AI Market and Regulation

A Silicon Valley-backed political machine and a fight over AI “agent” software are converging on the same question: who gets to set the rules. (sanjosespotlight.com) In Washington, the super PAC Leading the Future is backing candidates who favor lighter AI rules, while the Trump White House on March 20, 2026 urged Congress to preempt state laws it says place “undue burdens” on developers. The PAC’s supporters include Andreessen Horowitz, OpenAI president Greg Brockman, Joe Lonsdale of 8VC and Perplexity. (cnbc.com; politico.com; prnewswire.com) That push has landed in House races. San Jose Spotlight reported on April 18 that Representative Sam Liccardo is facing pressure from child-safety and watchdog groups to reject the PAC’s endorsement, while his office said he supports federal preemption “within reason” and does not expect PAC money. (sanjosespotlight.com) At the same time, AI companies are moving competition away from the raw model and toward the “harness,” the software layer that keeps an agent on task, manages tools, tracks costs and decides how work gets billed. The New Stack reported on April 18 that Anthropic charges $0.08 per session hour for one harness product while OpenAI has released an open-source alternative. (thenewstack.io) That billing layer matters because the company that owns the harness can decide whether customers pay by token, by hour, by task or through bundled software. It also controls which model gets called underneath, which can turn the model into a supplier and the harness into the storefront. (thenewstack.io) The desktop is becoming the next distribution fight. The New Stack reported on April 18 that Google and OpenAI are shipping fuller desktop AI products as Anthropic, whose Claude had built an early lead on developer desktops, has stumbled through recent launches and product missteps. (thenewstack.io) Cheaper models are adding pressure from below. PrismML said on March 31 that its Bonsai 8B uses about 1.15 gigabytes of memory for an 8.2 billion-parameter model, and MarkTechPost published a CUDA deployment tutorial on April 18 showing how developers can run Bonsai through an optimized GGUF stack for chat, JSON output and retrieval-augmented generation. (prismml.com; marktechpost.com) States are still writing their own AI laws while Congress has not passed a single national framework, which is why preemption has become the central political demand from industry groups and the White House. Legal analyses from Baker Botts, Freshfields and Ropes & Gray all describe a patchwork of state rules and a new federal push to override conflicting state measures. (bakerbotts.com; blog.freshfields.us; ropesgray.com) The result is a two-front contest: one over who writes AI law, and another over who owns the software layer that turns models into paid products. The companies and politicians that win those fights will have more say over how AI is distributed, priced and governed in 2026. (sanjosespotlight.com; thenewstack.io; politico.com)