AI: A high-level perspective on environmental aspects
This is part of a series focusing on the presentations from the Netnod Meeting 2025. Each blogpost covers the key points from the presentation and provides links for readers who want to know more.
The context
It is not obvious how AI will affect society. This talk focused on the question of how the enormous power requirements for building out large AI data centres will be met. I suggest that, among the AI hype, the details of power needs and the environmental impact should not be ignored.
The energy use
Data centres are growing quickly in size and power use: in 2013 we marvelled at Meta’s 30 MW data center in Luleå, while this year China Telecom’s IMIP draws five times that as much power at 150 MW. And the largest planned data centre is in the magnitude of 3 000 MW, and will require an investment of 35 billion USD.
Another aspect of this energy use is the water necessary in producing electricity. According to a recent report, global AI data centers require 4-6 times more water than used in Denmark each year. In France, the amount of water needed to cool their existing nuclear plants has become a national issue, with many of these reactors forced offline in the past two summers thanks to drought conditions.
While data centres use as much power as smaller cities, they can spring up in months rather than the decades it takes for a city to grow. This makes it extremely difficult for electricity companies to plan and build out power grids to avoid grid congestion issues. The end result is massive delays connecting major industrial users to the power grid. In Virginia, USA, the wait can be as much as 7 years.
Bring Your Own Power
As a consequence, the current trend for AI data centre energy is Bring Your Own Power (BYOP) – in the form of natural gas powered mobile generator trucks.
As an example: xAI’s Colossus data centre in Memphis hosts 100 000 Nvidia H100 GPUs in its first phase, which was brought online last year. This number of GPUs requires at least 150 MW of power, and so the company has deployed a large number (some reports claim up to 35) gas powered trailer generators. According to the Southern Environmental Law Center xAI only applied for permission for 15 of these generators and they also claim that these generators are pouring out more nitrogen oxides (up to 2 000 tons) than the Memphis International Airport (1 077 tons), the Draslovka chemical plant (743 tons), and the Valero oil refinery (342 tons) combined. And xAI has plans to increase the scale of this data centre by a factor of ten. They have not gone into detail about how they’ll power it – especially given that the trailer generators they use today now have a two year lead time from their manufacturers.
A similar, but less extreme example took place in Ireland, where Google’s plan for a data centre expansion was rejected due to power concerns.
When it comes to the huge energy requirements for AI data centres, there is no clear plan today on how to establish the necessary power production.
What Power Requirements Do AI Data Centres Have?
It’s pretty clear that the city-scale AI data centres being planned have 4 key criteria:
- Gigawatt scale
- 24/7/365 energy supply
- High density power generation
- Low CO2-emissions
Figure 1 shows the 9 traditional ways that electricity is generated. Which of these approaches can satisfy an AI data centre?
Figure 1: Power options for AI Data Centres. Source: Slide 38 here.
Considering the requirements of gigawatt scale, 24/7/365 energy supply, with low CO2-emissions and compact installation, only nuclear and geothermal remains viable options.
However, geothermal falls into two types: Conventional and Enhanced. Conventional Geothermal accounts for less than 1% of electricity generated today, because it only works where geological conditions are favourable – countries like Iceland, Italy, Turkey, Indonesia, Kenya and the USA. Some of these facilities operate on gigawatt scale.
Enhanced geothermal power uses fracking techniques to enable geothermal generation in far more locations. However, they only operate on Megawatt scales today, they may need a lot of fresh water, and generate significant pollutants as slurry from the bore holes. Nevertheless, a study by the Idaho National Laboratory (hopefully an unbiased sources) suggests that 100 GW of power could be created from EGS within 50 years in the USA alone.
Nuclear Power
Regardless of what you may think about nuclear power, it really does seem to be the only option for these mega scale AI data centers. But it’s clear that there are many issues – both real and perceived – that haunt the nuclear industry.
Let’s put aside the obvious issues of public fears about safety and nuclear waste. Traditional nuclear power has a massive problem because, here in the West, we have literally forgotten how to build large scale nuclear power plants. Let’s looks at challenges in 3 Western nations.
In the USA the Vogtle 3 and 4 nuclear plants finally came online in Georgia. But they are at least 8 years late and their cost has risen from $14B to $37B – that’s a 2.6X cost overrun.
In the UK they are building two new plants, including Hinckley Point C in Somerset. The original date of 2025 has slipped to 2029, but they are already talking about a further slip to 2032. Costs have risen from $12B to a staggering $56B – a 4X cost overrun.
