River AI secures $1.1 billion to build trainable personal agents
River AI, the startup founded by xAI co-founder Igor Babuschkin, has raised $1.1 billion in a seed/Series A financing round just two months after emerging from stealth. General Catalyst and AMP PBC led the investment, with Nvidia, AMD Ventures, Y Combinator and Temasek also participating.
Babuschkin has previously held AI roles at DeepMind and OpenAI. River’s stated objective is to rebuild the AI stack around personally trainable agents rather than pursue systems designed principally to replace human workers.
A full-stack approach to personal AI
In its launch material, River said the required stack spans training, models, the product layer and hardware that allows personal AI to operate close to the user. The company envisages capable agents as a normal part of everyday life: systems that know their users and act on their behalf.
The thesis arrives as enterprises seek more control over their AI choices, including the use of open-weight models alongside other systems. That direction also aligns with investment in enterprise agents that reduce dependence on frontier-model providers, such as enterprise agents without frontier-lab dependence, where control over the model stack is central to the proposition.
River has not yet demonstrated how its technology will differ from other efforts around locally run agents or AI-capable hardware. Nvidia, meanwhile, has partnered with PC makers including Dell, Microsoft and HP on hardware for AI workloads.
API targets post-training workflows
River’s initial product is an API billed per one million tokens, with prices varying by the open model selected. Developers can use the service for reinforcement learning and low-rank adaptation, or LoRA, fine-tuning.
The company positions the product as an alternative to prompt engineering. Its product literature argues that prompting only steers a model a customer does not own or improve, while River is intended to let users train open models and serve the resulting systems through an endpoint.
River says an enterprise can complete a complex reinforcement-learning run in 15 to 20 minutes without an infrastructure team. It also claims costs can be two to four times lower than closed-source alternatives, a claim that will be important for prospective customers to test against their own workloads and model choices.
What the funding changes
The size of the round gives a very young company substantial resources to pursue a broad technical agenda, from post-training infrastructure to a longer-term vision for personal agents. AMP PBC, an AI-focused investment firm founded by former Andreessen Horowitz general partner Anjney Midha, joins a group of strategic and financial backers that includes major chip industry participants.
For businesses, the practical implication is to evaluate post-training platforms not only on model quality, but also on the time, infrastructure effort and governance required to adapt open models for specific internal tasks.