- 18 hours ago
Every digital breakthrough currently carries a hidden debt paid in megawatts. As generative models scale, a single query can devour ten times the energy of a standard search, straining local grids and bloating consumer utility costs. This massive appetite threatens to accelerate carbon emissions and deplete vital water resources, yet a paradox exists: this technology is also a potent tool for optimizing crumbling power infrastructure and discovering sustainable materials. We are at a crossroads where AI infrastructure must either fund its own transition to carbon-free power or risk becoming an ecological liability. How do we ensure intelligence doesn’t come at the cost of the planet? Is a truly "green" algorithm even possible?
Category
🤖
TechTranscript
00:07Good afternoon, I'm Jennifer Shanker, founder and editor-in-chief of The Innovator, a global
00:12publication that connects business with technology. It's my pleasure to welcome you to our panel on
00:20the Hungry AI, Fixing the Energy Conundrum. So I'd like to welcome our panelists. I'll start over here
00:30with Sid Singh, co-leader on energy and AI at the International Energy Agency. And then next to him
00:44is Philippe Grumbach, Chief AI Officer from Schneider Electric. And last but not least, of course,
00:54is Nicolas Lefebvre-Martin, Group Vice President, Data Center Acceleration and Strategic Partnerships
01:02at Engie. So I'm just going to take a minute to set the scene here. AI runs on power, enormous
01:12quantities of it. Data centers already account for about 1.5 percent of global electricity consumption,
01:20and that figure is expected to more than double by 2030, according to the International Energy Agency.
01:29The compute required to train and run frontier models is only growing, and here Europe faces
01:37a structural disadvantage with no easy fix. In major European data center hubs, Frankfurt, Amsterdam,
01:47Paris, Dublin, grid connection queues average 7 to 10 years. That is not a bureaucratic inconvenience.
01:56It's a hard physical limit on how quickly new AI infrastructure can come online. Meanwhile,
02:05EU industrial electricity prices were 158 percent higher than in the US. In 2023, a consequence of Europe's
02:18dependence on imported fossil fuels against America's status as a net energy exporter.
02:25In America, there's a backlash from consumers because data centers are seen as driving up electricity
02:33prices, depleting local water supplies, and increasing emissions as the hyperscalers are deploying
02:41gas turbines to get data centers online quickly while awaiting grid connections. In India,
02:49the government's aggressive courtship of big tech data center investment is also generating a significant
02:57backlash on the ground. Solutions that depend on new nuclear plants, fusion plants, or data centers in
03:06space do not provide quick fixes. So how do we address the energy conundrum? Let's get right into the
03:16discussion, and Sid, I'd like to start with you. Give us a global view of AI's energy demands and real
03:25-world
03:25barriers to building more data centers and connecting them to the grid.
03:30Sure. Thank you, Jennifer. It's a pleasure to be here. I was here last year as well, and, you know,
03:36we had just
03:37finished working on a major flagship publication then that looked at the global picture on energy and AI, and since
03:44I'm here exactly a year later, I wanted to actually talk about what's changed in this year, because that'll give
03:49a kind of a good
03:51overview of where we're at currently. Yeah.
03:55Okay. All right. Thank you. And you, can you discuss how tech companies are securing the energy they need for
04:05data centers, you know, and some of the green washing that's going on?
04:11Yeah, sure, Jennifer. So first, just very quickly on what's changed in the last year, because it's very, you know,
04:18revealing on the pace at which this industry is moving forward.
04:21We find that in just one year, the global investment into data centers has shot up by over 60%. So
04:31in this year, in 2026, we expected about $700 billion are going to be spent as capital expenditure just by
04:38the world's largest tech companies.
04:41This is a huge number. For context, it's already larger than the global oil and gas investment. It's pretty significant,
04:49and it's rising very, very rapidly.
04:52All of this, of course, has a consequence on the energy demand, and therefore the ability of the system to
04:57supply electricity for data centers.
05:00On the other hand, of course, we also know that AI uptake is taking place very quickly.
