Three Theories of Victory is a comparative framework attributed to Nita Farahany, introduced in Class 25 of her introductory course (November 30, 2025). It reframes the "AI race" question from "who is winning?" to "what race is each jurisdiction trying to win?" Farahany argues that the United States, the European Union, and China are pursuing distinct theories of how to prevail in AI, such that the conventional single-axis "who's ahead" question presumes a shared race that does not exist.
Origin
Farahany develops the framework from two press conferences held in the same week of July 2025, which she presents as illustrating different conceptions of what AI competition is about. On July 23, 2025, at the "Winning the AI Race" summit, Trump described an American approach centered on deregulation, energy abundance, and keeping advanced chips out of China. On July 26, 2025, at the World AI Conference in Shanghai, Premier Li Qiang proposed a World AI Cooperation Organization headquartered in Shanghai and a Global AI Governance Action Plan framed around "equal rights for all countries" and capacity building for the developing world. Farahany's reading is that the same week and the same technology produced different visions of what the AI competition is about.
The three theories
European Union: be the floor everyone must meet
In Farahany's account, the EU has concluded that it probably cannot win the technology race, lacking both the scale of the US market and the centralized coordination of China. Its theory of victory is to become the global regulatory standard-setter: even if European companies do not build the best AI, European rules would shape how AI is deployed globally, an extension of the Brussels Effect to AI standards. This approach tolerates slower innovation, regulatory burden, and market friction.
The framework cites as supporting evidence the EU AI Act, GDPR's global influence on privacy practice, and the compute and capital gap reflected in Mistral's $640M raise against OpenAI's $8.3B raise at a $300B valuation. The identified vulnerability is a Brussels Effect operating in reverse, in which companies geo-fence and decline EU service rather than comply, with technology firms already delaying EU launches.
United States: be the frontier everyone depends on
The US theory bets on continued technological leadership: American companies build the frontier models, American chips power the training runs, and American capital funds the research. The theory of victory is to stay at the technological frontier, export models, and restrict competitors' access to hardware, treating the frontier as a moat. This approach tolerates distributed harms (with Harper and Sewell cited as examples), reactive individual remedy, slow legal response, dynamic pricing and discrimination, and concentrated benefits to those with frontier access.
Cited evidence includes the frontier-model position of GPT, Claude, and Gemini; the October 7, 2022 chip controls; the Foreign Direct Product Rule, under which roughly all advanced semiconductors destined for China require US permission; and the $500B Stargate buildout. The identified vulnerability is that controls can be circumvented (DeepSeek and the Huawei Mate 60 Pro are cited as cases), that algorithmic efficiency may matter more than raw compute, and that leadership disappears if the frontier becomes commoditized. The US-specific governance mechanisms are treated in more detail in Export Controls (AI), Three-Lane Standards-Based AI Governance, and Procurement-Driven AI Governance.
China: be unavoidable across multiple dimensions
China's theory, in Farahany's account, is to make itself unavoidable across several dimensions rather than compete on a single race, building leverage that matters regardless of who reaches AGI first. The framework groups the moves into four dimensions, drawing on the redirect described in embodied AI versus AGI:
- Rare earth control — roughly 90% of global processing and roughly 70% of production. October 2025 Announcement No. 61 added rare earth processing technology to the export control list with a foreign-direct-product rule mirroring US chip controls.
- Embodied AI dominance — China installed roughly half of all global industrial robots in 2024 (295K against 50K in the US). The 14th Five-Year Plan targets 70% AI-plus-manufacturing penetration by 2027 and 100% by 2035.
- Open-source flood — Alibaba's Qwen generated more than 100K derivatives on Hugging Face, and DeepSeek captured roughly 24% market share on OpenRouter in early 2025. Farahany frames this as an inverse Brussels Effect: standards spread through technological adoption and influence follows ubiquity.
- Global South strategy — positioning Chinese AI as the path of least resistance for developing countries, so that governance influence follows technological adoption.
This approach tolerates stringent domestic content control: the 2023 Generative AI Interim Measures require outputs to "embody core socialist values," the CAC requires registration for every public-facing model, and venture capital is comparatively low ($4.7B in Q2 2025, described as the lowest in a decade and attributed partly to regulatory uncertainty). The identified vulnerabilities include skepticism from Singer (Carnegie) that AI-plus will achieve 90% economy embedding by 2030; effective Chinese chip capacity roughly 170 to 180 times smaller than that of the US and allies when adjusted for yields (TSMC at 60 to 80% on 3nm against SMIC at 5 to 20% on 7nm); and a reliance on third-country arrangements that could be tightened.
The reframe
Farahany's synthesizing point is that asking "who is winning the AI race?" presumes everyone is running the same race, when the more useful question is which race a given actor wants to be running. She maps the three theories to three possible futures: if AI's future is AGI as disembodied superintelligence, the US-led frontier-model race is the one that matters; if the future is AI embedded in physical systems transforming manufacturing, logistics, and infrastructure, China's embodied AI investments may be more consequential; and if the future is determined by regulatory standards that shape global norms, the EU's approach deserves more credit than it typically receives. Farahany suggests all three may be partially right and that the AI future probably includes all of these dimensions, so that the governance choices each jurisdiction makes, and how those choices interact, will shape both who "wins" and what kind of AI future results.
Relationships
- introduced-by: Inside My AI Law & Policy Class 25: China's AI Governance — Are Different Players Running Different Races? (Farahany, November 2025)
- depends-on: Brussels Effect (EU mechanism); Export Controls (AI) (US mechanism); Embodied AI vs AGI (China's Race Redirect) (China's race redirect)
- related: US-China AI Competition: Different Races, Different Metrics, Three-Lane Standards-Based AI Governance (US sub-framework), Procurement-Driven AI Governance (US sub-framework)
- instance-of: AI Governance (umbrella)