Lennart Heim is a researcher focused on the governance of computing power as a lever for AI policy. He is currently at the RAND Corporation, having previously worked at the Centre for the Governance of AI (GovAI) at Oxford. His work centers on compute governance, training-compute measurement, and the political economy of chips, data centers, and power.
Affiliations
Heim is an associate information scientist at the RAND Corporation and a professor of policy analysis at the Pardee RAND Graduate School, where he leads compute research in the Technology and Security Policy Center within RAND Global and Emerging Risks (Source: rand.org). He co-authors compute and energy analyses cited across the US government. He previously worked at the Centre for the Governance of AI (GovAI), where he produced conceptual work on compute as a governable input. He has collaborated with researchers at Epoch AI, the Center for Security and Emerging Technology (CSET), the Center for Strategic and International Studies (CSIS), the Oxford Martin School, and Open Philanthropy-adjacent policy groups.
His academic background is in engineering rather than policy: a 2020 master's thesis on evaluating and deploying resource-constrained machine learning on embedded devices, completed jointly at ETH Zurich and RWTH Aachen University, and a 2016 RWTH Aachen bachelor's thesis on network virtualization for software-defined-radio wireless experiments (Source: heim.xyz).
Publications
Heim is a co-author, with Pilz and Mahmood, of AI's Power Requirements Under Exponential Growth (2025), published as RAND RRA3572-1 (see RAND — AI's Power Requirements Under Exponential Growth (2025)). The report is a policy-facing forecast of AI data-center power demand through 2030.
His earlier GovAI work argued that compute is a governable input, on the grounds that its measurability, excludability, and supply-chain concentration make it more amenable to governance than data or algorithms. He has also contributed to training-compute measurement methodology used in the Epoch Notable AI Models dataset and cited in US export-control policy.
His wider published record falls into four clusters.
Compute as a governance instrument. The anchor paper is Computing Power and the Governance of Artificial Intelligence (February 2024), a GovAI white paper on which Heim is second author behind Girish Sastry, with a nineteen-author list including Markus Anderljung, Miles Brundage, Gillian Hadfield, Yoshua Bengio, and Diane Coyle. Related work develops specific mechanisms: Governing Through the Cloud: The Intermediary Role of Compute Providers in AI Regulation (Oxford Martin School, March 2024, with Fist, Egan, Huang, Zekany, Trager, Osborne, and Zilberman); Oversight for Frontier AI through a Know-Your-Customer Scheme for Compute Providers (with Janet Egan, October 2023); and Training Compute Thresholds: Features and Functions in AI Regulation (with Leonie Koessler, August 2024, arXiv:2405.10799), which examines what a compute threshold can and cannot do as a regulatory trigger. The Lawfare commentary "To Govern AI, We Must Govern Compute" (March 2024, with Anderljung and Belfield) is the short-form statement of the position (Source: heim.xyz). See Compute Thresholds.
Export controls and the chip supply chain. Understanding the AI Diffusion Framework: Can Export Controls Create a U.S.-Led Global Artificial Intelligence Ecosystem? (RAND Perspective PEA3776-1, January 2025) is his sole-authored treatment of the AI diffusion rule, mapping the AI compute supply chain against its export-control points and arguing that the framework strengthens protections against chip diversion and smuggling, while noting that theft of model weights or algorithmic insights could undermine chip controls (Source: rand.org). Adjacent work includes Hardware-Enabled Governance Mechanisms: Developing Technical Solutions to Exempt Items Otherwise Classified Under Export Control Classification Numbers 3A090 and 4A090 (RAND Working Paper WRA3056-1, January 2024, with Kulp, Puri, Gonzales, Vermeer, Smith, and Winkelman); Accessing Controlled AI Chips via Infrastructure-as-a-Service (IaaS): Implications for Export Controls (with Egan, December 2023), a government information-request submission on the cloud-rental gap in chip controls; the Foreign Policy piece "Chinese Firms Are Evading Chip Controls" (June 2023, with Fist and Schneider); and The United Arab Emirates' AI Ambitions (CSIS report, January 2025, with Adamson, Allen, and Winter-Levy) (Source: heim.xyz).
