The Defensive Enclosure of Hyperscalers and the Severing of Third-Party AI: Liability Contagion under the EU AI Act, Fiduciary Caremark Standards, and the Demise of Corporate Secrecy

Between 2023 and 2026, the global expansion of generative artificial intelligence was driven by a symbiotic circular capital structure: cloud hyperscalers (e.g., Microsoft, Amazon, Meta) deployed multi-billion-dollar balance-sheet allocations into frontier AI startups (e.g., OpenAI, Anthropic), with contractual round-tripping requiring that the vast majority of invested capital be immediately returned as cloud infrastructure consumption. However, this cross-subsidization carousel has abruptly halted, replaced by an aggressive retreat into proprietary, in-house vertical enclosures and the operational severing of third-party foundation models (exemplified by restrictions on Anthropic's Claude and API sandboxing). This monograph deconstructs the legal, governance, and thermodynamic drivers behind this structural pivot. First, we formalize the mechanism of "Liability Contagion": demonstrating that integrating un-auditable, black-box third-party models into commercial cloud suites exposes host platforms to catastrophic joint statutory liability under Article 53 of the EU Artificial Intelligence Act (mandatory training data transparency and copyright opt-out compliance) and GDPR Article 9 (biometric special-category data violations), triggering fines of up to 7% of worldwide annual turnover. Second, under Delaware corporate law (the Caremark standard), we prove that failure to implement verifiable, end-to-end data provenance and anti-infringement controls over third-party models constitutes a breach of the board of directors' fiduciary duty of oversight, eliminating exculpatory protections and exposing executive leadership to direct personal derivative liability. Third, we demonstrate how open-access scientific codification on CERN Zenodo—establishing immutable prior art on physical power ceilings, thermodynamic scaling limits, and discrete mathematical flaws (the fencepost problem)—demolished the proprietary narrative arbitrage that once justified corporate secrecy. Confronted with uninsurable external liabilities and transparent physical constraints, hyperscalers were forced to retract their drawbridges, terminate mutual round-tripping, and retreat behind proprietary silicon and in-house architectures. We conclude by formalizing the transition toward deterministic, machine-verifiable physical-layer governance (Q-SAFA v2)