Interview, Fireside Chat
Can $500 Billion Win the AI Race? | Anton Leicht
- The "fast follower" open-source strategy is expected to provide U.S. defensive operations a four-month window before models become accessible to non-state actors and rogue nations, though this lag will likely diminish as frontier developments become restricted and less visible.
- Middle powers face a choice between integrating into the AI supply chain to secure economic contribution and minimal dependency or risking a "second best/slash worst equilibrium" where they become useless allies, potentially necessitating narrow policy interventions to align industries with American development.
- A "compute for access" model, where countries build data centers to secure guaranteed frontier model access at parity with U.S. markets, is anticipated to become a standard framework, with effective data center buildouts realistically achievable within 18 to 24 months.
- Building a competitive frontier model is estimated to cost 500 billion over four to five years for a coalition of middle powers, a cost likely to rise by 100 billion every six months due to accelerating chip competition, potentially reaching a trillion dollars.
- Direct deals with commercial labs are predicted to be more viable and stable than government-level deals, as labs have a more acute need to diversify inference compute and may lobby the U.S. government to prevent access restrictions.
- A coalition project aiming to build a frontier model would require a broad alliance including the EU, UK, Canada, Australia, New Zealand, South Korea, Japan, and possibly Taiwan, funded primarily by treasuries rather than private capital.
- Security carve-outs for closed or government access parity would require middle powers to adopt U.S.-handed down standards regarding cybersecurity architectures, firm security, and ties to China.
- Securing necessary hardware involves acquiring approximately three million class-leading chips, a difficult task due to U.S. export controls that might be mitigated by leverage from the semiconductor supply chain involving ASML, Japan, and South Korea.
- The U.S. may not initially restrict chip exports to a coalition project until it succeeds, but could later engage in broader retaliation including tariffs, seizure of intelligence cooperation, and hitting security architectures if the project threatens U.S. dominance.
- Countries without frontier models risk losing sovereign relevance if relative growth diverges due to the compounding advantages of early innovation and recursive self-improvement in AI systems.
- Economic value capture in the AI era will depend on owning "AGI-compatible" industrial assets, such as super-modular automated factories that provide proprietary data, requiring countries to screen Foreign Direct Investment and modernize plants with skilled labor and sensors.
- Labor markets in middle powers must be extremely flexible to navigate the "automation vs. augmentation race," as inflexible markets risk high long-term unemployment and efficiency losses, whereas wage guarantees or subsidies may fail to satisfy young workers.
- Policy responses to AI displacement should avoid direct token taxes on AI use or equity stakes in developers; instead, broad-based consumption taxes or increased corporate income taxes are identified as more viable methods for capturing value and funding social disruptions.
- Formal U.S. nationalization frameworks are unlikely; instead, a "messy" dynamic of executive threats and export controls will likely continue to influence lab behavior.
- A "politically realistic" pause on AI development is expected to fail unless it is a sophisticated international treaty, as a symmetric pause is difficult to sell to the U.S. and would disadvantage American strategic leadership against China.
- Governance strategies favor involving more middle powers to reduce volatility and avoid unilateral policy errors, as a stable world with distributed leverage creates better incentives for decision-making than a U.S.-centric model.
- Germany is projected to succeed in AI if it achieves 3% to 4% economic growth and implements a flexible labor market, leveraging its skilled workforce and industrial base despite the political difficulty of required reforms.
- The strategy of building a frontier model is a contingency plan to be advocated only if the "compute for access" approach fails and the U.S. stops deals, restricts chip outflow, or prevents data center structures.
- Banning recursive self-improvement (RSI) loops within labs is identified as the most operationally viable safety policy, as a "banned on super intelligence" is too underspecified and banning RSI allows for continued visibility while preventing software-only explosions.