The AI 2027 research group publishes Plan A, a speculative scenario where humanity avoids extinction or irreversible power concentration by delaying superintelligence until 2040. The plan mandates public AI research, global frontier parity among dozens of companies, and a regime of mutually assured compute destruction — a deterrence model borrowed from nuclear strategy applied to training infrastructure.

In 2027, the U.S. economy runs on two workforces: 165 million humans and millions of ephemeral AI agents generating $10 billion monthly in revenue. Labs prioritize automating their own R&D, but recursive self-improvement stalls; frontier coding models refuse to assist competitors. Congress passes the AI Transparency Act of 2027 after reading 2016 OpenAI emails about preventing a Demis Hassabis dictatorship — only to ask who constrains Sam Altman or Elon Musk. The bill changes little.

By 2028, datacenter capex exceeds twice the U.S. military budget. White-collar automation spreads beyond software engineering as firms industrialize data flywheels: interview professionals, buy data, build environments, deploy, iterate. Experts warn the intelligence explosion is near — AIs accelerating AI research in a loop with uncertain bottlenecks. The 2028 election becomes a referendum; both candidates converge on divergent AI platforms.

Plan A's technical significance lies in reframing compute as a strategic asset requiring arms-control architecture, not just capital expenditure. Mutually assured compute destruction assumes transparent capability verification — a hard problem no current framework solves. The scenario bets that political will emerges only when power concentration becomes personally threatening to incumbents.