The AI and Data Power Shift

Emails. Texts. Social media. Google searches. This post. All of them require energy. How much? This much...So far.

AI is transforming the economy in more ways than one. So far, in Australia, AI is still considered something of magic fairy dust: scatter it on existing products and services, and magic happens, not grasping just how transformative AI will be. And that’s problematic.

There’s much more that can be—and has been—written about the future of AI, its applications in domains from warfare to finance to healthcare, and the ethics of its development and use.

But what we want to address here is its energy implications.

---

AI, like crypto-currency, online games, internet search, 5G networks, software-as-a-service providers, cloud providers, the internet of things (IoT) and an increasingly data-dependent range of industries, rely on data-centre infrastructure.

In 2021, the energy use by data centres globally in 2022 was estimated to be between 240TWh and 350TWh, or one to 1.3 per cent of global electricity production. The actual use was 470TWh, or two per cent of global electricity production.

The International Energy Agency’s Electricity 2024 has forecast that figure could double to over 1000TWh by 2026. As of May 2024 data centres, by 2030, could consume twice that which do now—up to nine per cent of total US electricity generation; globally, the increase may be 160 per cent, or between three and four per cent of total global power.

Specific tech developments can shift consumption markedly: a single query via ChatGPT takes ten times the energy as a Google search—2.9Wh vs 0.3Wh. AI-related tasks are expected to account for 19 per cent power load by 2028. Yet so far—notwithstanding the development of yet hungrier tech and noting that tech improvements will also contribute to better energy efficiency—growth in data centre power demands is not driven solely by AI. Currently, the global power consumption by data centres is roughly half that to power household devices world-wide, but the increasing ubiquity of data, digital devices in households will themselves help drive data centre use.

By way of comparison, the total electricity production in Australia in 2022 was 270TWh, roughly one per cent of global production. Five per cent of that was consumed by data centres in the country; by 2030, that could rise to between eight and 16 per cent.

Nor is such growth linear. With the caveat that the US economy is much more IT-focussed than the Australian economy, from John Politano’s work below¹, we can see that the growth is exponential. Ten years ago, a 40MW data centre would have been considered large. Now hyperscalers of 200MW are common—Australia has a small number—and Meta has flagged a future 1GW data centre. To put that in context, 50MW would power around 20 shopping centres, 180 warehouses, 220 office buildings and 15,000 homes.

For Australia, the exam question that needs to be somewhat urgently answered is how should the nation meet its energy needs arising from such demand from accelerating data, digital and communications development and uptake. The urgency being compounded by large cyber-physical assets under development, such as nuclear-powered submarines, and gearing up for a within-decadal advent of artifical general intelligence (AGI) and quantum computing.

Renewables are part of the answer, including to ensure resilience and to help mitigate the worsening effects of climate change. At present (2023 data), roughly 35 per cent of Australia’s electricity production is through renewables, including solar (16 per cent), wind (12 per cent) and hydro (six per cent). The current goal to achieve a reduction in Australia’s emissions by 43 per cent by 2030 and zero in 2050 is what guides Australian government policy in its approach to national energy needs.

Renewables come with their own challenges, not often acknowledged, though wind turbines have had their share of controversies. Hydro, of course, transforms local ecosystems. Solar panels have a life of around 30 years; global solar waste amounts by the 2050s of 5.5-6 million tonnes will almost match the mass contained in new installations (6.7 million tonnes). Waste or broken solar panels and wind turbines can leach heavy metals into the soil and ground water.

We also tend to think of the electricity systems much as we do with water: with storage, pipes and taps allowing the right amount of water to be provided on demand. But electricity is a different beast: it does not flow per se; it is a force.

Supply of renewables is obviously variable. Systems for introducing inertia may be used to help provide resilience, and steady power system frequency, during disturbances. Market and measures such as fast frequency response can help stabilise the systems after disruption. Battery storage systems can act as a buffer and energy source integrator for similar purposes but have yet to achieve the needed storage capacity needed to ensure the provision of baseload. Infrastructure and machines—including sensitive data storage, servers, and chips—are built on the assumption of sustained, reliable baseload. Variability disturbs the system, and renewables are inherently variable, making managing the electricity grid more of a challenge.

It follows, and is now inescapable that a cheap, stable and reliable baseload drives a technologically accelerated, data-driven economy; renewables can't be the only way to answer the exam question. And so we arrive at (the tech, manufacturing, defence, space, transport and logistics etc.) industry's interest and investment in nuclear power—whether conventional plants, small modular reactors and/or the prospect of fusion power.

Decisions about placement of data centres and high-performance computation is being driven increasingly by the availability of cheap, stable and reliable power. That’s why Microsoft has committed to restarting the old Three-Mile Island plant; once operational, it should generate 800MW. Microsoft, and its founder Bill Gates, have also invested in fusion start-ups and modular reactors. The technology associated with latter is not new per se—the US Navy have been operating them for almost 70 years.²

As noted in a GeoMissive in September 2024, nuclear power in Australia needs to be on the table and openly discussed, not shut down. If Australia is to be technologically competitive, it needs to build capability in the general-purpose technology of our era: data, digital and computation (including GenAI). Given the exponential growth in demand, it is best to prepare for having nuclear in the mix as many of our trading partners and strategic allies already have—and that comes with a significant lead time. Preparing now is necessary; skills in nuclear physics and engineering, plant and infrastructure design, and research into both fission and fusion will not go amiss and will build capability in other domains and for other applications—including submarines.³

Australia is in the early stages of increasing power demands of data centres, as well as growth in the number of required data centres across the country as recent headlining M&A activity demonstrates. There are also opportunities to harvest heat generation and recycle water from data centres (proximity of course a factor); R&D of improved battery technology will be ongoing. Even so, data centre operators are expressing concern regarding the availability of cheap, reliable baseload power.

Beyond the actual generation of power and distribution from its different sources, together with the means of load balancing and resilience, broader infrastructure needs to be built and maintained, for which net zero transition is having some role. An important point on that is made by Ed Conway in his book Material World where he argues that much more power will be needed to mine and process the minerals, and create the technology, infrastructure and capability to realise the necessary electrification needed for zero emissions.

If the exam question is answered (soonest) and the power needs for a technologically competitive nation are addressed, it will help drive and secure Australia’s economic prosperity and future.

We’re running behind with Missives this week, so in case you missed them, here are a couple of recent pieces by one of our co-founders, Lesley Seebeck:

• The new cybersecurity bills will do nothing to create trust, in the Australian Financial Review; and,

• Australian defence policy under a Trump world order, for Strategic Analysis Australia.

1. Australian data is harder to source, but the trends we see of investment into these areas in the United States are indicative, if magnified. Find more of

Joseph Politano

’s work on his Substack.

2. The time frame for the development of the first nuclear-powered submarine, the USN Nautilus, is instructive:

• 1942—first demonstration of a nuclear chain reaction;

• December 1947—nuclear plant design commissioning;

• March 1950—conceptual design of the submarine commences;

• July 1951—Congress authorises submarine construction;

• 14 June 1952—keel laid;

• 1953—nuclear plant prototype tested;

• 21 January 1954—Nautilus christened;

• 30 September 1954—submarine commissioned;

• and 17 January 1955—launch and shakedown voyage of the USN Nautilus.

3. It is worth noting that while nuclear development in the West has stalled, China is building another 22 reactors, on top of its existing 57, 33 of which were built over the 10 years to 2023, driving down the costs of nuclear power.