These startups are building advanced AI models without data centers

Researchers trained New kind Major Language Models (LLM) use GPU It is scattered all over the world, providing private and public data. This is a move that suggests a dominant construction method artificial intelligence It can be confusing.

Flower Ai and oldtwo startups pursuing an unconventional approach to building AI, worked together to create a new model called Collective-1.

Flowers created a technology that can spread training to hundreds of computers connected over the internet. The company’s technology is already being used by some companies to train AI models without the need to calculate resources or data. VANA provided a source of data that includes private messages from X, Reddit and Telegram.

Collective-1 is small by modern standards, with 7 billion parameters (worthy to give the model its capabilities). chatgpt, Claudeand Gemini.

Nic Lane, a computer scientist at Cambridge University and co-founder of Flower AI, says the decentralized approach promises to expand well beyond the size of Collective-1. Lane is in the process of training a model with 30 billion parameters using traditional data, and plans to train another model with 100 billion parameters this year to fit the size offered by industry leaders. “It can really change the way everyone thinks about AI, so we’re following this pretty hard,” Lane says. He says the startups are incorporating images and audio into their training to create multimodal models.

Building distributed models can also destabilize the power dynamics that have shaped the AI ​​industry.

AI companies are now building models by combining vast amounts of training data with a large amount of computing concentrated within data centers packed with advanced GPUs networked using ultra-high speed fiber optic cables. It also relies heavily on datasets created by reduced-reducible reductions, including websites and books.

This approach means that only the wealthiest companies and countries with access to a large number of most powerful chips can viablely develop the most powerful and valuable models. Even open source models like Calling Meta and Deepseek’s R1built by companies with access to large data centers. A distributed approach could enable small businesses and universities to build sophisticated AI by pooling different resources together. Alternatively, countries that lack traditional infrastructure could network several data centers to build stronger models.

Lane believes it will turn increasingly into new ways that enable the AI ​​industry to remove training from individual data centers. A distributed approach allows for “computations to be much more elegant than data center models,” he says.

Helentner, an AI governance expert at the Security and Emerging Technology Center, says Flower AI’s approach is “interesting and potentially highly relevant” to AI competition and governance. “We’ll probably continue to struggle to keep up with the frontier, but it could be an interesting first follower approach,” Toner says.

Divide and conquer

Distributed AI training involves rethinking how the calculations used to build powerful AI systems are split. Creating an LLM involves supplying a huge amount of text to the model that adjusts the parameters to generate useful responses to the prompt. Within the data center, the training process is split, running parts on different GPUs and can be periodically integrated into a single master model.

A new approach allows for hardware to perform tasks that are typically done within large data centers. The hardware can be a few miles away and may be connected via a relatively slow or variable internet connection.