Regulatory Risk is Coming For AI | David Woo on AI Data Center CapEx and Iran War

Episode Summary

Executive Summary: The episode centered on David Wu’s bearish near-term view of AI equities and the AI capex trade. He argued that hyperscaler spending slowed in real terms in Q1, that inflation in data-center inputs is flattering nominal earnings, and that AI model access and monetization are increasingly constrained by regulation and commoditization. He also tied geopolitics—especially Iran, Trump, and oil—to his broader macro stance.

Main Topics: AI capex slowdown and real-vs-nominal spending (Priority: 5/5): Wu argued that the five hyperscalers’ Q1 capex fell QoQ and that higher component prices masked an even sharper decline in real spending, making earnings growth look stronger than underlying activity. Inflation in AI supply chain inflates earnings optics (Priority: 5/5): He said rising prices for memory and other data-center inputs create an accounting and margin illusion: hyperscalers pay more, suppliers book high-margin revenue, and aggregate nominal earnings rise even without more output. Token maxing and near-term AI revenue distortion (Priority: 4/5): Wu claimed Q1 AI usage was artificially boosted by employees and firms maximizing token budgets, implying Q2 growth could slow as budgets normalize and the initial surge fades. AI commoditization and frontier-model competition (Priority: 5/5): He argued that large language models will commoditize because many firms train on similar public data, and that products like Claude Code face rapid competitive pressure from Microsoft, Google, Chinese tools, and others. Regulatory and national-security risk to AI diffusion (Priority: 5/5): Wu’s main long-term bear case was that AI is becoming too powerful and too risky, prompting government restriction. He highlighted cyberattack concerns and the possibility that frontier models will be tightly controlled. Oil, Iran, and geopolitical macro positioning (Priority: 4/5): He linked his bullish oil stance to conflict and uncertainty around Iran, arguing that Trump needs a deal while Iran sees leverage, which keeps oil risk elevated. AI productivity gains and labor-market winners/losers (Priority: 4/5): Despite his bearishness on AI stocks, Wu was constructive on AI’s real productivity impact, saying it can empower small teams, threaten routine white-collar jobs, and heavily disrupt software engineering.

Key Arguments: Q1 hyperscaler capex was down quarter-over-quarter, and in real terms the decline was likely larger because component prices rose. Earnings growth in AI suppliers is partly an accounting artifact: higher prices for the same inputs transfer income from buyers to sellers and lift nominal profits. Q1 AI usage and revenue were boosted by 'token maxing,' so Q2 should show slower growth once budgets normalize. The AI market is moving from a winner-take-all narrative to a commoditized one, especially in LLMs and inference-related products. The biggest bearish risk is not that AI is weak, but that it is too powerful and will trigger regulatory backlash and tighter access controls. AI will likely replace many routine knowledge workers, but it is less capable at original thinking, complex judgment, or real-world autonomy. Small teams can use AI to compete with much larger organizations, increasing entrepreneurship and productivity even as some jobs are displaced. On geopolitics, Wu believes Iran is unlikely to give Trump an easy exit and that continued tensions support higher oil prices.

Data Points: Hyperscaler capex QoQ: Down in Q1 for the first time in three years - Wu said combined capex of the five hyperscalers fell quarter-over-quarter. Hyperscaler capex share of operating income: 135% - He said capex for the five hyperscalers now exceeds operating income on an aggregate basis. Token budget exhaustion: 4 months - He cited Uber reportedly running through its annual token budget in four months. Anthropic user access to Claude Mythos: 150 users - Wu said the model had only been given to a very limited number of users, implying de facto restriction. AI model approval period: 30 days - He said the new requirement shortened a previous 90-day government approval cool-off period to 30 days. German cyberattack cost estimate: 300 billion euros - He used this as an example of the scale of cyber risk when discussing frontier-model misuse. Anthropic ARR estimate: $9 billion to $42–44 billion - He referenced a large stated jump in annual recurring revenue, while warning that token maxing and product competition distort the picture. China’s AI chip position: 2 generations behind - Wu said China is behind the U.S. in advanced AI chips, citing H200 versus Blackwell/Rubin. Software engineers at risk: 50% within 3–4 years - He predicted major disruption to software engineering jobs from AI tools. Google ad revenue share: 60% - He said Google’s earnings are heavily dependent on advertising, making the company cyclical. Digital ad revenue share of Google/Facebook business: 80–85% - He argued that ad revenue now dominates the business mix and limits further market-share expansion. Projected AI model user base for controlled rollout: 500–1,000 users - He suggested broader access would make attribution and containment of misuse much harder.

Pivotal Quotes: "Ironically, the inflation of the components that go into data centers have actually given, ironically, a boost to basically nominal earnings growth." — David Wu: Explaining why rising AI hardware costs can make earnings look stronger even when real capex weakens. "The bear case for AI is not that it's not powerful, but that it's too powerful and it's gonna be reined in by governments worldwide, including the US government." — David Wu: Summarizing his long-term bearish thesis on frontier AI. "AI is very productive... I use AI... I cannot live without AI." — David Wu: Clarifying that he is not anti-AI, but believes the market is overpricing beneficiaries and underestimating risk.

Implications: Investors should separate real AI adoption from optical earnings boosts, watch for regulation and commoditization, and expect more volatility in semis and hyperscalers. AI may still raise productivity, but winners may be smaller teams and infrastructure sellers, not necessarily the current market darlings.