The Race remains Mixed-contested, with September strengthening public counterweights to concentrated AI scale. Governments, schools, unions and shared public institutions gained more ways to set terms for deployment, infrastructure and consequential decisions. Those gains are real but uneven. They constrain particular uses and bargains more often than they redistribute ownership of frontier models, compute or operating platforms.
The most important movement is the connection between public authority and the infrastructure AI needs. Massachusetts tied state permitting for large data centers to community benefits and energy obligations. California enacted a coordinated package addressing electricity, water and local review. Municipal restrictions and special-use review preserved opportunities to negotiate before development rights hardened. These are mechanisms through which communities can influence scale even when they do not own the facilities. Adopted permit gates differ from future implementing rules, temporary pauses and advisory consultations; those stages must remain visible.
Public capacity also acquired more durable support. NSF's five-year NAIRR operating award creates a shared connective institution. Canada's literacy funding and Mitacs placements address different needs: understanding AI and applying expertise inside organizations. USAi supplies a continuing federal shared service, while the September GSA–OpenAI agreement lowers procurement barriers for several levels of government beginning October 1. These routes can help smaller institutions operate at useful scale. Their value depends on whether learning, data, workflows and purchasing choices remain under institutional control. A subsidized route into one provider can expand capability and strengthen that provider's market position simultaneously.
The concentration pressures remain strongest at three points. Frontier cybersecurity access is allocated by provider and national-security gatekeepers. Platforms retain the ability to admit some assistants and exclude others, even as standardized interfaces create limited separations. Employers largely control how AI changes tasks, staffing and performance measures. California's September worker laws establish meaningful counterweights, but most protections and reporting systems are prospective and do not establish a broad shift in workplace bargaining power yet.
Smaller and open-weight models create an additional route to contestable scale. Taiwan's pharmacy deployment makes the practical mechanism visible: a professional institution can run a specialized system locally using public drug data and open weights. Licensing and model size alone do not establish independent operation; hardware, integration, maintenance and evaluation remain necessary. September improves the evidence for local capability without demonstrating parity with frontier systems or universal portability.
Public finance continues to lean toward dispersion in the narrower, demonstrated sense of setting terms for infrastructure costs and public exposure. Governments are developing cost-allocation, subsidy and community-benefit mechanisms. The larger question of capturing AI-generated surplus to sustain services and social insurance remains open. Research explains how a shift from wages to capital could alter receipts, but it does not show an adopted replacement tax architecture or an observed AI-caused fiscal collapse. The federal Ratepayer Protection Act's September 30 Senate setback illustrates the gap between broad legislative support and an enacted national mechanism.
The month-end judgment is therefore an improvement in the capacity to contest concentrated power, rather than a general redistribution of its ownership. The next test is implementation: enforceable obligations, funded public expertise, usable independent deployment, worker remedies and credible exit routes.