Modern, high-end GPUs have shattered performance barriers, achieving parallel processing capabilities measured in teraflops (trillions of floating point operations per second; FLOPS) and even petaflops (quadrillions of FLOPS).

For example, the NVIDIA GeForce RTX 3090 GPU performs at 35.6 teraflops, or in simpler terms, 35.6 trillion floating-point calculations every second.

In stark contrast, even Intel’s advanced 18-core Extreme Edition CPU, while capable of handling hyperscale data centers tasks such as web hosting and database management, falls short for complex AI workloads with its performance of around one teraflop.

This vast performance gap between CPUs and GPUs boils down to a numbers game of cores and transistors. High-performing CPUs comprise up to 24 cores with transistors numbering in the low billions (e.g., the Intel i7-9700K CPU contains three billion transistors).

In contrast, GPU chips house thousands of smaller, more specialized cores and hundreds of billions of transistors. For instance, the cutting-edge NVIDIA Blackwell B100 accelerator GPU boasts a staggering 208 billion transistors, making it the powerhouse of choice for AI workloads.

These significant AI hardware advancements have triggered a seismic shift in data center power requirements. Server rack power demands have increased from a modest five to 10kW to modern designs demanding 50-100kW or more, with state-of-the-art facilities pushing the envelope to an astonishing 200kW per rack.

Given the average American household uses about 30kWh per day, each modern AI data center cabinet consumes the equivalent energy of three to six typical American homes per day.

This unprecedented surge in energy consumption raises critical questions about power demands, delivery mechanisms, and power generation. For example, we may ask whether existing utilities can manage current AI data center energy demands. We could also consider alternate energy sources such as bioenergy, geothermal, and even nuclear.

As we stand at the precipice of the exascale era, with its promise of quintillions of calculations per second, it is imperative that we explore and debate potential energy solutions to fuel this computational demand.

AI data centers: Scale of energy demand

AI data centers create large language models (LLMs) through two critical machine learning stages: training and inference. The energy demands of these processes are staggering - training a single large model can consume over 50MWh of energy. In contrast, traditional data centers spanning 5,000 to 20,000 sq ft and housing 500 to 2,000 servers typically consume a mere one to five MWh of energy to operate.

This exponential growth in AI capabilities comes at a significant energy cost, compelling the industry to explore and implement a wide array of solutions. This includes not only optimizing existing methods but also pioneering innovative approaches to energy generation and management.

Renewable energy

Renewable energy sources such as solar, wind, and hydroelectric power present promising sustainable solutions for AI data centers, but they come with challenges such as energy inconsistencies, high capital expenditures, and location limitations.

Solar technology, which harnesses photovoltaic (PV) panels to convert direct sunlight into electricity, has seen significant efficiency improvements. However, powering a large AI data center would require substantial land area and careful consideration of deployment space, costs, energy storage, and proximity to high solar irradiance areas (e.g., deserts).

Offshore wind farms offer increased energy consistency, with modern turbines generating up to 15MW of power. While they present an attractive option, the inhospitable nature of offshore locations drives up costs and poses maintenance challenges.

Hydroelectric power offers a more reliable and consistent source of energy, but environmental concerns associated with dam construction limit suitable locations. As a result, AI data centers are increasingly considering run-of-river hydroelectric systems, which have a smaller environmental footprint but lower power output.

To fully harness the potential of renewables, the industry must address challenges such as energy storage and grid integration. Emerging technologies in these areas, coupled with hybrid renewable systems, could revolutionize power generation for AI data centers.

Alternative energy sources

Beyond traditional renewables, alternative energy sources like geothermal and bioenergy present intriguing possibilities for powering AI data centers, each with unique advantages and challenges.

Geothermal energy, tapping into the Earth’s core heat, offers a reliable and constant power source. However, high drilling costs and location constraints may limit widespread adoption.

Bioenergy, which converts biomass into electricity through combustion and anaerobic digestion, offers the dual benefits of waste reduction and decreased fossil fuel reliance. However, challenges include meticulous process management and ensuring a consistent biomass supply. The carbon neutrality of bioenergy remains a subject of ongoing scientific debate.

These alternative energy sources could play a crucial role in diversifying the power mix for AI data centers, potentially offering more localized and sustainable energy solutions. However, their adoption will depend on technological advancements, economic viability, and successful integration with existing data center infrastructures.

Nuclear Energy: Controversies and public perception

Nuclear power, with its high energy density and reliability, presents a compelling option for AI data centers demanding consistent, substantial power. Its relatively small land footprint adds to its appeal in an industry where space efficiency is of utmost importance.

However, nuclear energy faces significant hurdles: high initial costs, lengthy construction times, and persistent public safety concerns, particularly regarding radioactive waste management. The shadows cast by incidents like Chernobyl and Fukushima continue to shape public perception, creating resistance to nuclear adoption.

Yet, technological advancements, particularly in Small Modular Reactors (SMRs), are rekindling interest in nuclear energy. These innovative reactors, producing around 300 MW(e) per unit - about one-third of a full-sized nuclear power plant’s output – offer enhanced safety features and scalability.

Custom SMR designs could potentially meet the varied power demands of AI data centers (ranging from 300-1,000 MW), providing the low-carbon, high-density energy required for large-scale operations. This adaptability makes SMRs an increasingly attractive option for the AI industry's growing energy needs.

Conclusion

As we approach 2028, the anticipated launch of OpenAI and Microsoft’s ‘Stargate’ - poised to be the world’s largest supercomputer - underscores the critical role of energy innovation in AI’s future. Many predict a nuclear solution for this groundbreaking facility.

The relentless advancement of AI technologies will drive an ever-increasing demand for energy. To meet this challenge, operators will likely depend on a diverse mix of energy sources, encompassing solar, wind, hydroelectric, geothermal, bioenergy, and nuclear. The future of AI services is inextricably linked with energy innovation. Sustainable progress in this field will hinge not only on adequate energy supplies but also on advancements in cooling technologies and high-performance networking infrastructure.

Nuclear energy, with its high-density and reliable power supply, emerges as a compelling solution for future large-scale AI data centers. Despite challenges such as cost, public perception, and construction timelines, the benefits of nuclear power may outweigh these concerns as the industry grapples with its escalating energy needs.

As we stand on the brink of the exascale era, the AI industry faces a pivotal challenge: balancing unprecedented computational power with sustainable energy practices. The choices made today in powering AI data centers will shape not only the future of artificial intelligence but also our global energy landscape.