Nous Research raises $50M for decentralized AI training led by Paradigm – News

Nous Research, a decentralized artificial intelligence startup, today said it has raised $50 million in an early-stage round almost entirely funded by the crypto-focused investment firm Paradigm.

According to a report from Fortune, the Series A round follows seed rounds totaling $20 million from investors including Distributed Global, North Island Ventures, and Delphi Digital. According to the report, the funding brings the crypto token valuation of the company to $1 billion.

Novus uses the Solana blockchain network to enable distributed training for AI models by allowing users to commit idle compute from both consumer- and industrial-grade graphics processing units.

The company seeks to disrupt the current paradigm of centralized AI development of giants such as DeepSeek and OpenAI by allowing numerous participants to join the network and commit resources.

Although Nous did not start with the idea of adopting blockchain technology, it eventually sought it out because it can be incorporated to facilitate decentralized model training. It can also be used to reward participants for their unused GPU compute contributed to the system with blockchain-based cryptocurrency.

“We think of the incentive mechanism behind crypto to push people to actually utilize their idle compute less as a donation but more as a transaction,” Nous cofounder Karan Malhotra told Fortune.

Additionally, blockchain technology includes a mechanism for tracking transactions that allows the company to build decentralized training processes and produce complete, transparent historical audit trails of all data committed to a system.

Nous hopes this capability will make it possible to detect and avoid data poisoning, also known as AI poisoning, a type of cyberattack where malicious actors manipulate or corrupt training data used to build machine learning models. The goal is to influence the outcome of the model’s behavior and outputs leading to biased or harmful results by injecting false or misleading data, deleting data or altering labels.

Malhotra added that there is a lot of skepticism currently about blockchain technology and cryptocurrency, especially regarding scams and hacking, noting that there is “a lot of room for grift.” However, he said a lot of the research team is familiar with the crypto business and initially were not in favor of adding the blockchain element until a recent research breakthrough for decentralized training.

“We don’t want to get kind of bogged down by the traditional view of how crypto operates when we’re a very serious research lab and an academic lab,” Malhotra said.

Nous recently released Psyche, named for the goddess of the soul from Greek mythology, a network on the Solana blockchain that allows for the fault-tolerant globally distributed training of AI models. The company said it ran a successful test run of a 15 billion parameter model in December 2024 by scaling the network through 11,000 steps.

At the center of Psyche is a component that optimizes training called Nous Distro which lets each node train independently instead of coordinating at every step. The company said this greatly decreases the bandwidth needed to achieve comparable results to centralized training.

Image: News/Microsoft Designer