What about France? In the 1970s and 80s France deployed over 50 reactors and was a poster child for nuclear power. At one point over 80% of French electricity was generated by nuclear…but today that has fallen to just over 70%. The latest French reactor at Flamanville has been plagued by delays and design issues. It was supposed to be operational in 2012, but is still not running. Costs have spiralled from $3.4B to $19.6B – almost 6X.
So…it is impossible to build nuclear plants today? Actually countries like Russia, China and South Korea are doing just fine.
Probably the best model is that of Russia. Before 2007 the Russian nuclear industry was a mess – with over 350 individual companies involved in the whole supply chain. In that year Vladimir Putin himself ordered the total integration of these companies into a single entity – Rosatom. Over the past 18 years Rosatom has put together a complete end-to-end offering for anyone who wants to own a nuclear power station. From site selection, to plant design, staff training, operations support, supply of the reactor, supply of the fuel, disposal of waste and a dozen other key capabilities – including finance, all of them are handled by Rosatom. If you are a non-aligned nation and want your reactor delivered on time and budget then Rosatom’s offering is extremely attractive. In fact Russia provides over 40% of the world’s nuclear fuel – including to reactors in the USA and UK. These deals are exempt from sanctions, by the way.
While China does not specifically have an equivalent of Rosatom, they are planning to build 150 new reactors by 2040, in addition to the 57 they have operational today. As the most CO2 polluting nation on Earth this is a vital step to a cleaner environment.
Small is Beautiful?
If large scale reactors are challenging to build on time and on budget, maybe we should scale down the problem. Small Modular Reactors (SMRs) are designed to be built in a factory and then with final assembly on site. This reduces costs and project uncertainties, and helps to stay on budget. However, there are only a few SMRs in operation today in, of course, Russia and China! In the West, rather than focus on a couple of promising designs, our Venture Capital culture means that over 60 companies are competing with different designs, with none of them operational today.
Hyperscalers, such as Amazon, Google, Microsoft, Oracle and Meta, are all planning ahead in terms of power supply, and are all, in the long term, leaning towards nuclear (either larger plants or SMRs). In at least two cases they also have plans for Enhanced Geothermal.
Is Nuclear Power Dangerous?
This is a really complex question and, like many controversial topics, it is bogged down with misinformation on both sides. Is nuclear power dangerous? Of course it is…just like any other means of generating power.
But in the 70 years or so of commercial nuclear power there has only been one incident that resulted in a significant loss of life – and that’s Chornobyl (or Chernobyl in the Russian spelling). Let’s put aside the fact that the RBMK reactor at Chornobyl is a terrible design that lacked several basic safety features we demand in every Western reactor design. Let’s put aside the ridiculous and irresponsible way that the “safety test” was carried out that resulted in a reactor explosion. And let’s put aside the classic Soviet approach to secrecy that meant people were not treated for radiation exposure in a timely manner. A number of people died as a result of a nuclear reactor explosion. How many? The figure is extremely difficult to find out. The UN seems to accept a value of around 60, but other estimates run into the thousands. Which do we believe? I suppose we could say that, if there were really going to be thousands of deaths from radiation we would have seen them by now, given that the accident happened almost 40 years ago.
Another way to look at it is this: following on from Chornobyl and Fukushima (where it’s possible that one plant operator may have died as a result of radiation exposure), Germany decided to shut down its 11 nuclear plants that provided 26 GW of clean, safe energy generation. They had to replace that with something…and they chose to keep a number of coal fired power plants in operation past the date they were scheduled for closure. A report by the National Bureau of Economic Research concluded that as a result of this decision an additional 1 100 people die every year in Germany from the pollution caused by coal fired power stations. Even Greta Thunberg has stated that she believes it was the wrong decision to close down those nuclear plants.
What now?
I see three steps the hyperscalers should take right now:
- Grab any spare capacity on existing nuclear facilities before someone else does
- Restart any mothballed nuclear facilities – amazingly even as recently as 2022 around 25% of nuclear plants in the USA were at economic threat of closure.
- Build new, large-scale nuclear plants – assuming you remember how
In short, nuclear power has a lot of challenges, and for this reason it may not be a general purpose solution for green energy generation. But hyperscalers are not a general purpose audience – they need Gigawatt scale power and they need it now. Nuclear is not just the best option, it’s really the only option given the power needs of AI data centres.
You can watch Geoff’s presentation from the Netnod Meeting 2025 here and see the slides here
