Exclusive | Deep Cogito Aims to Put Companies in Control of Their Own AI

AI research lab Deep Cogito said Wednesday that it raised $43 million to fund the development of specialized artificial-intelligence models that companies can own, helping them control proprietary data and optimize performance while keeping costs in check.

The Series A round was led by TQ Ventures, with participation from other investment firms Benchmark, Nexus Venture Partners, Atreides Management and South Park Commons, plus cybersecurity company Zscaler. This brings Deep Cogito’s total funding to more than $56 million.

The San Francisco-based startup focuses on lowering the cost of intelligence and improving performance with so-called open-weight models, a hot area of growth as companies look to improve AI economics. China dominates the market for such models.

Deep Cogito was founded in 2024 by two former Google employees who worked on its AI Search products, Drishan Arora, who is chief executive, and Dhruv Malrana, chief product officer. Their research is focused on the post-training phase, after a model has been broadly trained on enormous amounts of more general information.

Deep Cogito said its post-training engine powers its Cogito family of frontier open-weight models as well as specialized models trained on an enterprise’s proprietary data and outcomes. Unlike proprietary models, open-weight models publicly share underlying numerical parameters, or “weights,” and are generally cheaper to run.

Deep Cogito can play an important role in building an open-weight ecosystem in the U.S., according to Schuster Tanger, co-founding partner at TQ Ventures.

“We think of them becoming one of the defining companies building open-weight models and specialized enterprise intelligence. We think there’s an aperture for a highly proficient American solution,” Tanger told me.

Arora believes that companies will become increasingly open to the idea of building their own models, as long as they have external help. Currently, only a limited number of companies have the necessary in-house skills and resources.

“The idea that you are going to use open models…for your own proprietary intelligence, and optimize it for your particular content has resonated with a lot of folks. I believe there’s an inflection point here. This is going to happen more and more,” Arora, a former senior software engineer at Google, told me. He held a number of roles at the tech giant, most recently modeling Google’s generative search.

“Cost is one driver, but the equally important one is performance—specialized models trained on product-specific data often outperform general frontier models on that product,” he says.

Deep Cogito applies several techniques, including large-scale reinforcement learning, in which a model learns through a combination of penalties and rewards. It also uses recursive self-improvement, in which a model teaches itself how to improve, producing a compounding series of gains in intelligence.

One Deep Cogito research initiative repeatedly allows a model to use additional computation to produce answers beyond what it could generate directly, “then distills those improvements back into the model’s weights.”

The long-term goal, according to Deep Cogito, is to build models “that progressively improve their own capabilities and ultimately move beyond the limits of human-generated training data.”

I asked Arora how Deep Cogito’s approach differs from other techniques such as retrieval-augmented generation, which retrieves relevant proprietary data and business context and provides it to an AI model at inference time.

RAG is useful for retrieving facts, but lacks the capacity for intuitive expertise, according to Arora. Not everything can be written down as a set of data, skills, instructions or context. For example, if you wanted to use AI to share your general job experience with a new intern, today’s AI would be of limited use. This is where the model learning new capabilities can potentially be of use.

“As the problem gets a little harder—cybersecurity, coding assistants are a great example—it’s hard to write down exactly the kind of code structure you want at all times, but these are the kind of skills that the model can learn over time from your data,” Arora says.

Zscaler, a strategic investor, is testing Deep Cogito with the goal of deploying it on its own platform and giving its users additional tools that they can use to secure AI, according to Dhawal Sharma, executive vice president of AI security and strategic initiatives.

“You hear it all the time. Everyone is worried about the cost efficiency and model economics. Closed-weight models are very efficient, but the cost economics are very expensive, too,” Sharma said.

He said Deep Cogito has the potential to lower costs and help Zscaler build models to its own specific needs, independent of a third party’s usage rules.

It is too early to know how much success Deep Cogito or any one company will have in AI over the next few years. But collectively, they are accelerating the pace of development in AI, and that is an important signal in and of itself.

“One of the biggest common topics right now is this idea of recursive self-improvement,” Aaron Levie, CEO of cloud content-management company Box, told me. “I think there’s a very real chance that it will be the most meaningful part of the acceleration that we see…All of the evidence from the labs and the researchers in the labs suggests that we have reached a kind of escape velocity or we’re getting close to an escape velocity.”

That opens up a big opportunity for enterprises looking for ways to maximize their return on investment in AI. It creates a significant challenge, too. The effort only begins with creating models. They also will need to hire and retain the right business minds, especially leaders who can structure the company around a rapidly moving target.