05:05There's a service called Open Router, which lets people call AI tokens, and we find that activity on this platform
05:13has grown six-fold in just the last 12 months. Again, very, very rapid.
05:17And as a consequence of this, electricity consumption from data centers globally has grown 17% in just one year.
05:2717% for context, the total global electricity consumption has grown by 3%.
05:32So this is far outpacing the rate at which electricity in the data centers are being consumed and increasing.
05:41And so this space at which the AI industry itself is growing comes up against the barriers that we see
05:51in terms of how real-world constraints are not letting this industry grow faster than it wants to.
06:00So the appetite to increase is huge. The appetite to grow is very, very large.
06:06But real-world barriers of grid connections, of the ability to scale up new infrastructure for data centers.
06:14My colleagues here, of course, work in that field, and they'll be able to throw some more light on it.
06:18But the ability of the industry to meet that demand is a constraint, as is even in the tech, even
06:25in the computation side, the production of memory, you know, chips for data centers itself, that is also becoming a
06:33constraint.
06:33The supply of gas turbines for onsite generation is proving to be a constraint.
06:39So all of those constraints have come into the picture, which is what's proving to be a challenge for the
06:44data center industry to overcome.
06:46Okay, thank you. So now let me turn to you, Philippe. Are we overestimating the energy problem of AI and
06:56underestimating the potential to use AI to actually fix the grid?
07:03I would probably say yes to both of those questions.
07:07You said in your introduction that AI today is estimated to be 1.5% of total electricity consumption.
07:16Knowing that electricity is one-fifth of total energy of the world, AI today is 0.3%.
07:22In the most crazy scenarios, we may multiply that by 10.
07:27At best, we're going to get to 2, 3% of the total energy consumption of the world.
07:33So, yes, it's big. Yes, it's important. Yes, it's a fast move. Yes, it's an acceleration. Yes, it takes the
07:41reaction of all of us.
07:42But no, it will not exhaust the energy of the world.
07:46Many, many other things are still using much, much more energy.
07:50So, we need to have that in mind. And then, how can AI help solving that?
07:56The first thing I would like to say and to remember is we're lucky.
08:01Or maybe we are not, but we're lucky because GPU, chips, memories run on electricity.
08:08They don't run on diesel engine. They don't run on gas engine. They run on electricity.
08:13And electricity is not always green. Electricity can be green, can be de-carbonated.
08:19We see some recent announcements in France of people choosing France to install data center because, yes, energy is abundant.
08:27Yes, energy price is at least more or less controlled.
08:31And, yes, energy is quite de-carbonated for all the reasons you know why and how we create electricity in
08:37France.
08:37So, the first thing is it is electricity-based, which is the only energy we can de-carbonate.
08:44Then, I will share a bit on what we work.
08:47If you look at the energy side and the need to be more sustainable, to de-carbonize the world and
08:52so on,
08:53most people, and they are right, will speak about production.
08:57How can we put more solar panels? How can we put more wind generation?
09:02And that is needed and that is good.
09:04And here, AI will be needed. I will not go into details, but operating a grid with solar panels, with
09:11wind generation,
09:12is much more difficult to operate than a grid without because those sources of energy are not, let's say, self
09:19-equilibrated.
09:20I will not go into details, but AI needed here first.
09:23But then, even more important, what we all underestimate is the importance to work on the demand side of the
09:30energy equation.
09:31And here, to the risk to look stupid, the energy that impacts the less of the planet, the energy that
09:40costs the less, is the one we don't use.
09:43And using AI to optimize how we cool buildings, how we heat buildings, how we power data centers, how we
09:49manufacture stuff,
09:50by using AI, and by the way, not always generative AI, a quite cheap energy consumption AI, we can save
09:57today 15 to 20, up sometime to 30% of the energy for the same outcome.
10:05The second place where AI will play a massive role is, as I said, electricity can be fully decarbonized.
10:12Electricity can be fully green, but electricity is not always green, it's not always decarbonized.
10:18Notably, and you mentioned the gas turbines, notably at peak demand, when the grid needs to provide what we want,
10:27usually, in many, many countries, the solution is to start gas turbines, diesel generators, because that's what is easy to
10:35start, response fast.
