The Inference Report

April 18, 2026
From the Wire

The infrastructure bill has come due, and the people writing the checks are starting to ask harder questions about what they're actually buying. Data center construction is slipping behind schedule while energy constraints tighten. Component prices are rising fast enough that Meta is passing costs directly to consumers through more expensive Quest headsets. OpenAI, meanwhile, is aggressively consolidating: shedding Sora, folding its science team, buying finance apps and talk shows, and watching senior executives like Kevin Weil and Bill Peebles head for the exits as the company pivots toward enterprise-only work. The pattern is unmistakable. The venture-backed moonshot phase is over. What remains is the question of who actually pays for the infrastructure these models require and whether the returns justify the spend.

Cursor is raising $2 billion at a $50 billion valuation on the back of enterprise adoption, while Recursive, a months-old startup founded by former DeepMind and OpenAI engineers, just closed a $500 million round at $4 billion with backing from Google Ventures and Nvidia. The money is flowing toward tools that integrate into existing workflows, not toward research labs or consumer applications. Anthropic launched Claude Design as a visual tool for founders and product managers. Oracle is selling semantic search without LLMs, positioning vector-based retrieval as the enterprise-grade alternative to RAG systems. The message is consistent: builders who can ship something enterprises will actually use are getting funded. Those chasing incremental research gains or consumer experiences are not.

Meanwhile, Sam Altman's World ID is expanding its Orb-based verification system into Tinder and other partnerships, betting that proof-of-humanity infrastructure will become essential as AI agents proliferate. The UK government committed £500 million to build sovereign AI champions, a rounding error against OpenAI's $852 billion valuation but a signal that governments are watching concentration happen and moving to hedge it. The gap between AI insiders who understand the cost structure and everyone else is widening. Developers chasing token optimization are writing more code that costs more to run and requires more rewriting. Newsrooms are adopting AI-assisted writing under efficiency banners while journalists push back on the actual tradeoff. Federal agencies are preparing to deploy Anthropic's Mythos model despite warnings it will compress the time between vulnerability discovery and attack. The industry is spending like it's certain of returns while operating like it's uncertain of anything else.

Sloane Duvall