Igor Babuschkin (@ibab), the former xAI co-founder who helped build AlphaStar at Google DeepMind, has raised $1.1 billion for River AI, two months after publicly introducing River.
The financing, announced on August 11th and reported by Reuters, combines River's seed and Series A rounds. General Catalyst and AMP PBC led the funding, with strategic investments from Nvidia and AMD Ventures. Y Combinator and Temasek also participated. River did not disclose its valuation or divide the $1.1 billion between the two rounds.
That structure makes the financing difficult to compare with a conventional Series A. It gives Babuschkin a large pool of capital before River has published customer names, revenue or retention data, and before its first product has moved beyond preview access. Investors are underwriting his record of building large AI systems and his argument that companies will eventually own customized models instead of relying entirely on general-purpose systems rented from closed labs.
Babuschkin has worked near the center of that closed-lab model. At Google DeepMind, he was part of the team behind AlphaStar, the reinforcement-learning system that reached grandmaster-level performance in StarCraft II. He later worked on large-scale model training at OpenAI before co-founding xAI (@xai) in 2023.
River is his attempt to take techniques developed inside well-capitalized research labs and package them for enterprises that lack dedicated training infrastructure.
River starts with the training layer
The first product, the River API, manages LoRA fine-tuning, reinforcement learning, inference and deployment for open-weight models. Customers supply training data and reward functions while River handles weight transfers, elastic compute and consistency between the systems used to generate samples and update the model.
River says developers can use the API with models ranging from roughly 35 billion to 1 trillion parameters, then retain the resulting checkpoints. The service is intended to let developers move a trained model into production without rebuilding the application around a new interface.
The distinction River is selling is ownership. Prompting a general model can alter its output for a session, while training changes the weights and produces an asset that a customer can keep, version and tune again. That pitch is aimed at organizations with proprietary workflows or data that produce poor results from a generic model but do not justify assembling an internal reinforcement-learning infrastructure group.
River claims a complex reinforcement-learning run can finish in 15 to 20 minutes without an infrastructure team. River also says the service costs 2x to 4x less than closed alternatives. Neither claim is supported by an independent benchmark, and the relevant cost depends heavily on the base model, context length, training tokens and inference volume.
The preview pricing makes River's commercial approach clearer. River charges for training and inference tokens instead of exposing customers directly to GPU-hour pricing. River's preview pricing ranges from approximately $1 to more than $15 per million training tokens, depending on the model and configuration.
Token billing can make experimental training easier to budget, though River still carries the underlying risk of securing and efficiently using expensive accelerator capacity. The participation of both Nvidia and AMD ties River to the two chip suppliers with the clearest incentive to expand demand for model customization. Each training run and production deployment creates additional accelerator usage, regardless of whether customers ever see a GPU-hour line item.
Babuschkin is funding a larger ownership bet
Babuschkin publicly introduced River on June 10th with a broader thesis than enterprise fine-tuning. He wrote that AI agents should learn from their users, understand their preferences and remain under their control. The API is the first layer of a plan that River says will eventually span continual learning, personalized products and hardware capable of keeping personal AI close to its owner.
"AI should be open, freely available, and affordable," Babuschkin said in River's funding announcement. He argued that intelligence should work for the person using it rather than remain controlled by the lab that trained the original model.
That ambition explains why River raised well ahead of demonstrated commercial scale. Building a managed training service already requires scarce researchers, distributed systems engineers and compute commitments. Expanding from infrastructure into consumer software and hardware would put River into several capital-intensive businesses at once.
The immediate product is narrower and easier to evaluate. River must show that enterprises want to train and own model checkpoints, that its reinforcement-learning workflow is materially easier than existing services, and that token-based pricing can support healthy margins after compute costs.
A heavily financed market
River is entering a category where established AI infrastructure providers are also raising large rounds around the demand for specialized models.
Together AI announced an $800 million Series C on July 1st for a platform spanning open-model training, customization, inference and compute. Fireworks AI announced a $1.505 billion Series D on July 15th, saying its customers increasingly use models specialized on proprietary data. Baseten raised a $300 million Series E on February 5th to expand training and production inference infrastructure.
Those financings validate the market River is pursuing while raising the performance bar. River is competing with providers that already operate substantial training, inference and deployment platforms. Babuschkin's response is to start with reinforcement learning, fast training cycles and customer-owned checkpoints, then build upward toward personal AI.
The $1.1 billion gives River the resources to secure compute, hire across research and engineering and subsidize a rapid product build. It also places a large financing benchmark around a preview-stage service, even though River has not disclosed its valuation.
Babuschkin has spent his career scaling reinforcement learning inside frontier labs. River's financing rests on the bet that he can turn that expertise into infrastructure that ordinary engineering organizations can use, then carry the ownership model beyond enterprise APIs. The capital is in place. River now has to prove that customers want to possess and continuously train their intelligence badly enough to move away from the convenience of closed, general-purpose models.