AI companies face a fundamental business model crisis that neither critics nor boosters fully grasp, according to new research from Princeton and New America fellows.

The analysis by Arvind Narayanan and Akash Kapur argues that frontier model providers like OpenAI and Anthropic cannot sustain profitable businesses selling inference tokens. Models remain largely undifferentiated, switching costs stay low, and competitive dynamics push pricing toward marginal costs.

This creates what the researchers call a "commodity trap" — similar to historical infrastructure buildouts in railroads, telecom, and aviation that destroyed trillions in investor capital despite creating enormous societal value.

The Enterprise Software Escape Route

The labs' likely response involves migrating up the technology stack through vertical integration and enterprise deployments. This mirrors successful strategies from cloud computing and semiconductor fabrication — the only infrastructure businesses to escape commoditization.

Enterprise software sustains 75% gross margins through three mechanisms: zero marginal reproduction costs, deep switching costs, and non-ephemeral value that amortizes fixed costs over decades. AI companies are already importing these properties through embedded deployments and managed services.

The shift explains recent moves by major labs toward specialized enterprise offerings rather than pure API access. Companies like Harvey for legal work and Hippocratic AI for healthcare represent this vertical integration trend.

Competition and Lock-in Concerns

While current AI markets appear competitive with interchangeable models, the researchers warn this could change rapidly as companies build switching costs and proprietary integrations.

Historical precedent suggests infrastructure builders during "installation periods" rarely capture long-term value. The telecom buildout of the late 1990s saw capacity explode 186,000-fold in seven years before prices crashed and $2 trillion in market value vanished.

The analysis suggests AI's current competitive dynamics may be temporary. As labs move toward enterprise software models, concerns about monopolistic concentration could materialize before switching costs become apparent to customers.

The research challenges both AI critics focused on current losses and boosters citing revenue growth, arguing both camps miss the industry's structural transformation toward higher-margin, stickier business models.