- il y a 2 mois
AI & Energy The New Power Couple
Catégorie
🤖
TechnologieTranscription
00:00As it has just been mentioned, AI is going to be big, and if it's big, it's going to need
00:05a lot of energy.
00:06So basically, the question we're going to try to answer today is, is AI the enemy or a friend of
00:13the climate?
00:14And the two guests have been introduced, Luc from EDF and Anna from IBM.
00:20Luc is big in energy, nuclear power, and Anna at IBM is a huge player in IT and especially in
00:28servers.
00:29So you build, both of you, the infrastructure that will be very crucial for the success of AI.
00:36Let me start with just maybe a question about figures or how big is the challenge when it comes to
00:44AI in terms of growth, in terms of needs of energy?
00:48What are the key metrics maybe we should have in mind when we talk about AI?
00:51Anna, if you want to start?
00:53Yeah, let me start.
00:54Let's think about the reality that we have been, you know, data centers, been growing in the past few years.
01:02And right now, the figures that we're considering are about 1 to 1.5% of all the energy that
01:08we consume in the world is directed towards data centers.
01:12Now, we have a new application coming up, right?
01:15And that's artificial intelligence that is completely transforming the way the world operates, that has a tremendous impact on how
01:22we consume data, on the growth of applications that we are going to see coming up in the next few
01:27years.
01:28We have a prediction that by 2028, there will be more than 1 billion new applications that are going to
01:33develop through AI.
01:35And that will, of course, increase the demand of energy for data centers.
01:39Right now, our prediction is that every 100 days, the needs of energy to power AI is going to double.
01:47So, growth of about 25% every three months, which is significant.
01:53Huge opportunity, maybe for those who sell electricity, but do you always already see the impact of the need for
02:00electricity for computing powers in France or in other part of the world when you talk with your friends of
02:05the energy sector, Luke?
02:08Well, the use of electricity for data is already something that does exist and is visible in the portfolio of
02:15electricity users since a couple of decades.
02:18It appears to be quite concentrated on not that many countries so far, which means that the infrastructure for data
02:26started, of course, in the mature markets, and now it's rising in emerging markets as well.
02:34What AI is bringing is an acceleration of this type of needs, a demultiplication, which is going to come in
02:40the next decade.
02:41But it starts on an existing infrastructure that has been built over the last two decades.
02:46We were talking about data centers that were central, talking in hundreds of megawatts at the maximum.
02:54Now, we are going to reach data centers that probably will reach the gigawatts of needs in a stable baseload
03:02need of electricity.
03:03That's the change.
03:04And, of course, for countries, when you look at the needs of electricity at the country size, you will need
03:09to address this need in the middle of the other needs of electrification.
03:14That is a global trend that we see everywhere for decarbonation.
03:18So, in a few words, the need will be more power for a stable baseload for AI and data in
03:25an environment that will need stable baseload, decarbonated power.
03:29In terms of challenges, is it a big challenge, a huge challenge to be able to cope with this increasing
03:37demand in AI computing power and energy consumption at the same time?
03:42How big a challenge is it, do you think, Anna?
03:45The challenge is big, but at the same time, it's very optimistic to see that the evolution of the technology
03:53to cope with those challenges is also progressing very fast.
03:56So, you can use AI today to predict demand, to forecast how the energy consumption is going to affect the
04:04grid.
04:04We are working, for example, with a company in Australia, the Western Power, to project in the next 30 years
04:12what is going to be the curve of consumption that their base of users is going to present.
04:19Other techniques that are very interesting, so, for example, today the consumer of energy can also be a supplier of
04:27energy if you create these two-side markets in the way that you manage the grid.
04:34And then other capabilities that I think, in addition to AI and combining AI with quantum, for example, to identify
04:41new materials, to identify new sources of energy that can be, you know, green, that can be renewable.
04:48I think if we add all these advancements that are happening, we are going to address this challenge much faster
04:55than it's going to create a problem for us.
04:57Okay. And on your point of view, Luc, in terms of a challenge, producing the electricity for this revolution is
05:05going to be a challenge or it's still also a manageable opportunity for you?
05:10Well, it will certainly be a challenge for some and an opportunity for others.
05:14And the opportunity will be for the regions of the world who have available low-carbon electricity, and we are
05:23fortunate to be in that category in France because of our 57 nuclear power reactors plus renewables,
05:30making our mix in this country 97% low-carbon already, and the average of carbon emission for EDF group
05:40at 37 grams of CO2 per kilowatt-hour, where the average in Europe is 251 and the average in the
05:48world is 451.
