The do-everything AI just became a liability

Disney found out Sora was shutting down less than an hour before the rest of us. A billion-dollar partnership, vaporised by a phone call that arrived roughly when OpenAI's blog post was already in the CMS.

Sora's app closes on 26 April, with the API following in September. The service was costing an estimated $1 million per day while active users dropped below 500,000. In the same week, OpenAI unveiled GPT-Rosalind, a model built specifically for life sciences, with Amgen and Moderna as launch partners. The company that just killed its do-everything creative tool is now building models you need a lab coat to appreciate.

This isn't a pivot. It's a confession. The generalist consumer AI product, the one that does a bit of everything for anyone who wanders in, has no viable unit economics. Sora proved it with a $365 million annual burn rate and a user base smaller than most mid-tier mobile games. GPT-Rosalind proves the alternative: charge serious money to serious buyers solving serious problems.

The market agrees. Cursor is raising $2 billion at a $50 billion-plus valuation, nearly doubling its last round from six months ago, on the strength of doing one thing well: AI-assisted coding. Over $1 billion in annualised revenue, with expectations to hit $6 billion by year-end. Cursor doesn't generate videos or help you plan dinner. It completes your functions and understands your codebase. Enterprises are writing cheques for that specificity at a pace that makes Sora's burn rate look like a rounding error.

Meanwhile, Novo Nordisk signed a strategic partnership with OpenAI to embed AI across drug discovery, clinical trials, and manufacturing, with full integration planned by the end of 2026. The deal comes with strict data governance protocols, because when you're racing Eli Lilly for weight-loss drug dominance, you don't hand your molecular data to a chatbot that also writes limericks.

The conglomerate lesson

There's a parallel from corporate history that maps cleanly here. In the 1960s and 70s, conglomerates like ITT and Gulf & Western bought everything on the theory that professional management could run anything. Hotels, insurance companies, film studios. The market loved it, then punished it savagely. By the 1980s, investors demanded breakups, not acquisitions. Focused companies outperformed.

The way I see it, AI is hitting the same inflection. The "we do everything" model was plausible when the technology was novel and no one knew which applications would stick. Now we know. Coding assistants stick. Domain-specific scientific models stick. Consumer video generation at $1 million a day does not.

For builders, the implication is uncomfortable but specific. If your AI product's value proposition needs the word "and" to make sense, you're building a conglomerate in an era that rewards specialists. The $50 billion isn't flowing to the model that does everything. It's flowing to the product that does one thing so well that switching costs more than the subscription.

Pick the lane, or become the next Sora: impressive, expensive, and shutting down on the 26th.

---