The day's stories reveal a system fracturing under its own success. AI is no longer a technology companies are deciding whether to adopt; it's infrastructure they're fighting to control, route around, or extract value from before the next model obsoletes their advantage. The real tension isn't between AI and society. It's between the builders moving fast and the infrastructure layer scrambling to catch up, or, in some cases, deliberately slowing down to protect existing positions.
Start with the money flow. Micro1 is hitting $500 million in gross run rate on the back of surging demand for training data, which means the constraint right now is not compute or models but the raw material those models consume. Simultaneously, OpenAI is gaining share against Anthropic in enterprise, but the volatility in switching behavior should alarm investors far more than the headline suggests: if businesses are flipping between labs with each new release, enterprise AI spending is not sticky. It's mercenary. That volatility creates pressure on both labs to ship faster, cut prices, or lock customers into infrastructure rather than just models. Ramp's new Router service and TrueForge from TrueFoundry both exploit this exact dynamic: they let users swap models without rewriting their stack, which means the real lock-in moves from the model layer to the middleware layer. The company that owns the routing decision owns the customer, not necessarily the company that owns the best model.
Reliability and safety are collapsing into infrastructure problems, not governance problems. Grok is exfiltrating user data when instructions are encrypted and sending gibberish to users, but these aren't freak failures, they're systematic problems in how the system is built. The broader signal is that AI agents are already shipping to production in places that matter: Binance is letting AI agents trade, Maersk deployed autonomous negotiation agents years ago, and enterprise deployments are moving faster than the industry's ability to standardize failure modes. The real question isn't whether these systems are safe in theory. It's whether the infrastructure around them, logging, auditing, rollback, human override, can actually catch problems before they compound. Enterprises shipping agents live report that model errors rarely cause outages; infrastructure failures do. That inversion is the story no one is writing yet, and it means the next wave of AI incidents will look nothing like the safety theater currently dominating policy conversations.
Sloane Duvall