AI Race: Is US Dominance Dangerous?

In a candid conversation, Ben Thompson, a prominent tech analyst and author of "Stratechery," shared his nuanced perspective on the global AI race and its potential implications, particularly for the United States. He argued that a scenario where the US "wins" the AI race outright, especially from a military or national security standpoint, could be "very problematic for the US to win."

The Perils of AI Supremacy

Thompson posited a fantastical scenario where controlling AI grants a nation unparalleled military superiority. In such a state, he questioned the game theory optimal response of a rival like China, suggesting that the destruction of critical infrastructure, such as TSMC (Taiwan Semiconductor Manufacturing Company), could be a logical, albeit extreme, counter-move. He elaborated, "If we get to a place where we have a meaningful superiority, particularly from like in terms of a military national security perspective, I think that's very dangerous for the world."

The full discussion can be found on Invest with the Best's YouTube channel.

This perspective stems from a fundamental disconnect Thompson observes with much of the prevailing rhetoric in Silicon Valley. He believes that achieving significant AI superiority carries substantial global risks, rather than being an unambiguous positive outcome.

The Current AI Equilibrium and Its Challenges

Looking at the current state of AI development, Thompson likened it to the geopolitical complexities surrounding Taiwan. He suggested that the status quo, where companies like OpenAI and Anthropic lead the frontier while Google and Meta chase, and China remains approximately 6-9 months behind by distilling models, feels like a "pretty good equilibrium that I think is generally favorable to the US." However, he questioned the long-term sustainability of this balance, especially as AI begins to improve itself.

The conversation also touched upon the economic realities of AI, particularly the costs associated with running inference. Thompson dismissed the notion of open-source models being truly "free," highlighting that while R&D costs are recouped differently, the inference costs are very real. He noted the significant difference in costs for users who simply query models for basic information versus those tackling complex problems.

Geopolitical Dependencies and Manufacturing Realities

A significant portion of the discussion revolved around the underappreciated dependency on China for manufacturing precursors and components. Thompson argued that fixing this dependency is "going to be dramatically like dumb" from a purely economic standpoint if competitors continue to source from China at a lower cost. This leads to a situation where such shifts only happen when there is "literally no choice." He used Apple's diversification of iPhone manufacturing to India as an example, noting that Apple is not truly moving out of China but rather diversifying, a process that is astronomically expensive and akin to paying an insurance policy one might not need.

Thompson expressed skepticism about the feasibility of a scenario where the US could completely detach from China's supply chain, to the point where geopolitical events like a conflict over Taiwan would have no impact. He believes there's a "bit of facing reality" missing in current conversations about de-risking and supply chain resilience.

The Role of Competition and Openness

Addressing the competitive aspect of the AI race, Thompson pushed back against the narrative of simply trying to "be like China." Instead, he advocated for a different approach: "America succeeds by being on the leading edge and by leading into that. You said probably the US being purely dominant and AI is not the right end state for the world. What is your ideal equilibrium for how this goes worldwide?" He emphasized more openness, innovation, less top-down control, and fewer restrictions on speech as the path for American success.

Capitalizing on AI's Potential

Looking ahead, Thompson acknowledged the immense economic opportunity presented by AI, even if its models do not improve significantly from their current capabilities. He views himself as a "reluctant accelerationist," believing that progress is inevitable and the focus should be on navigating it rather than resisting it. The sheer scale of potential economic activity, even in "verifiable domains" like coding and math, is massive.

However, he also raised concerns about the timing mismatch between generating revenue to fuel investment and the current capital curve. With companies burning through debt markets and issuing equity, the question of "what's after that? Where's the money come after that?" looms large. He drew parallels to the railroad era, where the long duration mismatch between building and profitability led to financial strain, and suggested that the current AI boom might face similar capital challenges if revenue streams don't catch up to the investment needed for compute, electricity, and talent.

The conversation also touched upon the critical role of semiconductors, the dependence on companies like TSMC, and the potential for a "bull whip" effect in the industry. Thompson concluded by emphasizing that even if economic bubbles burst, the underlying AI technology would persist, underscoring its long-term impact on humanity.