The Heat 2024. Electrify Everything: The Power Challenge Behind Data Centres
23 September 2024 | Irena Spazzapan
The surge in AI capex plays into the so called “capex super-cycle”, a combination of technology, energy transition, nearshoring, geopolitics, demographics and public investment. AI requires significant upfront investment in physical, digital, and human capital, with overall capex expected to rise to $200 billion globally by 2025.
For the next couple of years, the physical world that powers data centres is going to be the biggest constraint to faster growth of next gen models.
Today, electricity remains only about 20% of primary energy demand globally. Until recently, the “electrify everything” narrative was all about transport and heating taking electricity towards 80% or more of primary energy demand. With the advent of gen AI, information has become the third and fastest growing variable in electrification.
There are three alternative narratives behind the physical world that powers AI.
I won’t spend much time on the first narrative. This is the belief that AI is the work of god as we push towards AGI, so adding another 1-2 GT of CO2 emissions is nothing compared to the gains we will see later down the line.
The second narrative is the one we hear most often in climate circles, pointing at ballooning carbon budgets from big tech and AI being an unwelcome addition to an already tight grid, making things worst. Lets explore this side of the story.
In Europe and the US, power demand has been flat for 20 years. This year, for the first time we are up 1.5-3%. About 1% of this due to increasing load from data centres. You will have seen many reports predicting this growth continuing into 2030. If this becomes true, then by 2050, internet could be 6% of global power consumption (from about 1.5% today). This would make internet the fastest growing segment in recent history.
AI power demand is split between model training, and AI searches. Lets look at some interesting data: Chat GPT-4 used 40x more power by processing 10x more parameters than Chat GPT-3. Now, model training has taken most power use today, but AI queries will start taking up most growth in electricity by 2030: search queries could surpass 1BN/day this year, 10BN/day in 3 years (comparable to Google searches today), 350BN/day in 20 years (comparable to today’s daily email traffic). Today, an AI search takes 10x more power than a google search and 4x more than an email. See why this is worrying for already congested grids?
Yet, this surging power demand is very different to transport and heating electrification – where cars and heat pumps are not continuously used assets. Data centers target 99.7-99.995% load factor Vs average 38% across the US grid. Renewables have an average 20% load factor, so they are only a small part of a solution that needs other assets too, taking the supply to 24/7 – but today at very high costs, about 300 $/MWh, as opposed to the classic 60-70 $/MWh PPA (where you offtake the generated hours only).
Quoting Jensen Huang, data centres will go where baseload power is cheapest. In the US, people in the fossil fuels industry are glowing at the opportunity of the largest build up of CCGTs in 3 decades, transforming the Permian basin into a giant power generator. In Europe, Scandinavia and Spain where a mix of respectively wind/solar and hydro can get to cheap baseload should be the winners, plus France with nuclear. The UK might be in the game if we scale offshore wind fast, and most importantly if we allow for locational pricing making power in Scotland among the cheapest in the world, as opposed to being forced to take the national price dominated by tight grids around London. If Europe wants to compete with the US (or China, adding new nuclear reactors every year in record timing), cheap power is crucial, and it can only be unlocked by rapid deregulation of the permitting process bringing large-scale renewables and batteries to market quickly. Nuclear is unfortunately not on the list because in the West it takes us 15-20 years to build a new power station, compared to 7 years in China.
Finally, the third narrative is one of efficiencies. Earlier I told you that power demand has been flat in the West since 2000. This wasn’t because of declining demand but because of efficiencies, estimated at 1.5-2%/year ie an impressive 35% over the past 20 years.
AI is going to accelerate efficiencies massively. Think about manufacturing, logistics, mobility, agriculture – AI will enable the so called “physical world” ie the world of atoms to go digital. Think a Tesla car that reboots its software whilst you sleep, or the latest John Deere tractors bringing machine learning to the fields, or Octopus Energy heat pumps optimizing for your personalized energy consumption.
In the world of bits - computers, which has dominated innovation with internet and mobile, you fundamentally have 3 layers: CPUs/GPUs manipulate the bit, storage stores the bits, and networks move bits around. In the physical world, the world of atoms, the equivalent are manufacturing which manipulates atoms, real estate that stores atoms and logistics that move atoms around. This is also where virtually all anthropogenic emissions come from. With digital twins and predictive AI, products get smarter in use, producing higher utilization and less waste across factories, buildings, and transport.
Further, efficiencies will continue also in data centres. Today, Nvidia GPUs are usually used at about half their capacity. We don’t have cooling that is good enough to avoid these chips from melting from overheat. The next generation of GPUs, Blackwell and then Ruben, will further improve density and efficiency. Over time, we might be able to move towards the miniaturization of data centres, making it easier to apply next gen cooling. Or perhaps we will might be able to move away from silicon chips to new forms of semiconductors, way more efficient at consuming energy. We don’t know yet where efficiencies will take us, but what I strongly suspect is that the linear growth in power demand projected from data centres is going to look very different by 2030.
Never bet against human ingenuity!
contact@
systemiqcapital.earth