05:50So, when you start with this, with available power, by definition, this is an opportunity because where the infrastructure will
05:58be, the users will be, which means that it will develop a new economy that is enabled by AI.
06:04So, I see this for France as a great opportunity, and of course, as EDF, we are ready to facilitate
06:11and welcome all the users of that sort.
06:13For other regions which may not have low-carbon electricity, it's going to be a bigger challenge, and that's probably
06:20one of the reasons why we see more and more countries taking a very wise approach of their future of
06:26electricity,
06:27balancing the intermittent sources of low-carbon electricity with renewables, with dispatchable, commandable ones like hydro and nuclear.
06:36But at the same time, you were mentioning that the global society is moving into a form of electrification.
06:43So, in terms of investments, aren't we facing a mountain of investment to come in the renewing of our electricity
06:52production in France, but also in other countries, to face AI needs, but not only?
06:57We are certainly facing an amount of investment that is significantly ahead of what has been done in the last
07:06two decades.
07:07Still ahead of what was done at the very beginning of the electrification of the countries, but compared to the
07:13amount of investment that is made on the AI itself, it's not that big.
07:17I mean, this is, you know, we just look at the electrification portion, if you see the trillions of euros
07:25that will have to be invested in AI itself,
07:28this is, of course, higher than what is just deserved by electrification.
07:33So, yes, this is fully doable.
07:35Ana, you mentioned the word quantum computing, but is the future also maybe to travel a leaner computing industry, you
07:44have, are you working on servers that would be much more energy consumption efficient?
07:50Is it possible, or the fact that you need so much computing power means that it's going to be an
07:56ever-increasing need for electricity?
07:58Yeah, one of the focus of our research in the past few years was exactly to create more efficient computers,
08:05right?
08:05So, computers that can run the same workload at a fraction of previous generations required.
08:11So, that's certainly one area.
08:13When you think about AI in particular, I would touch on two aspects.
08:18First, we will not need massive large language models for all applications.
08:23So, if I would give an advice today, the most environmentally conscious decision that somebody can take when it comes
08:32to AI is making the decision to use a model that is better targeted to the application that you want
08:38to run.
08:39Because super big models are not going to be the answer for everything.
08:43The third element is how you also create chipsets that are more adequate to run these smaller models, and that
08:54don't require so much energy, and that don't require so much consumption, not only of energy, but there are other
09:02resources that enter into the equation when you talk about data centers.
09:06It's the energy, it's the water to cool the systems, so there are many aspects about the environment that you
09:12need to think about when we are designing those technologies.
09:16There used to be a rule in the computing and the semiconductor industry, which was something like every 18 months,
09:23the computing power doubles for the same price.
09:26And when it comes to energy and when it comes to AI, is it going to be the same, or
09:31do you think that we are at a point where actually this moral law is not going to work and
09:35that we will need more electricity, at least until we come up with those quantum computing and the lean kind
09:41of computing that you mentioned?
09:43Yeah, one of the things that is certain is that the Moore's law, which is the law that you just
09:48talked about, is reaching its limits, right?
09:51That's why we are developing new architectures like quantum computing, because we really need to find a new path and
09:58we need to find a new model that is going to compute this massive amount of data at a more
10:04cost-efficient way, in a way that is more environmentally friendly.
10:10Electricity also has a cost.
10:12You said that even though if you compare to all the investment that we need for AI, it will be
10:16a fraction of the total investment, but still, green energy is sometimes seen as, at least because it's intermittent, I
10:26don't know if it's the word in English, but it's more costly.
10:30The new nuclear is also more costly.
10:33How confident are you in the fact that you will be able to supply the quantity, the quality, but also
10:39energy at a cost-efficient price for all the new applications that we need?
10:46Well, it starts by the energy that is already available.
10:49And today, you know, after a crisis that has been unprecedented in 22 and 23 in Europe, mainly due to
10:58a Russian invasion in Ukraine and its consequences on gas price, and due to the fact gas price comes as
11:07the very last element making the price of electricity when gas is needed to supply the whole electricity, we had
11:15this crisis for two years.
11:16But the outcome of these two years is that we are now entering into a new phase where there is
11:23enough electricity available and enough gas available for what we need to maintain mid-term prices at a very competitive
11:32level, divided by three or four compared to what we had only two years ago.
