We need energy for AI, and AI for energy, former Google CEO Eric Schmidt argues

By Eric Schmidt*; (this opinion piece is written from a US perspective).

In 1903, Mark Twain wrote that “It takes a thousand men to invent a telegraph, or a steam engine, or a phonograph, or a photograph, or a telephone or any other important thing.” This observation still mostly holds true. The invention of artificial intelligence required decades of work by thousands of scientists, engineers, and industry leaders. It will require many more men and women to develop the technology in the years ahead.

As the march of AI accelerates, a new requirement has become apparent: the next breakthroughs will consume colossal quantities of energy. AI guzzles electricity: a single ChatGPT query requires ten times as much as a conventional web search. As AI usage increases, its energy requirements will rise, and if demand outstrips supply, the technology’s development will be strangled.

The data centres that underpin AI development at scale – powering GPT-4, Gemini, and other frontier models – need around-the-clock access to power. They already account for roughly 3% of annual US electricity consumption, and this share is expected to more than double in the next five to ten years. More broadly, AI’s electricity usage is projected to increase from four terawatt-hours in 2023 to 93 TWh in 2030 – more than Washington State used in 2022. And that’s a conservative estimate; AI could consume this much power as early as 2025.

Though the dates may vary, the direction is clear: demand for energy will skyrocket. Securing sufficient access to electricity thus has become a top priority for AI companies. But while they are doing what they can, they will not succeed without government help. Building a sustainable supply of power to drive the AI revolution is in America’s interest, and it will also benefit other countries by delivering immense improvements in health care, education, science, national security, and other critical sectors.

To allow AI’s development to taper off because of insufficient energy supplies would be an act of national self-harm. With future needs already apparent, the US government must get ahead of the challenge now, while it can. By the time data centres’ energy demands double, our supply of energy should have at least doubled, too.

Greenlighting the revolution

A study by researchers at Princeton University finds that for the United States to decarbonise its economy by 2050, it may need to triple its electricity-transmission system’s capacity. Much of this can be done domestically. We can build data centres in the US, secure supply chains for energy technologies, and invest in modernising our energy infrastructure.

To that end, the government can unleash private-sector development by simplifying and accelerating siting and permitting. Complicated, decades-long approval processes are a drag on the transition to clean energy, with damaging consequences for AI development as well. Building a new transmission line in the US – from planning and permitting to land acquisition and construction – currently takes an average of ten years. That must change.

We must also take another look at nuclear energy, either by restarting existing plants already connected to the grid, or by building the next generation of smaller, safer, and more efficient reactors. Here, too, reforming regulations, which were written for 1970s technology, is vital to accelerating innovation and deployment. When Georgia’s Vogtle 3 plant came online in 2023, it was the first new reactor design to be approved in the nearly 50-year history of the US Nuclear Regulatory Commission.

That pace is not sustainable in an energy-hungry AI economy. The US must tap every safe, clean, and reliable energy source that it can. A more agile regulatory regime would reflect the reality that not all reactor designs are alike, and that not all review processes should be identical. For example, experts have proposed a more timely review of nuclear projects that reuse brownfield sites – allowing for a smooth transition from coal to nuclear projects.

While the benefits of nuclear power are well known, US energy innovation should also focus on fusion energy, the process that powers the sun. Fusion generators use the immense energy released when light atomic nuclei combine to form heavier ones, making it a potentially abundant, reliable, and clean power source. In 2022, scientists at Lawrence Livermore National Laboratory in California achieved the milestone of fusion ignition, producing more energy from a fusion experiment than was required to drive it.

If the US can develop fusion power at scale, AI’s electricity demands can easily be met. Doing so will require substantial investment and new partnerships between startups and national laboratories to advance concept designs and plan the path to commercialisation. But while new and existing technologies can start to bridge the gap, the US must also seek energy abroad.

AI allies

Looking overseas poses a different set of challenges. While some regions – such as the Middle East – are endowed with vast stores of reliable, low-cost energy, relying on these resources involves geostrategic and national-security trade-offs. It is no good building a data centre abroad to benefit from low-cost energy, only to find that its contents have been compromised by hostile foreign-intelligence services; that it is being used by third-party vendors as a means of avoiding export controls; that its servers are being used to exert leverage over the US; or that it is being used in ways that violate privacy.

AI technology is so valuable, and so critical to US national security, that potential partner countries can only be those that share common interests and values with the US, and which have formidable defenses to safeguard their AI systems. These risks can be managed, but doing so will require deliberate and careful diplomacy and technical engagement with our allies in the Gulf.

Of course, deepening our energy partnerships with European countries poses fewer national-security or reputational concerns; but their energy is often too expensive. In 2023, for example, power in European countries cost twice as much as in the US. In fact, energy is so expensive by international standards that virtually all European AI models are being trained abroad, often in the US.

Fortunately, other US partners offer possible solutions. Japan maintains over 20 gigawatts of idled nuclear capacity, more than enough for even the largest projected computer clusters. And closer to home, Canada offers the prospect of abundant hydropower, which is both renewable and reliable. Both options offer short-term, clean solutions from close American allies.

Advanced AI is both the objective and part of the solution. AI systems see things that humans do not, as DeepMind demonstrated when it successfully reduced energy use for data-centre cooling by up to 40%. We must take advantage of these capabilities. AIs should be deployed to identify and develop new ways to improve data centres’ efficiency and realise fusion energy’s potential. Scientists at the National Fusion Facility in San Diego have already used reinforcement learning to prevent instability and disruption to the complex fusion process.

The path to AI has been marked by unexpected achievements and paradigm shifts. The task now is to harness the power needed to propel this new era of innovation.

*Eric Schmidt, a former chief executive and chair of Google/Alphabet, is chair of the National Security Commission on Artificial Intelligence. Author photo by Guillame Paumier CC-BY-3.0.

Copyright: Project Syndicate, 2024.
www.project-syndicate.org