The week reveals a market fracturing along lines of control: closed models betting on regulatory moats and enterprise lock-in, open-weight alternatives gaining traction precisely where they matter most, and the infrastructure layer becoming the real prize. OpenAI's copyright settlement with Anthropic closes one legal front but leaves the core question unanswered, while simultaneously the company signals fear of open-weight competition from China, a tell that the moat is narrower than the rhetoric suggests. Trump's AI czar position has become a revolving door, with David Sacks departing after public fights with major labs over Chinese models, revealing that even within the administration there is no consensus on whether the threat is foreign competition or domestic incumbents defending turf. The practical answer is arriving faster than the policy: when a Hugging Face breach required forensics, defenders reached for an open Chinese model because US frontier guardrails locked them out of their own tools.
Infrastructure and efficiency are where the real competition is moving. Google is building custom silicon for Gemini efficiency. Oracle faces a $7 billion collateral bill for its Wisconsin data center as power costs mount. NVIDIA released Cosmos 3 Edge, a 4-billion-parameter model that runs on-device for robotics. Meta open-sourced Astryx, its internal design system with 150 components and an agent-ready CLI, after running it across 13,000 apps for eight years. Anthropic donated MCP to the Linux Foundation after it reached nearly 100 million monthly SDK downloads across more than 10,000 servers. These are not gestures of altruism but rational moves by companies that realized the real value is not the model itself but the infrastructure layer that makes models useful at scale. The companies giving away tools are the ones that understand deployment wins over architecture.
Enterprise adoption remains shallow and conservative. UK businesses are not deepening AI use; free tools dominate, and the focus is efficiency gains rather than new products. California's XPRIZE tests firefighting drones. YouTube clarified policies on AI-generated content and low-quality video monetization. The EU's AI transparency deadline arrives August 2, requiring explicit disclosure when users interact with AI. These are not barriers to deployment; they are the minimum floor. What matters is that the companies that control the infrastructure layer, the protocols, the design systems, the local models that run without upstream dependencies, are the ones positioned to survive tightening regulation and shifting geopolitics. The companies betting everything on model scale and closed weights are discovering that control is cheaper when you give it away.
Sloane Duvall