11:37And this is more or less the economics of the electricity system is not just one asset making the price
11:46of electricity is that the totality of the assets that you need to supply 24 seven in average is making
11:53the price of electricity.
11:54So this is where we are now in the 60 or 70 euros per megawatt hour for France, where this
12:02is representing a very attractive mix, fully decarbonized, and with a lot of electricity available because day in, day out,
12:10we basically export 10 gigawatts to our neighbors.
12:13So it means that we have 10 gigawatts available for more usage for all kinds of users in France for
12:20the decade to come.
12:22And in the meantime, we do a lot of capex to increase the amount of renewables for the short term
12:28and the amount of nuclear for the long run.
12:30And the cost of that, the more we do, the more we drill the cost down, as any industry, if
12:37we do a lot, we manage to get the cost down.
12:39So that's all this equation that will enable us to be available at competitive cost now, while continuing to build
12:46for the long run, a competitive and low carbon environment.
12:50Anna?
12:50Anna, we've been talking a lot about the infrastructure and the hardware, but I think it's also interesting to look
12:56at the software aspect.
12:58So one of the things that we have been doing in the past couple of years is developing technologies that
13:05optimize the consumption of computers.
13:08Through a software layer?
13:09Add a software layer, right?
13:11So for example, how you make your workloads on the clouds to run in a more effective way.
13:17How you optimize how our server is consuming the technology or the applications.
13:22So when you combine all these elements, right, making more effective and efficient infrastructure,
13:29adding a software layer that optimizes the way the workloads are running on your platforms.
13:35And then you bring AI to provide better predictability on how you're going to consume.
13:42I think that those are elements that are really going to address those challenges quicker.
13:47When you talk today to the chemical industry or those who are building battery plants for the car industry,
13:54they mentioned the cost of energy as a key element for making a decision of where to invest.
14:01And today they are saying that there is a bigger incentive to invest in the U.S.
14:05or even in China where they have coal or oil that make electricity less expensive.
14:10When you talk to those who build the huge data centers, the hyperscaler or yourself,
14:16do you see also a world dividing itself between countries where electricity will be really cheap?
14:22Maybe not that green, but really cheap.
14:24And it's where you need to invest and where the investment will come?
14:27Or do you think that there will be needs all around the world anyway?
14:30And we have only two minutes to answer to those questions.
14:32I will ask the same question to you.
14:34At IBM, we have a very clear commitment to our carbon zero goals, net zero goals.
14:41So right now we are already running more than 70% of our data centers in renewables.
14:47And that's going to be certainly our focus going forward.
14:51Luke, do you see a world dividing itself between attracting attractive countries for huge investment
14:57because they have a cheap electricity that might not be green, but that will be cheap
15:01and that will make it more efficient for the hyperscaler to invest?
15:06Well, I see a big difference between the situations you mentioned.
15:09Typically, chemical industry, this is a global industry selling globally on the market
15:15with, let's say, the US, Europe and China competing on this market as the place to welcome the industries.
15:23Among these three, Europe has a carbon price.
15:26China and the US don't have a carbon price.
15:28So, yes, we are comparing carrots and apples here.
15:33And at some point, we need to come to a level playing field on carbon price approach
15:38for these industries that are global.
15:40Now, on the data industry, there is still a rule for the data that it has to be close to
15:45the use.
15:46So you may choose Ireland versus the UK or France or Belgium versus Germany and so on and so forth.
15:53But you need, at the end, your data infrastructure, not 6,000 miles away.
15:58So, clearly, I think the data industry, on top of it, globally, has taken very, very clear commitments
16:06to go for low carbon.
16:08And we see that as a very comprehensive engagement from most of the players,
16:14which has consequences on the type of infrastructure they choose,
16:18which is why we want to help them to be successful.
16:21And do you see long-term contracts of green energy with those clients?
16:26And do they see nuclear as a green energy for them?
16:28Or sometimes there is a debate?
16:31Yes, we are happy to provide long-term contracts for anyone who wants to go low carbon, basically.
16:39And this will help, I think, to achieve the goals.
16:42And all the industries who want to go low carbon will find EDF as a potential partner for a long
16:48term.
16:49Thank you very much to you, too.
16:51And I will pass the floor to our next speaker.
16:53Thank you very much for being here.
16:54Thank you.
Commentaires