Transcription
Hello and welcome to People and Profit. I'm Charles Belgrave. Every month, every week, every day, the artificial intelligence wave covers more and more ground, radically changing how businesses operate and promising to boost productivity and profit margins.
The focus used to be on training AI models. Now, increasingly, it's on inference, the actual execution of the tasks we ask these systems to perform. And that's what's driving demand for energy even further: energy to fuel the data centers springing up worldwide to execute these tasks.
The International Energy Agency forecasts that global energy generation to supply data centers will grow from 460 terawatt hour to over 1,000 terawatt hour in 2030 and 1,300 terawatt hour in 2035. A lot of that generation is driven by so-called hyperscalers, the large tech groups like Amazon or Google that have the financial ability to build these sprawling data centers, and that's triggering pushback from communities worried about environmental impact, water use, and rising electricity prices. But does it need to be so huge in scale?
Well, our guest today is David Gurle. He's a French tech entrepreneur who previously held senior roles building up infrastructure at Microsoft Teams and Skype, among others, and who is the man, or one of the men, behind Antimatter. The startup aims to service this inference boom with energy solutions: mini data centers that fit inside a container, and appropriate software as well, for only a fraction of the cost and time. At least, that's the promise.
David Gurle, thanks for being with us.
Thanks for having me today, Charles.
So, what technological shift makes these smaller data centers viable now, when the industry has spent years moving towards these massive hyperscaler facilities? Is it that inference is less energy and computing intensive than AI model training?
Actually, it's, um, it's, it's about the type of model that you are using. You know, the world of frontier AI models is splitting in kind of two dimensions, you know? First dimension is the ones from Anthropic, you know, from xAI and from OpenAI. And we call them proprietary frontier AI models. They need massive amount of GPUs to be trained, which I talked about that, and then massive amount of GPUs to run for inference use cases. Um, but the problem with those models is that they are not sovereign. And their cost per token keeps increasing due to a number of energy shortage and chip shortage issues.
So, the world has kind of created a new alternative: more frugal, actually more sovereign. It's the open-source models. So, as the open-source models became more prevalent and, more, I would say, better in quality, um, enterprises started to adopt these tools. And these tools are far, far more frugal in the way they consume GPU cycles, CPU cycles, memory, and especially energy. So, the reduction of these models consumption of this compute and energy capacity is enabling a new world of inference to emerge.
Um, you mentioned frugality. I want to focus on one particular example that I've seen related to Antimatter and Entropycloud in particular. Um, for example, using the heat generated by, uh, the, uh, the, the, the containers in the data center, um, to heat up greenhouses in farms in rural France. Is that actually like a genuinely scalable, uh, model and infrastructure, or is it that more of a, a niche demonstration project?
>> Actually, it's really scalable, and believe it or not, we are not the initiator of this project. Um, you know, actually it's the arc culture, uh, you know, cooperatives, uh, have come to us and they said, "Look, you know, we have these huge greenhouses." Uh, you know, and, uh, we have invested over the last 10, 15 years in this renewable energy farms. We have batteries. And then we have these mechanical heat pumps. They cost a fortune to buy and a fortune to maintain, and, um, and that is to take the heat and inject it into greenhouses so that the tomatoes can grow the way we like them.
And they want to swap those mechanical heat pumps into digital heat pumps. And by doing so, not only they're going to be able, uh, to heat, uh, in a constant way all of the greenhouses, but also generate additional revenue. And, uh, and so we are talking about, you know, dozens of dozens of sites just across southwest of France. And in France alone, we are talking about roughly over 200 sites which can take those micro data centers, the polyclouds, and then produce, you know, use the produced energy from, uh, solar or wind, um, and then inject the generated heat into the greenhouses so AI makes tomatoes.
I see.
Um, another, uh, issue here, obviously, uh, at the forefront of people's minds, is water consumption. Uh, quantitatively, how much lower is the water footprint, um, for these, uh, smaller, uh, centers?
>> Zero.
Which is zero.
>> Yes, zero. We don't use any water, uh, because the, um, the heat generated by these units, first of all, they are cooled by air. And so, we have air conditioning units in those micro data centers, and they are, um, very sufficient in order to cool down to the optimal temperatures, uh, those servers, which are is very important.
Um, and then because we are using less, uh, power hungry servers and less power hungry chips, most importantly, um, therefore, we are able to, uh, generate much less heat than the big servers and the big chips do generate. So, in other words, if you think about it, if I may make a parallel, you know, we are not running a Ferrari engine. You know, we are running maybe a Toyota hybrid engine. And, uh, and therefore, it's far more efficient for what the Toyota can do.
