River AI raises $1.1 billion to sell custom agent training as an API
The xAI co-founder's pitch inverts the GaaS default: instead of renting a vendor's agent, enterprises train agents they own, on infrastructure that makes reinforcement learning a 20 minute job.
River AI, the startup founded by xAI co-founder Igor Babuschkin, announced Monday it has raised $1.1 billion in a round led by General Catalyst and AMP, with strategic investment from Nvidia and AMD Ventures and additional backing from Y Combinator and Temasek. The company did not disclose a valuation.
The number would be striking for any AI startup. For one that emerged from stealth in June, it is close to unprecedented. TechCrunch notes the company is roughly two months old. Babuschkin, who worked at DeepMind and OpenAI before co-founding xAI, left Elon Musk's lab to build what he describes as AI owned by the people who use it rather than the labs that train it.
What River actually sells
Underneath the mission language is a concrete product: an API for training. River lets developers run reinforcement learning and LoRA fine-tuning against open-weight models, and the company claims an enterprise can complete a complex RL run in 15 to 20 minutes without an infrastructure team. River also claims its service is two to four times more cost-effective than closed-source alternatives, a figure that comes from the company and has not been independently verified. Billing is token-based per million tokens, with rates varying by the underlying open model, according to TechCrunch. Reuters' coverage via Yahoo Finance quotes Babuschkin directly: AI should be open, freely available, and affordable, and it should feel like it is working for the person using it.
The investor list tells its own story. Nvidia and AMD both took strategic positions in a company whose entire premise is that customers will run large numbers of training jobs on open models. Every River customer is a compute customer, whoever ends up supplying the silicon.
A direct shot at the GaaS default
The dominant commercial model in agentic AI today is rental. A vendor trains and operates the agent, the enterprise pays per seat, per task, or per resolution, and the accumulated learning belongs to the vendor. River's pitch inverts that: the customer's data trains a customer-owned agent, and the vendor sells the training run, not the worker. Babuschkin told TechCrunch that capable agents will be a normal part of everyday life, will know you well, and will be yours, not someone else's.
That framing lands in a market already moving toward open and decentralized training. It rhymes with the RL infrastructure thesis behind Prime Intellect's $130 million Series A and the open-model economics that pushed Nous Research to a $1.5 billion valuation. The open question is whether enterprises actually want to own agents, with the maintenance, evaluation, and liability that ownership implies, or whether ownership talk is a wedge and the durable business is metered training compute. Either way, $1.1 billion buys River a long time to find out, and it puts a well-capitalized bet on the table that the next phase of the agent economy is sold by the training run rather than by the seat.