The real leverage in AI infrastructure is shifting away from Washington's policy ambitions and toward whoever controls the power and chips. Nvidia's announcement of a $150 billion annual investment in Taiwan represents a direct rebuke of the Trump administration's vision of an American AI hub, signaling that the company sees Taiwan's existing ecosystem and proximity to TSMC as more valuable than policy incentives or subsidies. This isn't a vote of confidence in US manufacturing capability; it's a calculation that Taiwan offers the lowest-friction path to scale. Meanwhile, Snowflake's $6 billion deal with AWS for AI CPU chips and DigitalBridge's $1 billion acquisition of ArcLight energy firm show the same pattern repeating: the money flows to whoever can guarantee reliable power and compute, not to whoever promises the best regulatory environment.
The gap between what companies claim AI does and what it actually delivers is widening just as adoption accelerates. Google's inability to spell its own name in AI-generated search results, alongside reports that AI agents ignore evidence and struggle to learn from mistakes, exposes a fundamental brittleness beneath the productivity narratives. Yet companies like Remote are reporting 50% revenue-per-employee gains from AI adoption, and Cognition reached $492 million in annualized revenue run rate, suggesting the gains are real even if the technology is flawed. The resolution of this tension lies not in the AI getting smarter but in organizational structure shifting to absorb the failures: humans reviewing AI outputs, agents running in sandboxed environments, and platforms like Robinhood restricting AI agents to pre-loaded wallets. The productivity gains are real. The safety is purchased through constraint.
Regulation is becoming a competitive advantage for those already dominant. Illinois passing America's strongest AI safety bill requiring third-party audits of models from OpenAI, Anthropic, and Google creates compliance costs that lock in incumbents while raising barriers for new entrants. The EU's push for tech sovereignty and the US law enforcement warning about "anti-tech extremism" both reflect a shift from innovation-first policy to control-first policy, but neither changes the underlying power distribution. Cognition's $25 billion valuation after eight months, Kirkland & Ellis committing $500 million to build proprietary AI technology, and OpenAI's foundation allocating $250 million to research AI's economic impact all point to the same conclusion: the winners are determined not by regulation but by who can move fastest and lock in users before the rules harden. Five days until Computex will likely confirm this pattern in real time.
Sloane Duvall