Mhm.
Um, you're up against giants. We mentioned it earlier, the US hyperscalers. Uh, you're aiming to get 1,000 poly clouds, uh, by 2030 from 100, uh, this year. Can a company like yours realistically, uh, remain independent in the market dominated by the hyperscalers, or are you bound to eventually join them in some way or another?
>> That we can, and we want to be, uh, independent. I think we have a very unique opportunity to create, um, a complementary solution. I don't want to say supplementary. I, I would like to position ourselves as kind of the last mile of inference, uh, which is sovereign, which is frugal, which is responsible, and, and which is, uh, at the edge, you know, where people are. And, uh, and give them the opportunity to, uh, to own this computing infrastructure and serve their communities and innovate at their own pace, you know, with their, um, you know, their mind at peace with respect to data and what I call inference sovereignty.
Um, any AI story is automatically met by a fair amount of, uh, skepticism and fear from the wider public about what it means for their jobs, but what it means for the environment. And data centers really crystallize those fears. Around the world, communities are coming together to stop these new developments. Brian Quinn tells us the story, the story of one such community in Cape Town, South Africa.
In Cape Town's Gugulethu Township, impoverished residents of this informal settlement struggle to weather proof their improvised homes. They also struggle with access to water and electricity.
What was his comments?
If they install water taps here, the criminals steal them. So, we have to travel far to fetch water. Even the portable toilets get stolen.
Now, they have a new worry. A massive data center project has sprung up nearby. With its projected power draw of 160 megawatts, community organizers fear added strain on both the power grid and the city's already notoriously precarious water supply.
In 2017 and 2018, we had drought. It was days here where people didn't have access to water. We had rolling blackouts happening. What is the rush to put in place technology that, that we don't know what the knock-on effect it will have on the lives of ordinary South Africans?
As tech firms rush to build up computing power amid the AI frenzy, data centers are facing increasing backlash around the world, including in the US, where they've been blamed for soaring consumer electricity prices, and Spain, where water shortages are common. The Cape Town community groups, along with a UK-based nonprofit, have filed a formal objection with city planners. They say the project has been short on transparency from the start.
This is a huge project that is nothing about water, nothing substantial about emissions. It's incredibly limited on electricity. There's nothing about the diesel generators. There's nothing about air pollution. There's nothing about noise.
The US-based company behind the project already operates a site in Johannesburg, which it says it's fully powered by renewable energy. It has 30 days to respond to the complaint, after which the Cape Town city government has 6 months to make a ruling. The South African government, meanwhile, is pushing ahead with digital infrastructure projects, planning to boost investment via tax incentives and policy reforms.
So, David Gurle, there's such a rush to build these systems that social and environmental concerns often seem to be secondary. Is it acceptable to actually slow this down, even if, even if it means slowing down AI innovation?
I think there are two dimensions to it. First of all, this is not the first time I hear this. I think in 2024, four of the big data center projects were either paused or canceled. In 2025, that's 25. And probably in 2026, the number is going to be much bigger.
Is it 20%?
>> Uh, numbers.
Numbers of big data centers.
>> Okay. So, it's not percentage. It's the total, uh, number of number 4 to 25, correct?
>> Yeah.
Yeah, exactly. And I'm sure it's going to be double in 2026. There are multiple reasons for that.
Um, first of all, the population is waking up to the consequences of what it means for their community. And, and I must admit that the companies who are going and building these giant data centers are not doing a great job in working with the communities and explaining, you know, the benefits or the issues and how they're going to cope with that. After all, it is part of their, you know, sites. And, and I think there is an awakening that's needed, you know, from those companies who are building those data centers.
Um, you know, should we slow them down? Um, I think there is no need to slow them down because they are naturally slow down. Um, you know, the energy shortage is real. Um, obviously a number of projects will be cancelled or paused due to population, um, you know, rejection of those, uh, projects, but many more are actually slow down due to energy shortages. And, uh, and this massive energy demand is not available today everywhere.
David Gurle, unfortunately, that's all time we the time we have for, uh, thank you so much for sharing your insights, uh, with us on France 24. You are the CEO and co-founder of Antimatter, uh, and, uh, it was a great honor to have you on the program.
Pleasure was mine. Thank you very much.
Well, that's all we have time for this week. As usual, make sure you subscribe to People and Profit on the podcast platform of your choice. You can also catch all of our previous episodes on the France 24 app. Thanks for watching and stay tuned to France 24.