River AI raises $1.1B to bring personalized, self-owned AI to companies and consumers

River AI raises $1.1B

Igor Babuschkin has raised $1.1 billion for River AI, the company he co-founded with Brannin McBee. Andreessen Horowitz (through partner Anjney Midha) is a named investor.

Babuschkin's path to the raise runs through almost every major AI institution of the last decade. He joined DeepMind roughly ten years ago, worked on WaveNet and a competitive StarCraft agent, then moved to OpenAI where he focused on reasoning and coding. He co-founded xAI with Elon Musk before departing to start River.

The ownership thesis

River's central argument is that the current model, where a few large corporations train powerful models and everyone else pays by the token, is not where AI ends up long-term. The alternative Babuschkin is building toward is one where companies and individuals fine-tune and own open-weight models trained on their own data, with inference eventually running locally, in a home or office rather than a centralized data center.

The River API is live now at river.ai/api. Companies can log in, select from a range of open-weight models, and run fine-tuning jobs charged by compute used rather than a flat subscription. Babuschkin says some customers end up training small, fast, specialized models that outperform much larger frontier models on their specific task, simply because their proprietary data is so well-matched to the eval.

“We just managed to raise $1,100,000,000 for River AI. We wanna figure out how we can distribute AI to everybody in a way where you own it, you're able to shape your own AI systems. Maybe you have the inference running in your home or in your office. We're helping companies build AI with the River API — you can log in and you can start training models based on open weights.”

Personalized agents

Beyond the API, River is building personal AI agents that grow more personalized over time. Babuschkin describes these as the next evolution after coding agents, with demand for tokens expected to increase significantly as they become widespread.

He acknowledges the hardest research problem in this area: today's best agents are trained in synthetic environments and never interact with the real world during training. River is working on online learning, updating model weights in real time as agents solve actual tasks for real users, but Babuschkin is candid that nobody has solved this in general yet.

Compute as the binding constraint

The $1.1 billion is partly a compute story. GPU prices are rising and capacity is constrained, and Babuschkin describes a new dynamic emerging where investors pool GPU access across portfolio companies, with allocations shifting based on which company has the heaviest demand in a given month. River is scaling GPU capacity for both research and to power the API as customer demand grows.

On custom silicon, Babuschkin sees room for inference-specific chips optimized for personal AI agents, potentially with transformer architectures baked in. The current ceiling is roughly 8 billion parameters for weight-baked chips like those Talos demonstrated before its AMD acquisition, while the best frontier models run into the trillions. Closing that gap is, in his view, a prerequisite for efficient local inference at scale.

The neo-lab question

Babuschkin's position is that the field is just getting started. The arrival of capable coding agents and APIs has opened space for entirely new business models and research directions, and he argues that the next breakthrough is as likely to come from someone outside the established AI institutions as from a ten-year veteran.