Posts & Publications

Konstantin F. Pilz, James Sanders, Robi Rahman, Lennart Heim

April 2025

Frontier AI development relies on powerful AI supercomputers, yet analysis of these systems is limited. We create a dataset of 500 AI supercomputers from 2019 to 2025 and analyze key trends in performance, power needs, hardware cost, ownership, and global distribution. We find that the computational performance of AI supercomputers has doubled every nine months, while hardware acquisition cost and power needs both doubled every year. The leading system in March 2025, xAI's Colossus, used 200,000 AI chips, had a hardware cost of $7B, and required 300 MW of power, as much as 250,000 households. As AI supercomputers evolved from tools for science to industrial machines, companies rapidly expanded their share of total AI supercomputer performance, while the share of governments and academia diminished. Globally, the United States accounts for about 75% of total performance in our dataset, with China in second place at 15%. If the observed trends continue, the leading AI supercomputer in 2030 will achieve 2e22 16-bit FLOP/s, use two million AI chips, have a hardware cost of $200 billion, and require 9 GW of power. Our analysis provides visibility into the AI supercomputer landscape, allowing policymakers to assess key AI trends like resource needs, ownership, and national competitiveness. (Blog post | Paper)

Konstantin Pilz, Yusuf Mahmood, Lennart Heim

January 2025

Larger training runs and widespread deployment of future artificial intelligence (AI) systems may demand a rapid scale-up of computational resources (compute) that require unprecedented amounts of power. We find that globally, AI data centers could need ten gigawatts (GW) of additional power capacity in 2025 alone, which is more than the total power capacity of the state of Utah. If exponential growth in chip supply continues, AI data centers will need 68 GW in total by 2027—almost a doubling of global data center power requirements from 2022 and close to California’s 2022 total power capacity of 86 GW. Given recent training compute growth, data centers hosting large training runs pose a particular challenge. Training could demand up to 1 GW in a single location by 2028 and 8 GW—equivalent to eight nuclear reactors—by 2030, assuming that current training compute scaling trends persist. The United States currently leads the world in data centers and AI compute, but unprecedented demand leaves the industry struggling to find the power capacity needed for rapidly building new data centers. Failure to address current bottlenecks may compel U.S. companies to relocate AI infrastructure abroad, potentially compromising the U.S. competitive advantage in compute and AI and increasing the risk of intellectual property theft. More research is needed to assess bottlenecks for U.S. data center build-out and identify solutions, which may include simplifying permitting for power generation, transmission infrastructure, and data center construction. (Link)

Zachary Arnold, Daniel S. Schiff, Kaylyn Jackson Schiff, Brian Love, Jennifer Melot, Neha Singh, Lindsay Jenkins, Ashley Lin, Konstantin Pilz, Ogadinma Enweareazu, Tyler Girard

October 2024

AI-related laws, standards, and norms are emerging rapidly. However, a lack of shared descriptive concepts and monitoring infrastructure undermines efforts to track, understand, and improve AI governance. We introduce AGORA (the AI Governance and Regulatory Archive), a rigorously compiled and enriched dataset of AI-focused laws and policies encompassing diverse jurisdictions, institutions, and contexts related to AI. AGORA is oriented around an original taxonomy describing risks, potential harms, governance strategies, incentives for compliance, and application domains addressed in AI regulatory documents. As of its launch in July 2024, AGORA included data on several hundred instruments, with new entries being added continuously. (Find database at: agora.eto.tech | Link to paper)

Claire Dennis, Stephen Clare, Rebecca Hawkins, Morgan Simpson, Eva Behrens, Gillian Diebold, Zaheed Kara, Ruofei Wang, Robert Trager, Matthijs Maas, Noam Kolt, Markus Anderljung, Konstantin Pilz, Anka Reuel, Malcolm Murray, Lennart Heim, Marta Ziosi

November 2024

As artificial intelligence (AI) advances, states increasingly recognise the need for international governance to address shared benefits and challenges. However, international cooperation is complex and costly, and not all AI issues require cooperation at the international level. This paper presents a novel framework to identify and prioritise AI governance issues warranting internationalisation. (...) We find strong benefits of internationalisation in compute-provider oversight, content provenance, model evaluations, incident monitoring, and risk management protocols. In contrast, the benefits of internationalisation are lower or mixed in data privacy, data provenance, chip distribution, and bias mitigation. (Link)

Girish Sastry, Lennart Heim, Haydn Belfield, Markus Anderljung, Miles Brundage, Julian Hazell, Cullen O'Keefe, Gillian K. Hadfield, Richard Ngo, Konstantin Pilz, George Gor, Emma Bluemke, Sarah Shoker, Janet Egan, Robert F. Trager, Shahar Avin, Adrian Weller, Yoshua Bengio, Diane Coyle

February 2024

Recent AI progress has largely been driven by increases in the amount of computing power used to train new models. Governing compute could be an effective way to achieve AI policy goals, but could also introduce new societal risks. Our paper gives a broad overview of the properties that make compute a promising governance tool and discusses the benefits and risks of various compute governance proposals. (Link to arXiv)

Konstantin Pilz, Lennart Heim, and Nicholas Brown

November 2023

Training advanced AI models requires large investments in computational resources, or compute. Yet, as hardware innovation reduces the price of compute and algorithmic advances make its use more efficient, the cost of training an AI model to a given performance falls over time — a concept we describe as increasing compute efficiency.

We find that while an access effect increases the number of actors who can train models to a given performance over time, a performance effect simultaneously increases the performance available to each actor. This potentially enables large compute investors to pioneer new capabilities, maintaining a performance advantage even as capabilities diffuse.

Since large compute investors tend to develop new capabilities first, it will be particularly important that they share information about their AI models, evaluate them for emerging risks, and, more generally, make responsible development and release decisions.

Further, as compute efficiency increases, governments will need to prepare for a world where dangerous AI capabilities are widely available — for instance, by developing defenses against harmful AI models or by actively intervening in the diffusion of particularly dangerous capabilities.

Preprint on arXiv

Konstantin Pilz & Lennart Heim

July 2023

Data centers are industrial facilities that efficiently provide compute at scale and thus constitute the engine rooms of today’s digital economy. As large-scale AI training and inference become increasingly computationally expensive, they are dominantly executed from this designated infrastructure. Key features of data centers include large-scale compute clusters that require extensive cooling and consume large amounts of power, the need for fast connectivity both within the data center and to the internet, and an emphasis on security and reliability. The global industry is valued at approximately $250B and is expected to double over the next seven years. There are likely about 500 large (> 10 MW) data centers globally, with the US, Europe, and China constituting the most important markets. The report further covers important actors, business models, main inputs, and typical locations of data centers. (Link to arXiv)