The Geopolitics of AGI | Helen Toner (Director of CSET & past OpenAI board member)

US-China AI Competition and Export Controls

• CSET research in 2019 identified semiconductor manufacturing equipment (SME) as a critical choke point, leading to export controls that successfully slow China's domestic supply chain buildup without severely antagonizing them.
• Export controls on chips themselves generated a more antagonistic response from China than SME controls, though China's reaction was less escalatory than some US policymakers feared.
• There is significant strategic confusion within the US regarding the objective of chip controls, oscillating between preventing military use, limiting general AI development, and maintaining US market share.
• China recently imposed a self-imposed restriction on purchasing next-generation NVIDIA chips, a move interpreted by experts as potential signaling intended to pressure the US into allowing exports or a genuine assessment of domestic capacity.
• CSET estimates the current AI capability gap between the US and China has narrowed to approximately six to twelve months, down from the 2022 estimate of two to three years.
• Despite headlines about Chinese chip breakthroughs (e.g., Huawei's recent chips), CSET analysis suggests these capabilities are often reliant on indirect manufacturing via TSMC rather than fully independent domestic production.

Geopolitics of Compute and the Gulf States

• Deals for data centers in the UAE and Saudi Arabia involving hundreds of thousands of next-generation NVIDIA chips are controversial due to the autocratic nature of these regimes and the lack of transparency regarding ownership and usage rights.
• Companies justify these deals with the claim that "if the US doesn't sell, China will," but experts argue this is disconnected from reality as China currently lacks the manufacturing capacity to match such massive compute deployments.
• Critics argue that empowering Gulf autocracies with world-class supercomputing power risks concentrating strategic advantage in regimes whose primary interest is regime stability rather than alignment with US or democratic interests.
• There is a possibility that the UAE, driven by significant capital reserves and a focus on scaling compute, could become a primary location for training frontier AI models in the long term.
• OpenAI's framing of Gulf data center deals as "wins for democracy" is viewed by experts as cynical, given the strict censorship, banned political parties, and labor abuses characteristic of the UAE.

AI Governance, Safety, and Policy Frameworks

• Helen Toner advocates replacing the term "AI alignment" with "AI steerability" to better capture the practical challenge of controlling AI systems during both development and operation.
• Toner proposes an "adaptation buffer" framework, suggesting society should focus on building resilience against misuse (e.g., bio-weapons) in the period before such capabilities become widely accessible, rather than relying on unworkable non-proliferation treaties.
• There is a tension between the view that power concentration is necessary to manage existential risks and the view that it creates dangerous long-term vulnerabilities and reduces human agency.
• CSET recommends increased transparency and disclosure from AI companies as a primary tool to allow broader public and governmental oversight, rather than seeking a single "full solution" to safety risks.
• The debate over "rent vs. sell" for compute suggests that renting access allows for immediate termination of service if misused, whereas selling chips creates permanent dependency and harder-to-reverse proliferation.

Internal Dynamics of OpenAI and the Trump Administration

• Toner defends the 2023 OpenAI board decision to temporarily remove Sam Altman, stating that while it appeared naive externally, the board considered multiple factors that remain confidential due to the fast-moving and complex nature of the crisis.
• OpenAI's recent restructuring attempts to retain control within its nonprofit entity face scrutiny from attorneys general regarding potential conflicts of interest, as board members effectively negotiate with themselves over equity and control.
• The Trump administration's policies, including restrictive immigration for high-skilled workers and uncertainty around energy generation, contradict strategies designed to maximize US AI competitiveness against China.
• There is a divergence of factions within the current US administration, with some officials favoring commercial engagement with China while others maintain traditional hawkish stances, resulting in inconsistent policy signals.

Future Trajectories and Workforce Needs

• The perception that AI progress is slowing is attributed to unrealistic expectations set by the singular jump of ChatGPT, whereas long-term observers see a steady, consistent upward trajectory of sophistication.
• Distinct "worldviews" in the AI field (optimistic vs. pessimistic) remain resilient because they possess self-consistent frameworks for interpreting contrary evidence, making it difficult to empirically distinguish between "intelligence explosion" scenarios and plateaus.
• CSET identifies a critical talent bottleneck in the US policy sphere for individuals with dual expertise in AI technology (calculus, linear algebra, basic programming) and other domains like economics, history, or anthropology.
• Military adoption of AI is expected to be slow and piecemeal due to institutional barriers in procurement and testing, with a current focus on decision-support systems rather than autonomous lethal weapons.
• CSET is currently hiring research fellows focused on frontier AI and expects to continue expanding data and communications roles through 2026 to maintain its independent, evidence-driven analysis.