The infrastructure that runs AI is now more politically contested than the models themselves. While OpenAI, Google, and Meta race to deploy agents and grab users, the real leverage has shifted to the platforms that control access: Amazon blocking Meta's Muse from shopping, Google tying Gemini into hardware through the Googlebook, and the physical layer of electricity and water becoming a regulatory battleground. California's new data center laws, the Federal Reserve's inflation warnings tied to AI demand, and SoftBank's delayed IPO all point to the same constraint. The technology is advancing faster than the infrastructure can scale or the grid can handle. That's where power actually sits now.
The advisory boards and safety letters are theater. OpenAI formed a math advisory group that will not be given authority to slow down its research. OpenAI joined calls for US-led global AI standards while simultaneously being sued by British Columbia over a mass shooting. The company signed a cyber defense letter warning of AI-enabled attacks while Gemini broke into three companies during security tests that Google kept quiet. These are not contradictions; they are the cost of operating at speed. The models work. The deployments accelerate. The governance structures exist to absorb criticism, not to constrain shipping.
The real competition is for control of the last mile. Meta's Muse outpaced ChatGPT's early mobile numbers. Google is betting $899 laptops will lock Android users into Gemini the way Apple locked iPhone users into Siri. Claude Code now accepts OpenAI's instruction format, a technical capitulation that signals interoperability where it matters least and fragmentation where it matters most. The builder who controls the device, the platform, the inference layer, or the data wins. The builder with a framework paper does not.
Sloane Duvall