10:37So, at peak demand, electricity is expensive, and sometimes quite carbonated.
10:42With AI, again, we can help shift the demand, we can forecast the consumption, we can forecast the needs.
10:48And then, we can push the demand out of peak, and help the grid provide the green electricity, a cheap
10:55electricity, a reliable electricity.
10:57And let me go to my wildest dreams. When I see all these data centers, of course they are part
11:02of the problem, but they also install those gas turbines,
11:06they also install those batteries, and today, they install it in order to be sure that they will always operate.
11:13So, it's a reliability question.
11:15But tomorrow, in many, many cases, we could use these installations to balance the grid, and use them as reserve
11:22energy,
11:22to be even more green.
11:24So, I'll stop here.
11:26To be very clear, yes, energy for AI is a big question.
11:30We have to work on it.
11:31There is a question of grid adaptation, there is a question of locations to find, all of that.
11:35But no, we are not exhausting the...
11:39I start again?
11:41So, we are not exhausting the energy available.
11:44And yes, AI can play a massive role on the demand side, reducing the demand, and helping us being on
11:51the green side of energy consumption.
11:53Okay. Thank you, Philippe.
11:55So, now I'm going to turn to you, and I'm going to ask you, Nicolas, to give your point of
12:01view on how AI can help to balance the grid,
12:04and to tell us if you believe that the benefits outweigh the negatives.
12:11So, I think it's a very interesting and important topic, whether AI and the growth in data center load can
12:19contribute to balance the electricity grids.
12:23I think we talk a lot about the growth in the demand, but at some point, you're reaching some of
12:28the physical limits that, Sid, you mentioned.
12:30And so, having AI and data centers, in particular, play some sort of a response, a demand response, and flexibility
12:38is becoming a very hot topic.
12:40For a very long time, data centers liked to be left alone, be able to run, you know, as much
12:47as they needed.
12:48But now, because of these physical limits, there's increasingly demands on data centers to start to flex their use.
12:56There are some software solutions which are coming up. People talk about workload orchestration. You hear about NVIDIA coming up
13:03with software solutions, other actors doing so.
13:06So, this is really emerging. I think mostly because of the necessity from the grid constraints.
13:11But then, there are also solutions from the energy space, such as batteries. You mentioned batteries can play a really
13:18very critical road in being this buffer between the data centers and the grid.
13:24And so, when the grid is under significant stress, the batteries can actually play a role. So, all of a
13:31sudden, you don't need these firm connections.
13:33You can have these partially firm connections because the battery can be that buffer. So, I think there's a lot
13:39to do there.
13:40And my sense to your second point of your question on, you know, what are the pros and cons of
13:46this, frankly, I really do believe that there are a lot of low-hanging fruits here.
13:51And I don't think it'll be such a big economic impact on these actors. This is just something which they
13:56were never asked to do in the past.
13:58And now, they're being kind of led to do because of these physical constraints. And I think there are many
14:04solutions that you'll see come up.
14:05And I think there will be some very positive surprises in the month and years to come.
14:10So, let me come back to you, Sid, and ask, like, that all sounds great.
14:18But, you know, we're hearing a lot of announcements here in France and other places about, oh, you know, we're
14:28putting money into building new data centers.
14:31So, we have commitments to have these new data centers built in our country.
14:36But how quickly can that actually happen?
14:40And, you know, yes, AI can help balance the grid and maybe even bring, you know, help renewables to play
14:48an even more important role in balancing the grid by, you know, via storage and batteries, et cetera.
14:55But are we talking about short-term?
14:58Because you mentioned, like, there's this, you know, need, a tremendous pressure in the race to get these data centers
15:08up to speed.
15:09How do we overcome all of these barriers for the licensing, all of the things, you know, getting them connected
15:19to the grid, et cetera?
15:20You know, how fast can we do this?
15:23Sure, Jennifer.
15:24I think, firstly, you know, we must acknowledge that the pace at which the industry is growing, this is turning
15:32out to be a global issue, even though data centers today are concentrated in a few regions.
