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

David Woo argues that combined capex across Microsoft, Google, Amazon, Oracle, and Facebook declined quarter over quarter in Q1 for the first time in three years, and that year-on-year growth also slowed relative to Q4. Because component prices from suppliers like Micron and Samsung rose during the same period, real capex volume fell even more than nominal figures suggest. The capex-to-operating-income ratio for the five hyperscalers has reached 135 percent, meaning they can no longer self-fund investment from cash flows, with Google reportedly looking to raise 80 to 85 billion dollars in equity to cover capex needs.

Woo identifies an accounting asymmetry he says creates an optical illusion of AI earnings strength. Hyperscalers expense only a small depreciation fraction of capex as cost, while suppliers like Micron book near 100 percent margin on price increases. This income transfer from hyperscalers to component makers inflates aggregate index earnings growth without reflecting real output expansion. Google and Amazon partially offset their capex burden by booking large revaluations of Anthropic holdings, which Woo characterizes as artificial earnings support. Microsoft pulled back capex partly because corporate adoption of Copilot disappointed and because it lost its exclusive arrangement with OpenAI.

Woo argues that Q1 AI usage figures were artificially inflated by companies such as Amazon ranking software engineers on AI token consumption, incentivizing token-maxing behavior rather than productive use. Uber's CEO stated the company burned through its entire annual token budget in four months with little to show for it. Anthropic's stated annual recurring revenue grew from roughly 9 billion dollars in December to 42 to 44 billion dollars, but Woo attributes much of that acceleration to Claude Code and to token-maxing, both of which he believes are now decelerating. Claude Code faces competition from at least five comparable tools, and Woo predicts a meaningful slowdown in both AI usage and capex in Q2.

Woo argues the central risk to AI investment has shifted from a frontier model capability plateau to regulatory risk stemming from models being too capable. Claude Opus, which he describes as capable of identifying network vulnerabilities and enabling cyberattacks, has so far been distributed to only 150 users. Anthropic cannot meaningfully monetize the model at that scale, but expanding access to 500 or 1000 users makes it practically impossible to prevent adversaries including China from obtaining it. He notes that Claude Mythos reportedly has recursive learning capability, meaning version one can create version two without human intervention. Trump signed a measure reducing a mandatory cool-off period for new large language models from 90 days to 30 days, which Woo characterizes as a reluctant but meaningful step toward broader regulation. He predicts AI will become a major political issue heading into the midterm elections, partly because those most exposed to AI-driven displacement are concentrated in Trump's political base.

Woo rejects the winner-take-all framing the stock market applies to AI companies by analogy to Google in search. He argues that every time one model surges ahead, competitors catch up within roughly three months, and that large language models will commoditize because they all learn from the same public data. On chips, he argues inference chips are far less complex than training chips, allowing Intel, AMD, and Broadcom to compete with Nvidia, and that agent-based AI runs primarily on inference chips rather than GPUs. He considers a compute glut replacing the current shortage to be very likely, and predicts high bandwidth memory prices will fall sharply once new capacity comes online around 2027 to 2028, with Chinese competitors matching current producers within two to three years. His current positioning is short the Nasdaq or semiconductor index, bearish equities overall, bullish oil, and bearish gold.

Despite predicting 50 percent of software engineers will be eliminated within three to four years, Woo frames the larger economic effect as democratization rather than pure displacement. He argues AI tools now allow a team of three people to replicate software infrastructure that firms like Renaissance Technology and Citadel built over decades with thousands of engineers. He also notes that Anthropic co-founder Jack Clark warned at Oxford that AI could bring about human extinction on a shorter timeline than decades, and that AI companies recently wrote a joint letter to the government requesting controls on ingredients that could be combined with AI to engineer bioweapons capable of a mass extinction event.

On oil, Woo says he cannot find a single hedge fund client currently long oil despite the Strait of Hormuz being effectively closed. He predicts the Iran conflict will not resolve quickly and that most plausible endgame scenarios result in higher prices, with oil executives warning Trump that prices could reach 150 dollars per barrel if the strait remains closed through the end of July. He argues Trump's repeated public statements that a deal is days away have backfired by convincing Iranian negotiators that Trump is desperate, hardening their position. Iran has fired missiles directly at Israel and targeted a US helicopter in ways Woo interprets as signals that Iran believes it holds the stronger hand and is willing to terminate negotiations entirely.