The Geopolitics of AI 2040: Evaluating 'Plan A' and the Feasibility of Compute Arms Control

A recent analysis published on lessw-blog provides a condensed overview of the AI Futures Project's "AI 2040: Plan A," a comprehensive proposal advocating for an international AI-race slowdown treaty. For PSEEDR, this text represents a critical pivot in AI safety advocacy from theoretical alignment research to applied geopolitical arms control, raising significant questions about the feasibility of compute-monitoring and verification protocols in a multipolar world.

The Strategic Logic of a Managed Slowdown

The core argument presented in the lessw-blog summary is that the rapid, unchecked progression of AI capabilities drastically increases the probability of catastrophic outcomes. These risks are not limited to sudden AI takeover scenarios but also encompass gradual disempowerment of human institutions and extreme power concentration in the hands of a few state or corporate actors. To mitigate these threats, the AI Futures Project proposes "Plan A," an international arms control agreement designed to throttle the raw scaling of AI capabilities while explicitly permitting and encouraging ongoing research into AI alignment and control. This approach is positioned as a pragmatic alternative to "Plan S," a complete and permanent pause on AI development. The authors argue that a structured slowdown acknowledges the impossibility of halting technological progress indefinitely while providing the necessary breathing room to solve fundamental alignment problems. By treating advanced AI development as a global security threat rather than a purely commercial enterprise, Plan A attempts to align the incentives of competing superpowers toward mutual survival rather than unilateral dominance.

Geopolitical Feasibility and Compute Arms Control

Evaluating the geopolitical feasibility of an international AI-race slowdown treaty requires examining historical precedents in arms control, such as the Non-Proliferation Treaty (NPT) for nuclear weapons or the Chemical Weapons Convention (CWC). However, applying these frameworks to artificial intelligence introduces unprecedented challenges. Unlike fissile material, which requires massive, easily detectable industrial infrastructure to enrich, the primary resource for AI development-computational power-is inherently dual-use and globally distributed. Enforcing a treaty like Plan A would necessitate a highly intrusive, global compute-monitoring regime. This would likely involve hardware-level tracking mechanisms embedded in advanced semiconductors, cryptographic verification of large-scale training runs, and international inspectors granted access to sovereign hyperscale data centers. The geopolitical friction generated by such a regime would be immense. Sovereign nations increasingly view AI superiority as a foundational pillar of future economic and military security. Convincing competing superpowers to submit their most advanced computing infrastructure to international oversight requires a paradigm shift in how states perceive the immediate threat of unaligned AI versus the strategic risk of falling behind human adversaries. Furthermore, the current geopolitical climate, characterized by intense technological decoupling and export controls, suggests that achieving consensus on a binding, verifiable treaty will be exceptionally difficult.

Implications for the Global AI Ecosystem

If a framework resembling Plan A were to gain diplomatic traction, the implications for the global technology sector would be profound and disruptive. The locus of AI governance would shift decisively from domestic regulatory agencies to international oversight bodies. Hardware manufacturers, cloud service providers, and sovereign data centers would effectively operate under regulations akin to defense contractors, subject to stringent export, import, and usage controls. We would likely witness the emergence of a massive "compute compliance" industry, where cryptographic proofs of training compute and algorithmic audits are mandated by international law. This regulatory environment inherently favors massive incumbents and state-backed actors who possess the resources to navigate complex international compliance structures. Consequently, an unintended side effect of Plan A could be the severe stifling of open-source AI development and the marginalization of smaller research laboratories that cannot afford the overhead of treaty-mandated verification protocols. Additionally, capping the raw compute available for training runs would likely drive intense investment into algorithmic efficiency. Researchers would pivot toward achieving greater capabilities with less compute, potentially accelerating certain types of AI advancements in ways that evade the specific thresholds defined by the treaty.

Limitations and Verification Blind Spots

While the lessw-blog summary successfully makes the dense, comprehensive AI Futures Project proposal more accessible to key stakeholders, several critical blind spots remain in the text. The source explicitly mentions that Plan S (the permanent pause) possesses key limitations compared to a structured slowdown, but it fails to detail what those specific limitations are. More importantly, the concrete enforcement and verification mechanisms of the proposed international treaty are entirely omitted from the summary. How does the international community prevent rogue actors or non-state entities from utilizing decentralized compute clusters to bypass treaty limits? What are the specific penalties for treaty violations, and which international body holds the authority to enforce them? Additionally, the core differences in technical assumptions between the project's earlier "AI 2027" scenario and the current "AI 2040" timeline are not explained. A thirteen-year shift in projected timelines implies a massive recalibration of assumptions regarding hardware scaling, algorithmic breakthroughs, or economic investment. Without addressing these enforcement mechanics and timeline shifts, Plan A remains a theoretical policy construct rather than an actionable diplomatic strategy.

The introduction of Plan A signals a necessary maturation in AI safety discourse, moving beyond abstract technical alignment puzzles to confront the harsh realities of international relations, state-level competition, and arms control. However, the viability of an AI-race slowdown treaty hinges entirely on the development of robust, tamper-proof compute verification technologies and an unprecedented level of global cooperation. Until the mechanisms of enforcement are as rigorously defined as the existential risks they aim to prevent, international AI arms control will remain an ambitious but elusive objective, leaving the global community to navigate the rapid advancement of AI capabilities without a coordinated safety net.

Key Takeaways

• Plan A proposes an international AI-race slowdown treaty to mitigate existential risks while allowing alignment research to continue.
• Enforcing such a treaty requires unprecedented global compute-monitoring, presenting massive geopolitical and technical hurdles.
• A compute-capped regulatory environment would heavily favor state-backed actors and massive incumbents, potentially stifling open-source development.
• The source text lacks critical details regarding the enforcement mechanisms of the treaty and the specific limitations of a permanent pause (Plan S).