15:38So, 85% of all data center capacity globally today is between the US, Western Europe, and China, or Eastern
15:45China, to be precise.
15:48Therefore, the rest of the world doesn't have the same challenges at the moment.
15:52But even they, in various parts of the world, in Southeast Asia and South Asia, in certain parts of Africa,
15:59you are seeing requests for new data centers coming in that far exceed the ability of the local providers to
16:06meet that challenge.
16:07So, to give you an example, in Texas, Texas alone has connection queues from data centers that amount to 180
16:17gigawatts.
16:18For context, that's two times more than the global capacity of commercial data centers.
16:24So, two times of the global existing today is what requests are going into just one place, which is Texas.
16:31The question is, to what extent are some of them speculative?
16:35To what extent are the companies who want to put their, you know, put their, put themselves ahead of the
16:41queue, perhaps with the hope that they will get that request and they'll be able to scale up?
16:45There's a second and added challenge that AI-specific data centers bring to the table, which are slightly different from
16:54what the conventional data centers that we're used to today.
16:56So, the new AI-specific data centers tend to have workloads that are slightly different, and they're also concentrated in
17:06terms of how much power they can consume.
17:08So, to give you some context, there's the next generation of AI-specific racks.
17:15So, in a data center, you'd find like these, you know, racks which look like, say, large refrigerators.
17:22A single one of them can consume as much electricity as 50 to 60 households.
17:28Okay?
17:29And that's almost like a small neighborhood.
17:31And they can also have significant variability in demand.
17:34It can spike up very quickly, can double, then it can halve and double, and all in less than a
17:39second.
17:39This is where, again, the role of, you know, batteries come into the picture, which Nicholas mentioned.
17:45So, point is that this new dynamics are changing the game, which means that regulators and transmission companies will have
17:55to put their thinking caps on,
17:58start from scratch, you know, have better communication as to how and where data centers can come up,
18:05where it's able to scale up infrastructure, and which parts may be easier for them to provide green electricity to,
18:13and so on.
18:14So, there's a bit of reform needed on the side of regulators and the electricity sector, but also on the
18:21tech side,
18:21to be more transparent about their intentions to scale up data centers in those particular regions.
18:30This next question is really to all three panelists, anyone who wants to answer it.
18:36How do we soothe the concerns of, you know, consumers and locals in places like India?
18:44You know, they're worried that, like, all the village water is going to be used up by the data center.
18:48In the U.S., they're worried about, you know, how does this impact my electricity bill?
18:53How do we bring, how do we soothe their concerns, address those concerns, and make sure that the benefits of
19:05building data centers are not detrimental to local populations?
19:10And whoever wants to take that?
19:13I don't know. I mean, there's an answer, and there is no definitive answer to that, but I would say
19:18this problem is as old as the world.
19:21We all want progress. We all want technology, but never in my own backyard.
19:27And we have been addressing this issue as humanity since a lot of time.
19:31We have built railways. We have built highways. We have built a lot of things that, when they come near
19:38your home, you're not super happy, obviously.
19:40Still, as a society, we need it. So, the answer for me is not really AI specific.
19:44The question is really how are we able, as a society, to make sure that the global benefit brought to
19:51all of us by AI is compensating enough for the people who directly suffer from it.
19:57For whatever reason, they are close to a data center. They are close to a building.
20:01But, to some extent, we know the recipes to that.
20:04We have been doing that as a society for tens and tens of years of balancing individual constraints versus general
20:11need.
20:12And there is a balance, of course, as always, between the greater good of all of us and the impact
20:18on some of us.
20:18And we need a society to take care on this impact, on the one impacted negatively, so that they are
20:24not or they are at least enough compensated.
20:28Yvola, do you want to add to that?
20:29Yeah, I'm happy to.
20:32I just had a conversation just before coming here with the CEO of an industry association for the data center
20:39space called iMason.
20:40I encourage all of you to have a look.
20:42And she really shared that the data center space does come from a legacy of secrecy.
