The regulatory pressure on AI companies is intensifying just as the market's faith in AI infrastructure is cracking. The FTC is demanding safety incident records from OpenAI and Anthropic with the same force it uses against companies that hide data breaches, while eight state attorneys general are suing OpenAI over ChatGPT's impact on children. At the same time, investors are abandoning data center bonds issued by AI infrastructure providers, debt from companies like CoreWeave has sunk to stressed levels, and planned IPOs from Australian data center operators have been shelved. The signal here is not about regulation versus innovation. It is about who bears the cost when the bet goes wrong. Anthropic secured junk-rated debt in June and is now valued at $1.5 trillion in private discussions, down from $2 trillion. The company has also handed METR researchers employee-level access to its systems and given up the right to block their published assessments of risks and incidents. This is not a sign of confidence. It is a company preparing for a world where the absence of transparency becomes more expensive than the cost of admitting what you do not know.
The infrastructure that was supposed to be the bedrock of AI abundance is showing stress fractures. Google is now offering daily token quotas for BigQuery's AI functions because enterprises cannot predict their bills. Microsoft is shifting Copilot pricing from fixed to consumption-based models, which Gartner reports is among the most common questions it receives from clients trying to understand their costs. Amazon's devices chief is describing a future where ambient AI anticipates user needs, but Amazon is simultaneously selling new personal AI devices starting at $74.99 that summarize conversations and create to-do lists. The actual products in market are far more modest than the vision. Meta released Muse Spark Open, an 816-billion-parameter model that runs on Mac hardware, but this is an open-weight model that users download and run locally, not a service that generates recurring revenue. The cost structure of AI is becoming visible, and it is not the one that venture capital funded.
Accountability is the word that appears nowhere in the headlines but underlies all of them. When an AI agent makes a decision, who pays if it fails? When a model is compromised, who notifies users? When an employee records colleagues in an OpenAI office, who is responsible for security? ServiceNow is rebranding its AI Control Tower to Super Intelligence Control Tower and adding verifiable identities for agents, but the core problem remains: systems that act without a human in the loop create liability chains that existing law does not recognize. Researchers are calling for air-gapped systems to train models safely. Security vendors report that TypeSafe's Jev model is more vulnerable to adversarial attacks than traditional large language models. The builders are moving faster than the infrastructure for accountability, and the market is beginning to price that gap.
Sloane Duvall