In early 2023, Tome looked like a winner. Keith Peiris's startup, which spun up beautiful slide decks in minutes, became the fastest productivity tool ever to reach a million users, a figure that eventually climbed to 25 million, and it raised $80 million from the likes of Lightspeed, Greylock and Reid Hoffman. Then the growth stopped meaning anything. The users were students and small-business owners on free plans or $10 subscriptions, revenue flatlined around $3 million, and the professionals who might have paid real money never came because Tome could not reach their data. By late 2024, Peiris admitted he had built the wrong company. He shut it down in March 2025, laid off most of 70 employees, kept six, and eight months later launched Lightfield, unglamorous software that helps salespeople write follow-up emails and track client calls. It is now growing 80 percent a month with a thousand paying customers.
Tome is not an isolated story. It is the template. Across the first wave of lavishly funded AI startups, companies that raised enormous sums on a dazzling demo are quietly abandoning the thing they raised the money to build. Pika took $135 million to make an AI video generator and now builds agents and avatars. Poolside raised $620 million to train coding models, announced a giant Texas data centre with CoreWeave, failed to get the chips online in time, and split itself into two companies. What makes this strange is the timing. Pivots normally happen before a company has traction or a war chest. Here they are happening after both, because in AI the ground moves faster than the product can set.
The reason keeps coming back to the frontier models. "As frontier models end up taking more and more oxygen in the room," says Aditya Agarwal of South Park Commons, a subset of these hot companies will either make hard pivots or spin out the infrastructure they happened to build along the way. Capital, adds BCV's Christina Melas-Kyriazi, is what let them wander: raise on a hypothesis, and you can afford a "much more meandering path" than earlier eras allowed. The through-line is a migration from spectacle to plumbing. The demo that goes viral and the business that pays rent, it turns out, are rarely the same thing.
Some of the pivots are dressed as evolution, and some are survival. Character AI, once valued for a consumer chatbot and gutted when Google acqui-hired its founders and licensed its technology for $2.7 billion, has settled a wave of lawsuits, banned under-18 users at the cost of four million of them, dropped its own models for open-source ones, and reinvented itself around AI audio stories and microdramas. The evaluation startup Patronus abandoned catching model errors to build simulated "world models" for testing agents; those now make up 70 percent of its revenue and helped it raise $50 million at a $400 million valuation. Wispr killed a $14 million neural-signal headphone after Humane's AI-pin flop spooked it, and kept only the dictation software it had built as a side feature. That app now has millions of users.
Read one way, this is the market working: teams with talent and cash finding real demand and chasing it. Read another, it is a warning. When a company can raise nine figures on a prototype and still have to reinvent itself within two years, the lesson is that the application layer sits on ground the labs keep repaving. The survivors are the ones who noticed early, stopped chasing the sci-fi version, and started building what people would actually pay for. As Wispr's founders put it, they stopped chasing the dream and started building what people truly needed. It is a modest epitaph for an immodest amount of money.