AI Power Struggle: Coercion, Compute Sovereignty, and Platform Lock-In

The Trump administration's AI governance approach has crystallised into a coherent if legally contested strategy: reward aligned companies through federal contracts and DOJ interventions, and punish uncooperative ones through procurement restrictions and national security authority. The DOJ's filing to dismiss the NAACP's environmental lawsuit against xAI and the Pentagon's reported pressure campaign against Anthropic are not isolated episodes — they are the same instrument applied to different ends. Both cases use executive tools that bypass notice-and-comment rulemaking, congressional authorisation, and formal regulatory process.

The legal and legislative responses are now catching up. EFF's First Amendment challenge to the Pentagon's Anthropic posture, and the Cruz-Wyden JAWBONE Act creating a private right of action against government officials who coerce AI providers, together represent the first systematic architecture for constraining executive coercion as a regulatory substitute. Neither has yet succeeded — the litigation is active and the bill unenacted — but the framing they establish is significant: if courts accept that procurement leverage used to punish public policy positions is unconstitutional, any future administration will be forced to govern AI through formal rulemaking rather than informal pressure. For enterprise AI vendors, Washington alignment is now a material business risk variable, not a political preference.

Three distinct sovereignty postures are hardening in parallel. The US is tightening data residency and compliance requirements to keep AI compute within its borders — a posture that directly constrains European operators dependent on US hyperscaler infrastructure. China is assembling a closed-loop ecosystem spanning chips, rockets, and AI laboratories to build compute capacity physically beyond jurisdictional reach. Europe is pledging sovereign capacity while remaining structurally dependent on the US infrastructure that American policy is simultaneously pulling inward. Macron's G7 AI infrastructure commitments remain substantially at the announced rather than confirmed-build stage, and the gap between headline figures and contracted capacity is widening.

The infrastructure realities are compounding the policy fragmentation. Battery Energy Storage Systems are allowing data centres to come online faster than grid connections can support — decoupling buildout pace from transmission infrastructure in ways that create new grid stability risks. Ghost compute from networks like Pearl's alleged GPU-mining operation is inflating rental prices and contaminating the demand signals that infrastructure investors use to pace capacity decisions. And China's orbital compute ambition, while years from proven deployment, would create a category of sovereign infrastructure that export controls and data localisation frameworks are architecturally unable to address. Together these dynamics mean the global compute market is fragmenting faster than any bloc can replicate the supply chain chokepoints — TSMC, ASML, NVIDIA — it depends on.

Adobe's simultaneous public beta launch across its entire Creative Cloud suite — combined with Firefly's persistent project context and asset memory — is the clearest example of incumbents executing a platform enclosure strategy. The competitive threat to Adobe was never that a single image generator would produce better outputs; it was that the entire project workflow would migrate to a competitor's ecosystem. Persistent context memory directly addresses that risk by making Firefly's studio the place where professional creative projects live, not just where individual assets are generated. The same logic is playing out in enterprise software: Samsung's global deployment of ChatGPT Enterprise and Codex treats foundation model access as core operational infrastructure, generating workflow integration depth that creates durable lock-in long before any rival can displace the relationship.

Tencent's WeChat AI integration adds a distribution dimension to the enclosure dynamic. With over 1.3 billion monthly active users whose payments, commerce, and communication already run through the super-app, Tencent does not need to win an AI product competition — it needs only to make its AI assistant the path of least resistance for existing users. OpenAI's modular versioning strategy, deploying GPT-5.5 Instant for healthcare and reasoning models for rare disease diagnosis, reflects a parallel recognition that the competitive battleground has shifted from flagship model releases to domain-specific deployment speed and workflow integration. The window for standalone generative AI tools to establish defensible positions before incumbents close the feature gap is measurably narrowing, and capital allocators pricing software M&A must now account for Bain-style AI due diligence that can rapidly test whether a target's moat is genuine or feature-deep.