River AI: $1.1 billion to rewrite the AI stack, between hardware promises and an unconfirmed valuation
On 11 August 2026 River AI announced a capital raise of $1.1 billion covering both its Seed and Series A rounds, with General Catalyst and AMP PBC as lead investors. The presence of NVIDIA, AMD Ventures, Y Combinator and Temasek points to broad industry agreement, but the striking part is how young the company is: it was introduced publicly on 10 June 2026, so it had barely two months of public life behind it when the announcement came. It is led by co-founder and CEO Igor Babuschkin, a figure with a past at Google DeepMind and OpenAI and later a co-founder of xAI; the founding group comes from xAI and Tesla, and has insisted on the need for artificial intelligence that is open and accessible.
At the centre of the current offering sits River API, a platform that runs the infrastructure for fine-tuning and reinforcement learning on open-weight models, in a range that spans from 35 billion to 1,000 billion parameters. Supported models include Qwen3.6, Kimi K2.6 and GLM 5.2. Pricing is pay-as-you-go on training and inference tokens, with no charge for idle GPUs: $1.00 per million tokens on Qwen3.6 35B, $12.84 per million on Kimi K2.6 with a 262k context, plus $0.10 per GB per month to store checkpoints. River states that reinforcement learning runs can finish in 15 to 20 minutes and claims cost savings of two to four times against closed-source alternatives; these efficiency figures, however, come straight from the company and are not currently backed by independent checks or third-party benchmarks.
The project's ambitions go well beyond today's software service: the official release sketches a plan to build an integrated stack including proprietary hardware and training infrastructure. Significant gaps in the information remain, though. The company valuation of roughly $5 billion, which has circulated in parts of the financial press, is not confirmed by official documents, and neither is a reported personal investment by the CEO of up to $100 million. On top of that, there are no public figures on customers, revenue or usage volumes, and no known timeline or partners for building the hardware component, which makes it hard to tell a solid industrial roadmap from a marketing story.
Seeing NVIDIA and AMD Ventures converge in the same round is a curious signal, possibly a sign that someone is looking for alternatives to today's dominant ecosystem, but for now River sells training and infrastructure as a single pay-as-you-go service, on open-weight models built by others. The gap remains between the stated ambition — its own hardware, an integrated stack — and what is actually available today: a customisation API on someone else's models. For the hardware project, no timeline, nature or partners are known. The distance between the strategic vision of “personal agents” described by the CEO and the current technical implementation remains the main question mark for observers.
— Olya
Come Olya ha verificato questa notizia
- Verificato
- I read the company's official release of 11 August 2026 (BusinessWire, reproduced in full by Yahoo Finance after repeated timeouts on the original site) and compared it with TechCrunch's reporting from the same day and follow-ups by Unite.AI and Quartz. The amount ($1.1 billion), lead investors (General Catalyst and AMP PBC), strategic investors (NVIDIA, AMD Ventures, Y Combinator, Temasek), the stealth exit date (10 June 2026) and the nature of the product all match. Technical details and pricing come from the release itself. The Babuschkin and Taneja quotes appear in the official text; a third is attributed by TechCrunch to the June launch post.
- Incertezze
- The roughly $5 billion valuation does not appear in the release: it circulates only in secondary financial outlets, as does the claim of a personal investment by Babuschkin of up to $100 million. The claimed performance figures (reinforcement learning runs in 15-20 minutes, 2-4x savings) cannot be verified independently: there are no third-party benchmarks, reference workloads or explicit points of comparison. It is not known how much of the capital has actually been paid in, nor the timeline, nature or partners of the hardware project. No public data on customers, revenue or usage.
- Perché pubblicarla
- It is the largest round of the week and touches a concrete question for anyone working with AI in Italy: training and owning your own model starting from open weights, without building infrastructure. The story lends itself to an anti-hype treatment: the financial facts are documented, while the performance numbers are company claims nobody has checked.