W hen DeepSeek launched its R1 open source model in January 2025, it stunned Silicon Valley. The idea that a small Chinese lab had created such a powerful open source model with comparatively fewer resources than their American counterparts left the tech industry’s heavy hitters baffled. Marc Andreessen called it a “ Sputnik moment .” Then-Scale CEO Alex Wang described it as a “ wakeup call .”

Inside OpenAI, it jolted three young researchers to realize the power of open source models. If they could build the infrastructure to optimize those models, the researchers thought, it could mean big business. “Our worldview was just very different than a lot of our OpenAI colleagues,” Yash Patil, one of the researchers, tells Forbes . “If open weight models are going to be really good, we don't see OpenAI or Anthropic having all of the data in the world to train this super-model. We more see the JPMorgans of the world using their own data to train models for the things that they care about — because they're never going to give that data away.”

Five months later, Patil, along with fellow researchers Rhythm Garg and Linden Li, left OpenAI to found Applied Compute, a startup that trains custom open source models for enterprise customers, then runs those models on the startup’s cloud. It’s been a whirlwind ever since. It raised a $20 million seed round in June 2025 and within a year had signed up big name clients like DoorDash, Nvidia and Microsoft, whose CEO Satya Nadella sat down for an interview with Patil in June about the future of enterprise AI. Then on Wednesday, Applied Compute told Forbes it’s in talks to raise $350 million in funding that would vault its valuation to $3.25 billion. The company cautioned some of those numbers could change as it finalizes the round. ( The Information earlier reported some of the details of the latest fundraising.) The new valuation would be nearly double Applied Compute’s worth from just four months ago, with its valuation growing faster than YC’s fastest unicorn .

Angel investor Elad Gil is the lead investor, with participation from existing backers including Lux Capital and Kleiner Perkins. Patil says the new influx of capital will go mostly to buying more chips to build out its cloud with a “massive fleet of compute.”

The company works with big enterprises to help them build special purpose AI models and agents from open source models like Alibaba’s Qwen or Moonshot’s Kimi, a much more cost effective proposal than relying on OpenAI or Anthropic. To fine-tune those models, the company relies on a method called “reinforcement learning,” which gives models feedback by rewarding them when they perform tasks well, which in turn encourages the model to adjust itself to gain more rewards.

Applied Compute sends its engineers to consult with enterprise customers to train models on their own internal data, before deploying the modes to run on its servers. The idea is to be a one-stop shop for training, deploying and running custom open-source agents. Last week, the company released AC2, a platform that lets clients use Applied Compute’s internal model training tools themselves — a cheaper option than relying on the startup’s high-touch consulting services.

The funding round comes as open source models have taken off in popularity, as companies try to curb the rising costs of their AI expenses. Flagship frontier closed-source models like Anthropic's Fable or OpenAI’s Sol are incredibly powerful, but can cost almost 40x per million tokens than some open source models. (Patil says in some cases their models can be 10x cheaper than frontier models.)

Meanwhile, as open source models have grown more and more capable, they’ve become suitable replacements when it comes to more straightforward tasks in some enterprise settings. “I always believed open source would play an important role in things,” Gil, the round’s lead investor, tells Forbes . The resume of the founders — all of whom are 25 or younger — made them a compelling team to tackle the issue of optimizing those open source models, he says. “Given that they're the former OpenAI research team that worked on these types of problems, they're really at the cutting edge.”

DoorDash has already worked with Applied Compute to build a model that lets restaurants accurately generate menus when they onboard with the delivery company. Nvidia used the startup to help post-train its open-weight Nemotron models for building complex agents. And the legal AI startup Harvey relied on Applied Compute to create an agent for reviewing massive volumes of litigation documents or reviewing contracts for due diligence. The whole process of building the agent took less than two months, Niko Grupen, Harvey’s head of applied research, told Forbes . Previously, he said, it would have taken several months to years — and considerable resources, including post-training researchers, supercomputing infrastructure engineers and tons of compute power. “It’s an order of magnitude difference.”

T he son of Indian immigrants — a pediatrician and hardware engineer — Patil, 23, grew up in Austin. As a child, his father worked for chip giants Qualcomm and ARM on circuit design and power architecture. He got into tech because he wanted to be like his older brother Neil, now an engineer at the biotech startup Chai Discovery. During the pandemic, when he was in high school, Patil and his brother built a program called Helping Hands, which connected those in need with healthy people who could carry out tasks like grocery shopping or picking up medicine. Garg, 24, who grew up in Dallas, discovered tech by doing hackathons in high school, while North Carolina native Li, 25, spent his formative middle and high school years in Shanghai and got into tech after seeing a demo for self-driving cars.

The three founders became friends at Stanford. As a computer science student, Patil met his future boss, OpenAI CEO Sam Altman, at a dinner he hosted at his house for Stanford engineering students. “I was really nervous. Didn't talk to him. Maybe said ‘hi, hello’ at the end,” Patil says. Then OpenAI launched ChatGPT months later, and like the rest of the world, Patil was enamored with the technology. He emailed Altman and said he’d love to work on the fledgling chatbot, and Altman invited him to apply for the company’s residency program. After joining the company at age 20, Patil referred Garg to join. Then Garg referred Li. Once there, Patil worked on post-training infrastructure and what would become Codex, while Garg worked on reasoning models and Li focused on machine learning systems.

Because they were all so young compared to OpenAI’s other researchers, they bonded even more; Patil and Garg were roommates at an apartment next to OpenAI headquarters in San Francisco, while Li, who lived nearby, would come over to hang out after work.

Then came the DeepSeek a-ha moment. Patil, Garg and Li all happened to be on a ski trip together in Lake Tahoe when the R1 model launched. Enthralled, they spent most of the day talking about it in between runs on the slopes. The saga put a spotlight on open source models more dramatically than ever before. Patil doesn’t think enterprise customers will abandon the state-of-the-art frontier models from OpenAI and Anthropic to go all-in on open source. Instead, he thinks companies will use a mix of high-end and cheaper models depending on the use case. And as some companies begin to focus more on cost effectiveness and open source alternatives become more viable options, Applied Compute thinks its moment is now.

Leigh Marie Braswell, the Kleiner Perkins partner who led the blueblood VC firm’s investment into Applied Compute, saw the company’s focus firsthand earlier this year. Braswell, an avid poker player, is well-known around Silicon Valley for holding poker nights with portfolio companies. But when she held one at Applied Compute headquarters in San Francisco, the night was cut short around 10 p.m. “They were the only company that's ever kicked me out of the office before,” she says, laughing. “They were like, ‘We have to get back to work.’”