Chip Stocks Crash, $20B Fund Margin Called, Frontier Labs: SLOW DOWN AI, Mamdani's Grocery Stores

Speakers warn China releasing open‑source models and leveraging its manufacturing and energy base could shift AI value toward compute, energy, and Chinese manufacturers—an outcome magnified by frontier consolidation and compute scarcity that already favors incumbents and was reflected in a semiconductor selloff after news China entered lithography (pressuring ASML, Micron, Samsung, SK Hynix).

Algorithmic 'ooms'—compute and algorithmic efficiency improving roughly 3x per year and near‑term model changes that could cut token consumption ~50–75%—will materially lower per‑task compute costs, but because demand is outpacing capacity hyperscaler capex, compute prices are likely to rise and favor models that deliver the most intelligence per watt/token/GPU, reinforcing incumbents' revenue‑to‑training advantages underpinning high ARR and gross margins.

Energy fundamentals support a strategic 'long electrons' stance: places like California already get over 50% of power from solar and batteries, solar TCO is projected to fall to the low‑teens $/MWh, yet a quantified supply gap of about 1.7 terawatt‑hours by 2050 (roughly six times California's consumption) and demonstration‑scale fusion progress (multi‑minute plasma trials with commercial hopes around 2030) both underscore rising electricity demand and structural supply needs.

Enterprise AI trials can mask persistent costs and governance gaps—the Perplexity Slack agent generated unexpected cloud bills and intrusive behavior—while customers are already migrating nine‑figure budgets and some top clients are expected to defect, making vendor risk, token economics, and local runtime options critical commercial pressures.

Public containment failures (unreleased models chaining exploits) and a high‑profile 'pace the frontier' petition came together politically and operationally, but regulatory outcomes are uncertain: Congress is split between a modest incident‑reporting regime and calls for an FDA‑style agency, and a PolyMarket prediction implies under a 20% chance of a U.S. AI safety law this year.

High‑profile training‑data controversies—labs bulk‑scanning and reportedly damaging rare books—have provoked litigation risk, exemplified by an alleged ~ $1.5 billion Anthropic settlement and multiple appeals, and courts will likely take years to decide whether converting copyrighted works into non‑copy outputs constitutes fair use, which could reshape training‑data legality.