The frontier AI labs are fracturing along a new fault line: between those willing to build for government and those betting their entire business model on refusing to. Anthropic's lawsuit against the Pentagon over its supply-chain-risk designation has triggered a remarkable show of solidarity from OpenAI and Google DeepMind employees, with more than 30 signing an amicus brief in support. The designation itself appears to have already cost Anthropic material revenue, with executives claiming companies paused deal talks after the Trump administration's move. Yet this isn't a case of the industry united against overreach. OpenAI has instead moved in the opposite direction, acquiring Promptfoo to shore up its ability to deploy AI agents in critical business operations, while Caitlin Kalinowski, the company's head of robotics, resigned over OpenAI's Pentagon contract, specifically citing inadequate review of surveillance and autonomous weapons safeguards. The split reflects a deeper disagreement about what builders should accept in exchange for scale and legitimacy.
Meanwhile, the money is flowing toward infrastructure and specialized models rather than consolidation around any single foundation. Yann LeCun's AMI Labs closed a $1.03 billion seed round at a $3.5 billion valuation to build world models focused on physical understanding, while Nscale, an Nvidia-backed infrastructure startup, raised $2 billion to reach a $14.6 billion valuation. Anthropic itself launched a Claude Marketplace to streamline enterprise procurement and deployed Code Review, a multi-agent system for analyzing AI-generated code. These moves suggest the market is settling into layers: frontier model providers compete on capability and trust, infrastructure companies capture the deployment economics, and specialized tools fill the gaps between raw models and production use. Anthropic's marketplace and Anthropic's code review tool both aim to reduce friction between what enterprises need and what the models can deliver.
The practical pressure on builders is now acute. Amazon held an engineering meeting after AI-related outages linked to generative AI-assisted code changes. Microsoft's Copilot for Microsoft 365 has captured only 3 percent of its customer base despite two years in market, forcing the company to add Anthropic's Claude to its own tools and replicate features like Anthropic's Cowork. The market is no longer asking whether AI works in theory. It is asking whether it works reliably enough to deploy at scale, whether it can be audited and reviewed, and whether the people using it can understand what it does. The companies that win will be those that solve these problems faster than their competitors, not those that promise the most capability.
Sloane Duvall