation
18:53and also
18:54and again
18:54that's the same
18:55technical point
18:56it's not
18:56self-equilibrated
18:57which means
18:57that it is
18:58not self-regulating
18:59which means
19:00that you need
19:00to have much
19:01more control
19:01of your grid
19:02you need
19:03to much more
19:03decide
19:04what to stop
19:04what to start
19:05what switches
19:06to turn
19:06off and on
19:07and the grid
19:08is extremely
19:08complicated
19:09and here
19:10AI is playing
19:11or will play
19:12a key role
19:12to make
19:13the right decision
19:13and it will
19:14play the key role
19:14in many different
19:16at many different
19:17levels I would say
19:17first
19:18it will help
19:19globally
19:20forecasting
19:21production
19:22and consumption
19:23to always manage
19:24this balance
19:26power
19:27consumption
19:27because just to be
19:28clear for everybody
19:28in the room
19:29when the grid
19:30gets out
19:30it's always a problem
19:32of balance
19:32and demand
19:33always a balance
19:34of demand
19:35and production
19:35so forecasting
19:37optimizing
19:37to make sure
20:01the products
20:02the assets
20:03on the grid
20:03work well
20:04and for how long
20:05and the third
20:06one
20:07is really
20:07to use AI
20:08to make all
20:08this automated
20:09decision
20:10of
20:10I take the power
20:12coming from
20:12this solar panel
20:13I stop it
20:14I balance power
20:14from this region
20:15to that one
20:16and it's going to
20:17become less and
20:18less humanly possible
20:19without a strong
20:20supporter AI
20:21to automate
20:21those decisions
20:22make it fast
20:23and especially
20:24I don't know
20:24what happened
20:25in Spain
20:25but if things
20:26would have worked
20:27well
20:27you lose
20:27a village
20:28you lose
20:29a neighborhood
20:29for losing
20:30a whole country
20:31there was a cascade
20:32of events
20:33that were not
20:33controlled
20:34fast enough
20:35I'm sure
20:35so the storage
20:36issue
20:37will be critical
20:38in the next year
20:39no?
20:39super important
20:40one
20:40super important
20:41one
20:41if we find
20:42a solution
20:43to storage
20:43most of the issue
20:44is solved
20:45we will find
20:46a partial solution
20:47to storage
20:47because we will not
20:48be able to
20:50build batteries
20:50big enough
20:51but even
20:52with some batteries
20:53AI
20:53and the balance
20:54between production
20:55and demand
20:55we can do
20:56a lot of things
20:57at local level
20:58at home level
20:59at district level
21:00at countries level
21:01definitively
21:02Amy
21:02I'm sure
21:03that the
21:03Spain blackout
21:05also initiate
21:06reflection
21:07about ARM
21:08because of course
21:09it's an issue
21:10for all chips
21:12and all data centers
21:14because it's
21:15probably
21:16you only
21:16once you invest
21:19and once you install
21:20your system
21:20you only need
21:21one thing
21:21it's electron
21:22so if there is no
21:23electron
21:24what do you do?
21:26yeah
21:26that's a good question
21:27so I think
21:28you know
21:29to your point
21:29it's all about
21:30responsibility
21:31and really
21:32balancing
21:33sustainability
21:34with the need
21:35for more energy
21:36and so
21:37it's a fine
21:38balance
21:38around
21:39achieving that
21:40and not getting
21:41to a point
21:42where we have
21:42the Spanish blackout
21:43so some of the things
21:44as you mentioned
21:45in your earlier question
21:46around what are we doing
21:47from an AI perspective
21:48and we have
21:49kind of a different
21:50vantage point
21:50obviously being
21:51at the chip level
21:52and designing
21:53the overall system
21:54with our partners
21:55where we're using
21:56AI to actually
21:57design chips
21:58and so you sort
21:59of think about
21:59why are you using
22:00AI to design chips
22:02because that's
22:02the most computationally
22:04complex
22:04and requires
22:06the most compute
22:07in order to actually
22:08design the chip itself
22:09so by using AI
22:11you actually can
22:12reduce the amount
22:13of energy consumption
22:14because at the end
22:15of the day
22:15eventually over time
22:17AI will help
22:17will do that themselves
22:19and I also think
22:20at the data center level
22:21data centers
22:23will actually be
22:23designing themselves
22:24in a period of time
22:25it won't happen
22:26in two
22:26three years time
22:27but you fast forward
22:29to five years
22:29and the progress
22:31we're making
22:31with AI
22:32I truly believe
22:33that will help
22:34drive some of the
22:35innovations
22:36in terms of
22:37energy consumption
22:38and preserving
22:39some of that energy
22:41Olivier what is
22:42your analysis
22:44and evaluation
22:45of what happened
22:45there in Spain
22:46and what are the
22:48solutions
22:48you could deploy
22:50to avoid
22:51and to rule
22:53and to cope with that
22:57AI requires infrastructure
22:59requires electricity
23:00you remove electricity
23:02you cannot run infrastructure
23:03you don't have AI
23:04so
23:06and honestly
23:07I tried to read
23:08what happened in Spain
23:09I have no clue
23:10and I don't know
23:11if somebody knows
23:12what happened exactly
23:13but it makes
23:15some people
23:16some companies
23:17who have the
23:18power
23:19investment power
23:20to look at
23:21ok let's put
23:22a nuclear plant
23:24in my data center
23:25to run my AI
23:26and becoming
23:26autonomous as well
