Researchers at École Polytechnique Fédérale de Lausanne (EPFL) have developed a software-based approach intended to enable large AI models to run without relying on massive data centers. The solution, named Anyway Systems by its creators, removes the need to send data to third-party cloud providers when using AI for tasks and allows processing to occur locally on user-controlled networks.
Concept and capabilities
Anyway Systems lets users download open-source large models — the researchers cite the OpenAI-referenced GPT-120B as an example — and run them on local machines so that data never leaves the local network. According to the team, the system can be deployed to a network of local machines in about half an hour, and very large model files can be downloaded and installed on the Anyway Systems setup in a matter of minutes.
Requirements and operation
The researchers emphasize that the approach does not require expensive, specialized rack infrastructure. The setup needs no more than four machines, each paired with a conventional GPU. This decentralized arrangement splits computation across smaller units instead of concentrating everything in a single enormous data center.
Benefits: privacy, sovereignty, sustainability
A key promise of the system is enhanced privacy and data sovereignty: because data processing remains within the local network, dependence on major cloud providers is reduced. The team also points to sustainability benefits: optimized local deployments may consume fewer resources than running large centralized data centers, which the article notes are often wasteful in terms of raw materials, energy and water use.
Limits and applicability
The researchers acknowledge that Anyway Systems is not currently aimed at home users: running the system requires multiple relatively powerful machines, which ordinary households typically do not maintain. They also note, however, that computing history shows rapid optimization is possible (for example, in the capabilities of modern smartphones), and that small organizations and companies are likely already equipped with the hardware needed to run the system.
Social and economic implications
EPFL frames the idea as more than a technical optimization: it challenges the social, environmental and economic foundations of AI infrastructure. The team argues that open-source models, local execution environments and distributed computing can reduce centralization and reshape the meaning of a “supercomputer,” making massive data centers no longer the only model for heavy AI workloads.
Practical testing
Anyway Systems has moved beyond prototype stage and has been tested for months across Switzerland in companies and public-sector organizations, according to the researchers. Rachid Guerraoui, a researcher at EPFL’s School of Computer and Communication Sciences and one of the project’s originators, said people long believed large language models required enormous resources and that privacy, sovereignty and sustainability had to be sacrificed — but that smarter, more frugal approaches are possible.
Summary
EPFL’s Anyway Systems demonstrates that large AI models can be run on distributed local networks rather than centralized data centers. The approach promises quick deployment, operation on a small number of standard GPUs, and improvements in privacy, sovereignty and sustainability, while broader adoption will depend on further optimization and testing.