20:49To some extent in the past, building data centers was not something that was really advertised.
20:54It was something typically which was nondescript boxes on a highway.
20:59You didn't really know you were going by a data center.
21:01And clearly she recognizes, and I think very clearly the industry recognizes, that that time has changed.
21:07And it's really about engaging, explaining, and working with the communities at all levels to address some of these challenges.
21:15Because some of these challenges are real and some are more frights that folks might have.
21:21But I think, and this is more from our standpoint as an energy actor who is really very supportive, who
21:27has been working with this space for a very long time.
21:29This partnership that we can bring between the digital space and the energy space, I think can help address many
21:35of these challenges.
21:36It can help address some of the grid challenges that are faced, because there are many synergies between the energy
21:41actors and the data center actors.
21:44It can also address some of the acceptability challenges.
21:47You mentioned power prices.
21:49You know, we're convinced at Engie that, you know, you must bring not only load to an area, but you
21:56also have to bring generation capacity.
21:58We're convinced that the best solution is through renewables and batteries.
22:02They're the cheapest to market, which addresses the time to market challenges that they have.
22:06They are the fastest to market, and so it's a very competitive value proposition.
22:11And this is really how we're helping address some of these challenges, not all of them, but some of them.
22:16Okay, thank you.
22:17Do you want to add to that, Sid?
22:18Sure, actually.
22:19If I could maybe say two things about this.
22:23The first is that, you know, what are the main concerns when local communities think of a data center in
22:31their particular district or village or town?
22:35What are they really thinking of?
22:36They're thinking of things like jobs, like creation of local jobs.
22:39They may be thinking about the environment, you know, access to water and all of that.
22:44And they may be thinking about their impact on electricity prices.
22:48Since I work at the International Energy Agency, I'll focus on the third bit, which is power prices.
22:54So we did an analysis to understand this relationship and what is it that's driving these concerns.
23:02Typically, in the energy sector, there are very high capital costs.
23:07It costs a lot of money to put the infrastructure that's needed on the generation side to transmit the electricity
23:13and so on.
23:15It's very important for any industry when they come in, any consumer when they come in, to provide a certain
23:24sense of stable base load.
23:26Like when you're able to distribute costs, it can lead to lower pressure on prices.
23:32That is, you can actually have lower prices if the consumer is able to provide a certain predictable amount of
23:39electricity consumption across the year.
23:41But while that may be the case for data centers theoretically, what happens in practice is that because the data
23:49centers are coming in very quickly,
23:52and there is a little bit of a question mark on how much electricity ramping, that is, how quickly they'll
23:59increase their consumption over the years,
24:01that leads to a mismatch between the amount of investment that's being put into the infrastructure and the actual consumption.
24:09That mismatch can lead to upward pressure in prices.
24:13So, therefore, firstly, I think there are tools that governments have, that regulators have, that can help ensure there's greater
24:21synchronization between the demands from the data center industry
24:26and the provision of infrastructure for these data centers.
24:31That's one part of the answer.
24:33The other is that when you think about AI, so this is more a conceptual and broader, I think, point.
24:40When you think about AI, what is it that you're really talking about?
24:43You know, when people say, okay, there's a data center here, and these data centers are going to, you know,
24:48cater to some form of AI demand, what is it that you're talking about?
24:52I don't think that everyone currently has a positive view of what AI can do for them.
25:01You know, AI is equally associated with slop, with, you know, fake news, with videos that they may not.
25:07And is the presence of this infrastructure or this, you know, ugly building in my neighborhood just creating that, you
25:15know, fake news?
25:15Is it just doing the bad thing?
25:17I think it's very important, therefore, for AI to be used for things that are socially and economically desirable.
25:26One of them, because again, I work in the energy sector, for me is that energy sector optimization.
25:31Can AI be used to reduce emissions by making systems more efficient?
25:37Can it be used to increase the uptake of solar and wind in the grid by better weather forecasting?
25:46There are several examples of how AI can actually help optimize the energy sector.
