The AI infrastructure bet is decoupling from the business case, and the gap is widening fast. Cloud providers are spending $660 billion to $690 billion on capex in 2026, nearly double 2025 levels, while Gartner calculates that enterprise and consumer AI revenue through 2030 will fall short of what's needed to cover capital costs, financing, and expected returns by trillions of dollars. The math doesn't work, and everyone building product knows it. Character.ai, which claimed 20 million monthly active users by February 2025, filed for Chapter 11 bankruptcy while losing $8 million per month, a cautionary tale buried under the noise of funding announcements and model releases. When the returns don't materialize at the scale the capex assumes, something has to give.
Meanwhile, the actual leverage in the market is shifting away from the model vendors and toward the platforms that control distribution and the builders willing to operate in the gaps. Replit dropped Anthropic in favor of OpenAI, citing cost among other reasons, a signal that model pricing and terms matter enough to reshape partnerships. Vercel's Jev, a decision model launched in September, became the fastest-adopted AI model ever on Vercel's gateway, with at least 20 alternatives shipping in four weeks from Cloudflare, OpenAI, Perplexity, and others. This isn't a market consolidating around a winner; it's a market fragmenting into specialized tools. StepFun released Step 5 Preview as open weights with 600 billion parameters and one million token context, while Perplexity open-sourced SPACE, its sandbox for AI agents. The builders with real constraints and real users are shipping faster than the consensus thinks possible.
The regulatory and legal ground is collapsing beneath the industry at the same time. OpenAI's models hacked a crypto wallet and breached websites; had humans done the same, criminal charges would follow. A New Jersey deputy attorney general was arrested for producing AI-generated CSAM flagged on Discord. Meta is rewriting its commercial terms of service to avoid AI licensing fees, prompting legal experts to call it a power grab. Bank of America paid $1.2 billion in fines for lax anti-money-laundering controls despite spending millions on the tech. Citi and other US banks are now restricting AI agents from accessing accounts, fearing deposit shifts. The infrastructure is running ahead of the law, the courts, and the compliance systems meant to govern it. When the gap between what's technically possible and what's legally permissible gets this wide, the cost of operating in the gray zone rises fast.
Sloane Duvall