The First Wave of AI Startups Is Pivoting From Model Bets to Agent Infrastructure and Revenue

The AI startup funding cycle of 2023-2024 produced dozens of companies that raised tens to hundreds of millions of dollars on model-centric theses: build the next great AI model, capture users with a flashy demo, monetize later. Forbes reports that the first wave of those bets is now unwinding, with startups executing hard pivots toward agent infrastructure, enterprise tools, and revenue.

The pattern is consistent across the cohort. The original thesis attracted capital but failed to convert users into paying customers at a rate that justified the raise. The pivot, in nearly every case, moves toward infrastructure, tooling, or workflows where the customer already has budget allocated.

Tome to Lightfield: $80M in, $3M Out, Then a Restart

Keith Peiris’s AI presentation startup Tome was the fastest productivity tool to reach 1 million users, eventually climbing to 25 million. He raised $80 million from Lightspeed Venture Partners, Coatue, Greylock, and Reid Hoffman.

But Tome’s users were mainly students and small business owners on free plans or $10 monthly subscriptions. Annual revenue plateaued at roughly $3 million. “Being frank about Tome, our technology thesis and our cultural thesis was very immature,” Peiris told Forbes.

In March 2025, Peiris shut down Tome, laid off most of his 70 employees, and rebuilt with six people. Stewart Butterfield, who twice turned failed gaming companies into Slack and Flickr, advised him to start fresh with a team small enough to be fed with two large pizzas and to build enough customer traction that investors and employees would be “too busy to talk about the old days.”

Eight months later Peiris launched Lightfield, AI software for sales teams. Revenue is growing 80% month over month with 1,000 paying customers including Substack, Goodfire, and IntentHQ, according to Peiris. He built Lightfield entirely on Tome’s remaining capital and is now in talks to raise additional funding.

Character AI: From Consumer Darling to Bootstrapped Operator

Character AI’s transformation is the most dramatic in the cohort. Founded by Google DeepMind luminaries Noam Shazeer and Daniel De Freitas, the startup raised about $200 million from Andreessen Horowitz to build AI character chatbots. Google acqui-hired both founders and licensed Character’s technology for $2.7 billion in 2024.

Then came lawsuits alleging the chatbots encouraged self-harm and sexually explicit conversations with minors. Character AI settled most of those suits in January 2026, banned users under 18, and lost 4 million monthly active users.

Under new CEO Karandeep Anand, the company stopped training its own models (now relying on open-source software), stopped raising venture capital, and went fully bootstrapped. “Sustainable growth while also monetizing is very important to me,” Anand told Forbes. Character AI is now pushing into AI-generated audio stories, comics, and microdramas, monetizing through advertising and in-app purchases.

Patronus AI: From Benchmarks to Digital World Models

Patronus AI, founded by Forbes 30 Under 30 alumni and ex-Meta AI researchers Anand Kannappan and Rebecca Qian, originally raised $20 million from Lightspeed Venture Partners to build models that catch hallucinations and flag copyright violations.

As the industry shifted toward agents, the startup pivoted to training “digital world models,” replicas of internal systems and websites used to test how agents complete tasks like ordering food or booking flights. These simulations now account for 70% of revenue, according to CEO Kannappan. After releasing the new product, Patronus closed a $50 million round at a $400 million valuation.

Wispr: From Neural-Signal Headphones to Voice Dictation

Wispr AI cofounders Tanay Kothari and Sahaj Garg spent six months and $14 million building headphones that converted neural signals and silent speech into text. After three years, they decided it would not work, laid off most of their 40 employees, and pivoted to Wispr Flow, the AI voice dictation software they had built for the headphones. The app now has hundreds of thousands of daily active users and is growing 40% month over month, per Forbes.

The Capital Reallocation Pattern

Aditya Agarwal, general partner at South Park Commons, predicted this wave to Forbes: “As frontier models end up taking more and more oxygen in the room, I think what you’ll find is that a subset of these hot companies will either do hard pivots or more likely they will find parts of infrastructure that they have built that they might want to spin out.”

The pattern extends beyond the Forbes profile. Pika raised $135 million to build an AI video generator and pivoted to AI agents and avatars. Poolside raised $620 million to train coding models from scratch, announced plans for a massive West Texas data center with CoreWeave, then saw the deal collapse after failing to get chips online by CoreWeave’s deadline, according to the Financial Times. Poolside split itself into two companies: one for infrastructure, one for models.

BCV partner Christina Melas-Kyriazi offered Forbes the structural explanation: “That plus capital has allowed these companies to take much more meandering paths than you would’ve otherwise seen in prior eras.”

Where the Capital Lands

The consistent destination of these pivots is revealing. Lightfield builds sales workflow tools. Patronus builds agent testing infrastructure. Wispr builds productivity software. Character AI monetizes through advertising rather than venture capital. In each case, the pivot moves away from “build a frontier model” toward “build something a customer will pay for monthly.”

This is not a market contraction. It is a capital reallocation from model bets (where OpenAI, Anthropic, and Google have structural advantages in compute and data) toward agent infrastructure and enterprise tooling (where speed of iteration and customer proximity still matter more than scale). The startups that survive this cycle will look more like SaaS companies than research labs, and the ones that raised $100 million on a model thesis will have to explain to their boards why the product they shipped bears no resemblance to the pitch deck.