25:51I think, A, that needs to be implemented, and B, people need to understand that that infrastructure can also cater
25:58to positive things.
26:01Yeah, sure.
26:02I just wanted to give a couple, just building on what you shared, a couple of very concrete examples of
26:07how we use AI as an energy company.
26:10As you can imagine, forecasting demand, forecasting production is critical in order to plan the dispatch and the use of
26:19energy assets.
26:20And through AI, we've really been able to capture very significant improvements in our ability to forecast.
26:25And if you forecast better, you reduce the hedges and the contingencies that you need to take on, and that
26:32has a direct impact on costs.
26:34And so that's something which is very positive.
26:36Just recently, in the past year or so, we've managed to secure 15% improvements in our accuracy of our
26:42forecast.
26:42That's example number one.
26:44Example number two is on a battery that we just launched recently in Belgium, a 200 megawatt battery.
26:52And its entire operation are done automatically through algorithm.
26:578,000 decisions a day which are done and optimized through just AI and supported by these large models.
27:07Great. And just as a follow-up to that, I'd love to get your view.
27:12I had an interesting interview earlier this week about the evolution of the battery sector.
27:18And, you know, one scenario is that between now and, say, 2030, that we could, with improvements in battery, see
27:30a world in which, you know, all consumers and industry and data centers are all contributing to helping balance the
27:41grid.
27:42And even consumers would be paid when, you know, for helping contribute at the right time.
27:49Do you share that vision?
27:51I certainly share that vision.
27:53I think it's not going to be all or the other, but certainly this notion of distributed flexibility, which is
28:01fundamentally what you're referring to,
28:03is something that we're already seeing, notably in the Netherlands and in Belgium, where there's already high penetration, but we
28:10see that happening much more in the years to come.
28:13The cost of batteries have come down tremendously, actually faster than any other technology, so it's really quite remarkable.
28:20And the opportunity for distributed flexibility, I think, is huge.
28:24And at Engie, we're certainly looking to be an active player in that space.
28:28And as I shared before, batteries and data center can be a key answer to some of the challenges that
28:34you've seen with the grid and having grid access.
28:37Okay, great.
28:38And then I want to come back to you, Philippe, because, you know, Sid touched on this idea of, you
28:46know, AI being used for, you know, making businesses more sustainable.
28:53And I know that that is a subject dear to your heart.
28:56So I wanted to let you talk a little bit about that.
28:59Yeah, I'd love to take a very practical example on how, today, we are already able to drastically reduce the
29:09cost of energy, the end of energy, and the carbon used by everybody, by our customers.
29:16Take something as simple as the room in which we sit.
29:20There is certainly a thermostat, a room controller, that controls the temperature of the room.
29:26By the way, it's a bit cold, but there is certainly a room controller.
29:32Bringing this room controller years ago helped the people get the right comfort with the right amount of energy.
29:41Then what happened is we, or the industry, connected that room controller to the system in the building so that,
29:47during the night, we don't cool it or heat it for nothing.
29:50And we heat it or cool it only when used.
29:53Already, you save 20%, 30% energy.
29:56Then the next question is, let's say, VivaTech starts at 9am.
30:01You want to have the right temperature at 9am.
30:03When do you need to start your air conditioning?
30:05At 8?
30:068.15?
30:078.30?
30:08With AI embedded in something as simple as a few dollar room controller, you are able to learn the thermal
30:15behavior of the room.
30:16Decide exactly when to stop, when to start.
30:19Increase the comfort and save again 20%.
30:22And that applies to every office in the world, every hotel room, every hospital room, every building.
30:28When you know that building industry is more or less 30% of the consumption of the energy of the
30:33world, and things as simple as that can save 20%.
30:37Coming back to the example that you were taking on forecasting, helping to have a greener energy, I come back
30:44on this question of peak demand and how to avoid to use energy at peak demand when it's expensive and
30:51usually quite heavy in carbon.
30:55Take again a building like this one.
30:56I'm sure people who run VivaTech and other events are climate conscious, they put a lot of solar panels in
31:03the roof.
31:04Great.
31:05But maybe not that useful.
