A startup that has existed for just two months has raised $1.1 billion in funding, led by General Catalyst and AMP PBC, with participation from Nvidia, AMD Ventures, Y Combinator, and sovereign wealth fund Temasek. Behind River AI is Igor Babuschkin, a co-founder of xAI and formerly a researcher at DeepMind and OpenAI. His thesis: AI agents should belong to people and be trainable by them, rather than rented as intelligence from a handful of labs. The actual product River AI sells today, though, is still a fair distance from that vision.
Key takeaways
- River AI was founded in June 2026 and had already raised a combined $1.1 billion across seed and Series A rounds by August 2026.
- Founder Igor Babuschkin is reportedly targeting a valuation of up to $5 billion, according to Forbes, and is investing up to $100 million of his own capital.
- The long-term vision is personal, owner-controlled AI agents, which Babuschkin compares to “guardian angels”: present, loyal, and truly owned by the user.
- The current business is far more modest: a cloud API for reinforcement learning and LoRA fine-tuning on open models like Qwen3.6, Kimi K2.6, and GLM 5.2.
- Critics point out that the personal-AI vision remains unproven long-term, and that River’s claimed speed and cost advantages have not yet been independently verified.
The big idea: AI that actually belongs to you
Babuschkin’s core thesis is that the entire AI stack, from training through models to hardware, needs to be rebuilt around personal agents rather than around intelligence rented from OpenAI, Google, or Anthropic. In his own words, these assistants should function less like helpers you summon and more like “guardian angels”: constantly present, on the user’s side, and actually owned by them rather than merely rented compute time from someone else’s servers. That vision deliberately bets on open weights over closed models, a framing General Catalyst CEO Hemant Taneja makes explicitly geopolitical: American leadership in AI, he argues, urgently requires leadership in open-weight models too, a position Nvidia recently backed with an open letter of its own, one that OpenAI and Anthropic have so far declined to sign. For readers following the broader debate over practical alternatives to the big cloud providers, this connects to the case for small, open models as a strategic answer to the big three, only here with the added claim that users should ultimately own the models they train.
The real business: a training API for enterprises
What River AI actually sells today is considerably less visionary and, for now, aimed at enterprise customers rather than consumers: a cloud platform where open models like Qwen3.6, Kimi K2.6, or GLM 5.2 can be fine-tuned to a company’s own data through reinforcement learning and LoRA, billed by actual tokens used rather than by reserved, often idle GPU capacity. Concretely, a run on the 35-billion-parameter Qwen3.6 model costs around $1 per million tokens according to River, while Kimi K2.6 with its 262,000-token context window runs about $12.84 per million tokens. A full reinforcement-learning run on a math dataset is supposed to stay under $1,000 and finish in 15 to 20 minutes, with no dedicated infrastructure team required. River claims a cost advantage of two to four times over closed-source alternatives, though those figures haven’t been independently verified yet.
Why investors are getting in anyway
The fact that investors like Nvidia and AMD Ventures are simultaneously chipmakers and backers raises the obvious question of whether they’re primarily interested in River’s personal-agent thesis or simply in another large buyer of training hardware. For Babuschkin, what matters most is the team: a core group of engineers with experience at xAI and Tesla is meant to solve the infrastructure problems that have historically made large-scale reinforcement learning slow and expensive, such as transferring model weights between training and inference systems or maintaining sampling consistency. Whether this actually grows into a mass market for personal, owner-controlled AI hardware, as Babuschkin envisions, remains an open question, and the company itself acknowledges that the longer-term thesis is still unproven.
Outlook
River AI is currently selling two different things at once: a solid, if still unproven, training API for enterprises looking to fine-tune open models faster and more cheaply, and a much larger narrative about AI agents that will one day belong to people instead of being rented from them. The $1.1 billion in funding mainly buys time and compute to close the gap between that present and that vision. Whether it succeeds will depend less on the size of the war chest than on whether River AI can actually compete for training budgets and talent against the established frontier labs.