River AI raises $1.1B for personal AI stack

River AI closed a combined $1.1 billion Series Seed and Series A round ^[1], reportedly valuing the roughly 20-person, Palo Alto-based company at around $5 billion ^[2]. What stands out isn't just the speed of the raise relative to the company's age - it's the investor list. General Catalyst and AMP PBC co-led, but Nvidia and AMD Ventures, chip rivals that rarely sit on the same cap table, both wrote checks ^[3]. Both chipmakers have an obvious stake in the outcome: River's stated roadmap moves beyond software into custom AI hardware for a fully user-owned personal AI stack ^[4], meaning today's investors are also positioning for tomorrow's silicon customer.

River's first product, the River API, doesn't sell a model - it sells the ability to fine-tune one. Customers apply LoRA, a technique that trains small adapter layers instead of retraining a full network, combined with reinforcement learning, on open-weight models from 35 billion to 1 trillion parameters, including Qwen3.6, Kimi K2.6, and GLM 5.2, billed per token rather than by GPU-hour ^[5]. River claims a full RL training run can be stood up in 15 to 20 minutes without an in-house infrastructure team, at 2-4x the cost savings versus closed-source alternatives ^[1]. That pricing logic reflects Babuschkin's broader thesis, stated at the raise: "The way AI is built today is not how it will be built in the future. AI should be open, freely available, and affordable." ^[6]General Catalyst's Marc Bhargava frames the gap River is filling in similar terms: "River closes this gap, helping any company build models on their own data, tailored to how they actually work." ^[6]

Babuschkin's path runs through Google DeepMind's generative modeling and reinforcement learning work, large-scale training at OpenAI, and a stint as an xAI co-founder before he left in August 2025 ^[3]to start River within months. His public framing of the new company leans philosophical as much as commercial: he has described the end state as personal AI hardware that "is yours, not rented, and you have real control over it" ^[3]. General Catalyst CEO Hemant Taneja ties that framing to a geopolitical argument, calling open-weight leadership a strategic priority: "American leadership in AI urgently requires leadership in open-weight models, while maintaining a lead in closed frontier models." ^[6]

The commercial product is narrow and transparently priced: training runs cost $1.00 to $15.41 per million tokens depending on model size, with checkpoint storage at $0.10 per GB per month ^[5]. That's the entirety of River's shipping business today, run by a team of roughly 20 people ^[7]. Personalization and hardware, the other two pieces of the stack Babuschkin has described publicly, remain unbuilt: the plan is for models to adapt to individual users through ongoing feedback rather than serving one 'average user,' and eventually to run on hardware a person actually owns instead of in a data center ^[4].