23:28which
23:29well the idea
23:31might be great
23:31but practically
23:33I'm not sure
23:34it will happen
23:35anytime soon
23:35to be honest
23:36maybe
23:37I've got one question
23:38to all of you
23:39about the way
23:40you used AI
23:42in your companies
23:44to make
23:45saving
23:45on energy
23:46maybe we can start
23:48with you again
23:49Philippe
23:49we used what
23:51I was saying
23:51if you come to
23:53our flagship buildings
23:54the newest one
23:55we have built
23:56intensity in Grenoble
23:57we use AI techniques
23:59to optimize
24:01locally
24:01what we have
24:03in terms of
24:03power production
24:04power consumption
24:05we also use
24:07that in combination
24:08with our neighbors
24:09because actually
24:10sometimes we need
24:11less energy
24:12they need more
24:12so we can do
24:13better optimization
24:14and we add
24:15to that
24:16a lot of techniques
24:18from darkening windows
24:19to automatic detection
24:21of people
24:22and so on
24:22at the end
24:23this building
24:24consumes
24:2510 times less energy
24:27per square meter
24:28per user
24:29than a good building
24:31already
24:32to some extent
24:33you know
24:34we have the solution
24:35big part of the solution
24:36not the full solution
24:37but the big part
24:38of the solution
24:39of fighting climate change
24:42already today
24:43and maybe one thing
24:44where
24:44talking to you
24:45I'm talking about
24:46another idea
24:46one thing where
24:47AI can probably help
24:48also
24:48is how can we
24:49scale those solutions
24:51you say use AI
24:51to compute faster
24:52the design of the chips
24:54so you can do more chips
24:55and better chips
24:56actually the solution
24:57I described
24:58we need to find a way
24:59to bring in the market
25:00at a cost
25:00which is acceptable
25:01and here AI can help
25:03I will take a very basic example
25:04you come from California
25:05I think
25:06lots of center California
25:07the government is pushing
25:09to electrify
25:10put solar panels
25:11EVs etc
25:12the first step
25:14that you need
25:15is you need an electrician
25:16to come to your home
25:16open your electrical panel
25:18and check if you can do it or not
25:20there are not enough electricians
25:22in California to do that
25:23that is slowing the deployment
25:25of renewables
25:26with AI
25:27you have made a tool
25:28where by yourself
25:29you open your panel
25:30you take a picture
25:31and the AI tells you
25:33yes it works or not
25:34so AI is also a tool
25:36that can be used
25:37to reduce the cost of adoption
25:39reduce the cost of deployment
25:40of many of the techniques
25:41that will help
25:42with energy transition
25:43that will help fight climate change
25:45so I see AI helping reduce
25:47helping removing the peak
25:49and also helping building
25:50data centers more efficient
25:51she's more efficient
25:53less electrician visits
25:55in California
25:55but with AI
25:57we can deploy all of that
25:58at a much lower cost
25:59which obviously will help us
26:01I mean is AI useful
26:02to do that too
26:04to decrease consumption
26:05and to be able
26:06to find solutions
26:10yeah
26:11no you know
26:12absolutely
26:13and I think
26:13we'll be able to
26:15hopefully
26:15with all the future advances
26:17of what we're using
26:18applying AI for
26:19I think in the next year or two
26:20we'll probably see
26:21many more innovations
26:22around how we can use energy
26:24to decrease
26:26the amount of tax
26:27on the grid
26:27back to your earlier question
26:29around how we're using AI
26:31within our company
26:32for energy consumption
26:33or energy reduction
26:34and you know
26:35in our company
26:36it's where are we not using AI
26:37we're pretty much using it
26:38in pretty much
26:39every facet of the company
26:4090% of our company
26:42is based on engineers
26:43that write instructions
26:45for chips
26:45design you know
26:46the actual instructions
26:48that go into
26:48pretty much every chip
26:49that's used on the planet today
26:51and we're actually using AI
26:53to write a lot of that code
26:54and so we've seen
26:55you know
26:56a great deal of efficiency
26:58with actually using AI
26:59specifically for writing code
27:01and we'll continue to see
27:02advances in that space
27:03Olivier we have a few seconds left
27:05maybe we can go on
27:07the decarbonized aspect
27:08of energy
27:09do you
27:10is it still a priority
27:11when you develop
27:12your infrastructure
27:13and your energy
27:14and your data center
27:16to find
27:16the right electricity
27:18the right decarbonized electron
27:19of course
27:20sustainability
27:20is a key driver
27:22we cannot waste electricity
27:24we cannot waste energy
27:27okay
27:27even if coming from
27:29UAE
27:30United Arab
27:31Emirates
27:32and group 42
27:33has been
27:33to answer your question
27:34group 42
27:35has been created
27:36to serve a vision
27:36the vision was in 2016
27:38but by 2030
27:4020% of the GDP
27:42of the Emirates
27:43should be generated
27:43by AI
27:44using and reselling
27:46AI technology
27:47so this is the purpose
27:48of a company
27:49so yes
27:49we are using
27:50and developing AI
27:51for the goods
27:53of the Emirates
27:54and now the world
27:56thank you very much
27:57to the three of us
27:58and thank you
27:59thanks to all
27:59bye bye
28:00to the
28:05Merci.
Commentaires