Former OpenAI Researcher Founds Applied Compute to Negotiate $3.5 Billion Valuation Financing
Three Stanford classmates who worked at OpenAI have left to found a company called Applied Compute, which trains custom models for enterprises using open-source weights and runs them on its own cloud. According to Forbes, about 15 months after its establishment, the founding team in their twenties is negotiating $350 million in financing, with a valuation of approximately $3.25 billion.
The founders are Yash Patil, Rhythm Garg, and Linden Li. Patil stated that their worldview differs from many former colleagues: if open-source weights are strong enough, institutions like JPMorgan will train models of interest using their own data, as they will not share their data. The company completed a $20 million seed round in June 2025, signing clients like DoorDash, Nvidia, and Microsoft within a year. Patil has previously discussed enterprise AI with Satya Nadella.
The new round is led by angel investor Eli Gil, with existing shareholders Lux Capital and Kleiner Perkins participating. The company notes that figures may change before finalization. The new valuation is nearly double that of about four months ago. Patil stated that the funds will primarily be used to purchase chips and expand "large-scale computing clusters."
The product form involves privatizing the training of open-source models and hosting them on the startup's cloud, rather than sending enterprise data to a closed flagship model. The narrative targets industry clients who are reluctant to share core data but need custom models.
Mechanically, the buyers are enterprises that want to retain their own data while having ready-made training and inference clusters, while the sellers are teams that have left OpenAI and can connect open-source weights to client corpora. The event is driven by the sufficiency of open-source models and enterprises' reluctance to share data, with funding entering GPUs from the new equity round; the beneficiaries are new clouds that can sell both training and hosting, while the pressured parties are closed laboratories that insist on training a supermodel with "all data."
Public information has not disclosed whether this round has closed, nor the revenue and contract duration, and has not provided a price difference table with closed APIs.
Source: Public Information
ABAB AI Insight
This is a part of the wave of departures from OpenAI that is more about "selling shovels to banks": it does not aim to recreate a consumer-grade chatbot but acknowledges that enterprise data will not flow into supermodels. The three founders have packaged the training processes from the lab into a contractable open-source custom cloud, with valuation growth relying on client lists and computing power procurement stories rather than public benchmark dominance.
The capital path is clear: seed round validation contracts, followed by a new round for card expansion. The motivation is to write a standard contract of "your data, our cluster" before open-source weights catch up with closed models. Resource mobilization relies on the training experience of former employers, along with checks from Lux, Kleiner Perkins, and individual lead investors, turning computing power inventory into valuation steps.
Comparable entities include Databricks turning open-source engines into enterprise lakehouses, CoreWeave renting GPUs as new clouds, and Anthropic taking a different closed route after leaving OpenAI. In terms of industry positioning, enterprise AI has shifted from "accessing flagship APIs" to "retraining on proprietary data using open-source"; supermodel companies want data from all over the world, while banks want their own data to stay in-house.
The structural judgment belongs to the transfer of pricing power: pricing power has shifted from "whose general model is the strongest" to "who can cheaply create sufficient proprietary models on client data." The mechanism is that open-source weights have lowered the training barrier, making the scarce items client corpora and compliant hosting cards; departing researchers use methods from their former employers to sell data that they are unwilling to share with their previous companies.
ABAB News · Cognitive Laws
Enterprises will not share core data with supermodels
After open-source sufficiency, what is sold is training and hosting
People who come out of the lab often turn what their original company was unwilling to do into a business