Compute measurement and trends. With the Epoch AI group he co-authored Compute Trends Across Three Eras of Machine Learning (IJCNN 2022), Estimating Training Compute of Deep Learning Models (January 2022), Trends in Machine Learning Hardware (November 2023, with Hobbhahn and Aydos, published with an accompanying dataset), and Will we run out of data? An analysis of the limits of scaling datasets in Machine Learning (October 2022). Compute at Scale: A Broad Investigation into the Data Center Industry (with Konstantin Pilz, November 2023, arXiv:2311.02651) extends the measurement work from chips to facilities, and Increased Compute Efficiency and the Diffusion of AI Capabilities (with Pilz and Brown, published in the Proceedings of the AAAI Conference on Artificial Intelligence 39(26)) treats efficiency gains as a proliferation channel for dangerous capabilities. The Compute Divide in Machine Learning: A Threat to Academic Contribution and Scrutiny? (January 2024, with Besiroglu, Bergerson, Michael, Luo, and Thompson) applies the same data to the academic-industry gap (Source: heim.xyz).
Technical AI governance more broadly. He is a co-author on Visibility into AI Agents (FAccT 2024), IDs for AI Systems (July 2024), Open Problems in Technical AI Governance (July 2024), Responsible Reporting for Frontier AI Development (April 2024), Societal Adaptation to Advanced AI (May 2024), International Governance of Civilian AI: A Jurisdictional Certification Approach (Oxford Martin School, September 2023), and Towards best practices in AGI safety and governance: A survey of expert opinion (GovAI, May 2023). He also contributed to the OECD's A blueprint for building national compute capacity for artificial intelligence (OECD Digital Economy Papers 350, February 2023) (Source: heim.xyz).
Regulatory submissions
Heim has filed comments in several US and UK proceedings: a response to the NTIA AI Accountability Policy Request for Comment (June 2023), a response to the OSTP request for information on national AI priorities (July 2023), a submission to the NAIRR Task Force RFI (June 2022), two responses to the UK's Future of Compute Review (August 2022 and March 2023), a January 2023 comment on the Bureau of Industry and Security's October 7 advanced-computing and semiconductor-manufacturing-equipment rule, and April 2024 comments on the Commerce rulemaking under the malicious-cyber-activity national emergency (Source: heim.xyz). See Export Controls (AI).
Positions and statements
Heim argues that compute is the most tractable AI governance lever: unlike algorithms, which are infinitely copyable, or data, which is hard to track, compute is physical, concentrated, and visible. His more recent work argues that chip export controls are a partial solution and that power infrastructure, including grid capacity, interconnects, and generation, is becoming at least as important a chokepoint.
He advocates for verification infrastructure as the technical substrate for any future international AI agreements, including on-chip mechanisms, registration of large training runs, and disclosure regimes. He frames compute governance as simultaneously enabling, through industrial policy and safety assurance, and restrictive, through export controls and adversary limits, rather than as a choice between the two.
On enforcement specifically, he has framed the aim of compute controls as moving "the unit of governance from AI chips, which are hard to govern and can be smuggled, to computing power itself," on the reasoning that chips located outside a controlled jurisdiction can still be used from within it. He has also characterized export control as "a blunt tool" and identified the most likely failure mode of the diffusion framework as the assumption that a rule adequate today will remain adequate as chip performance improves and more hardware crosses the restricted threshold (Source: chinatalk.media).
Influence
RAND's power-requirements estimates are used as reference numbers for US federal energy-AI planning. Heim's GovAI-era compute-governance frame appears in BIS export-control reasoning and in the US AI Safety Institute's evaluation-thresholds logic. He has been a co-author or acknowledged contributor on multiple CSET and CSIS pieces on US-China compute dynamics. A 2026 Journal of Cyber Policy article on the geopolitics of compute characterizes his position on the AI diffusion rule as a claim that the action would cement the advantages of the United States and designated tier 1 countries (Source: tandfonline.com).
Relationships
- supports: Compute Governance — Heim's work is load-bearing for the concept page.
- related: Epoch AI — research collaborator; Heim's training-compute methodology overlaps with Epoch's Notable AI Models dataset.
- related: Export Controls (AI) — Heim's conceptual framework is upstream of much export-control analysis.
- related: AI Environmental Impact — the RAND report drives the US-share-of-electricity numbers on that page.
- depends-on: RAND — AI's Power Requirements Under Exponential Growth (2025) — co-authored primary publication.
- related: Compute Thresholds — the subject of the 2024 paper with Koessler.
- instance-of: compute-governance scholarship tradition (GovAI / RAND / CSET).
- related: America's AI Action Plan — his compute-governance framing is upstream of the plan's treatment of chips and power.