31:06Because when the sun is shining, the sun is shining for everybody.
31:11Everybody produces green electricity and maybe we are even in a situation where a negative price and where there is
31:16overproduction.
31:18So in itself, good to have solar panels, but maybe not that beneficial.
31:22Or you add a bit of AI in a building like this one.
31:26You forecast with weather forecast, with machine learning, the production of the solar panels.
31:31You forecast the consumption of this building.
31:34And we don't do that 8,000 per day, but four times per hour, deciding if it's better actually to
31:40use what is being produced.
31:41To store it in a battery to use it tomorrow.
31:44To sell it to the grid because for whatever reason, we are producing and not the rest of France.
31:49And we can sell to the grid to help the grid.
31:51Those solutions exist today.
31:54They save 20 to 30 percent cost on carbon.
31:57The question now is really, how do we scale them?
32:01How do we scale them fast enough?
32:03How do we get them adopted?
32:05And that's why here this discussion that we're having on energy, making people conscious that the responsibility of energy is
32:13to all of us.
32:14That the responsibility of helping with the demand is with all of us.
32:18Should help us deploy our solution, but of course, energy solution.
32:22The point is not what we build, but those technologies exist.
32:25They are simple, easy to use, and they save today already 20 to 30 percent.
32:30Thank you, Philippe.
32:31So, there's one thing that we haven't talked about so far, which I think is really important that we cover.
32:36And that is large language models, as we've discussed, use a ton of energy.
32:44But, you know, sometimes we are using, you know, models that are so heavy for trivial tasks.
32:57Let's address the discipline of, like, knowing when to use large language models, when to use, you know, small models.
33:07And, you know, do you see things moving in a direction where we will end up using smaller models, more
33:17than large language models?
33:19Yeah, very clearly, there is a strong responsibility of suppliers like us to choose the right AI for the right
33:27task.
33:28Many things we have described so far could be done using LLM.
33:33They can also be done with much cheaper AI in terms of energy consumption, most energy efficient AI.
33:39And it's extremely important that not only we don't always pick up the most complex, the most sophisticated LLM, but
33:46we don't also forget that many things can be done with much simpler, much cheaper AI.
33:52I will just tell you a small anecdote.
33:54At some point with some people in Schneider, I was working with them on building an AI solution.
33:59And while exploring the topic, I said, you know, at the end, you don't need AI.
34:03They were disappointed, but they were absolutely wrong.
34:05The point is not to bring technology for the sake of technology.
34:08The point is to achieve the target, reducing emission, reducing carbon.
34:13And a large LLM is not always the answer.
34:15So, yes, 100% agree with you.
34:17It is very important to not fall in the fascination of technology for technology.
34:22Many, many tasks can be done with much more simpler LLM, even small LLMs, even locally, no need to go
34:29to the cloud.
34:29And many tasks can be done without LLM, with more traditional AI that works extremely well.
34:35So, yes, it is important to not do tech for tech.
34:37It's interesting.
34:38This morning, I was talking to an AI scale-up company that's speaking here at VivaTech about their AI solution
34:49for corporates.
34:50And they were saying that, like, it depends on the application.
34:55And some solutions might only require 10% AI.
35:00Others, 90%.
35:02And some, maybe not at all.
35:04So, this idea of just slapping AI on everything is kind of folly because...
35:12I can only agree.
35:14Yeah.
35:14Sid, do you want to add to that?
35:16Sure.
35:17So, you know, these domain-specific models tend to be much smaller in size because they're trained on very specific
35:28data of certain activity or something that you're trying to optimize.
35:32Compared to some of these larger frontier models that are being developed by some of the big tech startups around
35:39the world.
35:40Those tend to be, you know, multi-modal in many ways.
35:44So, they generate not just text but also video, audio, images, and so on.
35:50They could consume, you know, anywhere from 100 to 1,000 times more electricity, you know, in the training, for
35:57example, compared to a domain-specific model.
36:00But having said that, I will also say that most users don't know and probably we cannot expect them to
36:08care about the differences between these models.
