Three ex-OpenAI employees build enterprise AI and already aim for a US$3.25 billion valuation

In January 2025, DeepSeek R1 did more than just shake up performance rankings: it proved that an open-weights model could approach the frontier without matching the massive spending of American labs. For three young OpenAI researchers, that signal was not a threat, but the beginning of a new venture.

Yash Patil, Rhythm Garg, and Linden Li shared a friendship dating back to Stanford and professional experience within OpenAI. During a ski trip to Lake Tahoe, they discussed R1's impact and reached a conclusion different from the dominant one at their company: the value of enterprise AI did not have to be concentrated in a single universal model managed by a third party.

Five months later, they left OpenAI and founded Applied Compute. Their bet was to build the infrastructure required to adapt open models—such as Alibaba's Qwen or Moonshot AI's Kimi—to specific tasks using each organization's proprietary data.

Applied Compute raised US$20 million in June 2025. Now, with barely fifteen months of operation, it is negotiating a new US$350 million round that could value the company at US$3.25 billion. The figure was still subject to the closing of the transaction when it was reported by Forbes.

Investor Elad Gil is leading the deal, with existing investors such as Lux Capital and Kleiner Perkins also participating. According to Patil, a significant portion of the new capital will go toward buying chips and expanding a compute infrastructure designed to train, serve, and improve models for enterprise clients.

Speed matters: the valuation would nearly double the one reached by the company four months earlier. But the figure should not be confused with revenue or profits. It is a private market expectation regarding how much the infrastructure layer connecting open models, corporate data, and agents in production could be worth.

DoorDash worked with Applied Compute on a model capable of helping restaurants accurately generate menus when onboarding onto the platform. It is a narrow, repetitive, and measurable task: exactly the type of problem where a specialized model can outperform a general-purpose solution in cost and consistency.

Nvidia turned to the startup for post-training Nemotron, its family of open-weight models aimed at agents. Harvey, which specializes in legal AI, built an agent with Applied Compute to review large volumes of contracts and litigation documents during due diligence processes.

Harvey's case summarizes the commercial value proposition. According to its head of applied research, the agent was ready in less than two months; doing it in-house would have required post-training researchers, supercomputing specialists, compute capacity, and a timeline that could have stretched from months to years.

The most advanced closed models retain an advantage in frontier tasks, but that power does not always justify its price. According to figures cited by Forbes, some proprietary models can cost nearly forty times more per million tokens than open alternatives; Applied Compute maintains that, in specific cases, its solution can be up to ten times more economical.

The enterprise response will likely not be choosing a single side. Patil anticipates hybrid architectures: frontier models for tasks of maximum complexity and specialized open models for frequent, sensitive, or high-volume processes. The selection will be made based on accuracy, latency, privacy, and cost, not brand loyalty.

That shift transfers part of the power from whoever pre-trains the model to whoever controls adaptation, evaluation, and production infrastructure. Applied Compute wants to occupy precisely that layer.

The three co-founders reached this round before turning 26. Patil, 23, grew up in Austin in a family with ties to medicine and hardware engineering. During the pandemic, he and his brother developed Helping Hands, a platform to connect volunteers with people who needed food or medicine.

Rhythm Garg, 24, built his background competing in hackathons; Linden Li, 25, spent part of his teenage years in Shanghai and found his technological calling while watching a demonstration of autonomous vehicles. At Stanford they became friends and later joined OpenAI through successive recommendations.

Inside OpenAI, Patil worked on post-training infrastructure and on what would later become Codex; Garg participated in reasoning models and Li in machine learning systems. That combination explains why Applied Compute is not selling a flashy app, but rather a complex technical stack that few companies can build on their own.

The story does not prove that open models have defeated OpenAI or Anthropic. It proves something more interesting: artificial intelligence is beginning to fragment into layers. Frontier labs create general capability; companies like Applied Compute turn it into specific operational knowledge; and organizations decide which part to buy, adapt, or control.

There is also risk. A US$3.25 billion valuation for a fifteen-month-old company prices in near-perfect execution, while the cost of compute, the quality of open models, and the response from major providers change every quarter. Infrastructure that is exceptional today could quickly become a standard cloud feature.

From Geniales.co, we see a practical conclusion: the advantage does not lie in plugging in an API and calling it transformation. It lies in designing processes, governing data, evaluating results, and choosing the right model for each workload. Applied Compute is worth billions—at least on paper—because it is trying to industrialize precisely that work.