36:11Models are coming out all the time.
36:13There are, you know, new applications, new paradigms.
36:19You cannot expect a typical consumer, an average consumer, to actually keep track of everything that's changing.
36:27Ultimately, companies respond to incentives.
36:30And here, the incentives will be the incentives of, you know, like having efficient workloads, meaning lesser requirements for data
36:42centers, meaning quicker access to the grid or more affordable prices.
36:46Those are the incentives that will hopefully drive companies to focus, you know, computations towards smaller and more niche models.
36:59So, let me turn to you, Nicola.
37:01We only have, you know, a few minutes left.
37:04And so, from your perspective as, you know, being in charge of data center acceleration at Engie, what is the
37:13one message that you'd like to leave for this audience?
37:18So, one, it's a message of optimism.
37:22And secondly, it's the recognition that the AI infrastructure challenges and the energy transition challenges are really two sides of
37:32the same coin.
37:33And that many solutions and there are many synergies between the two.
37:36And that we've got a very strong conviction that, you know, it's through the co-development of these industries and
37:41of these addressing these challenges that we can really go a long way.
37:45Okay.
37:46Thanks.
37:46Over to you, Philippe.
37:48I would say that, yeah, I share the optimism.
37:51Definitely.
37:52We have to work on both sides.
37:53We have to work on energy for AI.
37:55How do we power more efficiently, better, less, as little as possible data center, choosing the right model, all of
38:03that.
38:03And also, AI for energy.
38:05How we use AI to optimize the demand, optimize the production, reduce the emissions.
38:10And then I would leave with one recommendation for each of you and see you touch on that.
38:15I do expect you to understand.
38:17And you should expect from yourself to understand.
38:20You need to learn.
38:21You need to be trained.
38:23You need to spend a bit of time.
38:24Otherwise, you're going to just listen to everybody being said everywhere.
38:28This is our duty as citizens to understand a bit, how it does, how it works, what is good, what
38:34are the limitations, all of that.
38:35So my personal word for you is really, really, all of us need to learn and understand this technology.
38:42Sid, you have the last word.
38:45And I would say that, you know, as far as the energy world goes, energy, sustainability, climate change,
38:51there are multiple challenges that are, that are, that have parallelly kind of focused or appeared on our horizon.
39:00So energy security continues to be a challenge.
39:02Energy affordability continues to be a challenge.
39:05Sustainability continues to be a challenge.
39:08It's at this time that AI has emerged as this vector, this new technology that is disrupting the way we
39:15do things or that has, that promises, you know, large scale changes.
39:20We should use this as an opportunity to reform the various things that need reform at the moment.
39:27So this includes, you know, what we already talked about, the reformation of connection queues or infrastructure that's needed.
39:34We find that over the past many years, there has been an underinvestment on grids.
39:39There's been an underinvestment on the infrastructure that enables electricity or clean electricity to be supplied to the consumer.
39:47This should be an opportunity for us to work on those multiple challenges, use AI to the maximum benefit to
39:56make the system more efficient, to make it more resilient, to make it more sustainable.
40:01And therefore, use AI and use this moment in time towards meeting those larger social objectives.
40:08So, do the panelists agree that, you know, AI will help us get greener and that, you know, we just
40:21need to be a little bit patient and that it's not the villain that it's painted to be?
40:29That it won't happen automatically, that we will need to work hard towards actually ensuring the benefits appear.
40:36And it's not to say there are no downsides. There are plenty of downsides. We discussed many of them.
40:40How can we ensure that we overcome them and make this into an opportunity from a current challenge that it
40:46is?
40:46I would really say, if we use AI for energy, I'm with you.
40:49If we use AI for fake news, weird kind of movies and so on, there is no benefit, only bad
40:56side, let's be very, very clear.
40:57And the biggest part of the bill, unfortunately, will not be on using AI to save energy.
41:02The biggest part of the bill will go on things totally useless, up to us to do it or not.
41:09Anything to add to that?
41:11Okay. With that, let's give a nice round of applause to our panelists.
41:15Thank you all.
Comments