_Decentralized AI Networks)
Recently, the idea that centralized AI computing could be a threat to humanity has gained traction. Right now, AI computing is mostly in the hands of a few companies, like OpenAI, Microsoft, Google, and Meta. This concentration of power limits global access and may lead to unfair pricing, exclusion, or geopolitical issues. For instance, the US currently controls over 75% of the world’s GPU compute capacity.
OpenAI (ChatGPT) increased its revenue from $1 billion to $15 billion in just two years. This showcases the huge value of AI products. Companies are now fiercely competing for influence and market share, realizing that those with computing power will hold a significant advantage in the future.
But is there a solution or an opportunity to change the game? Some believe there is.
Liberman Brothers and Gonka
One of the most ambitious efforts to decentralize AI comes from the Liberman brothers, seasoned entrepreneurs from Silicon Valley. Their project aims to create an open, decentralized AI compute network where anyone with powerful GPUs can contribute their resources to a shared global infrastructure.
Inspired by Bitcoin’s proof-of-work model, their protocol has an important twist. Instead of wasting energy on meaningless hashing, 99% of the network's computation focuses on useful AI tasks, like running LLM inference or training new models. Only 1% goes to network security. In exchange for their contributions, GPU providers receive the project's native token, which incentivizes them to share their hardware.
Their vision opposes monopolism. They want to create a "viable open-source alternative, so that not only a handful of corporations or states control the 'brains' of AI." To avoid regulatory issues, they developed the protocol without venture funding, made the code open source, and released it into the public domain. This ensures that it remains community-owned from the start. The network expanded organically from 80 to 450 server-grade GPU nodes in its first month and a half, aiming for a long-term trajectory to compete with centralized giants.
Telegram Founder Pavel Durov and Cocoon
In Dubai, Telegram founder Pavel Durov announced Cocoon (Confidential Compute Open Network), a project based on the TON blockchain. While it shares the idea of a decentralized GPU network, Cocoon primarily focuses on user privacy and confidentiality.
In the Cocoon model, hardware owners lend their GPU power to the network for AI tasks like summarizing documents or drafting messages in Telegram apps. They are rewarded with TON tokens. The main benefit is that user data never goes to centralized AI providers like OpenAI or Google, reducing risks of data misuse, profiling, or manipulation.
Telegram will be the first major user and promoter of Cocoon, planning to invest heavily in its promotion across its vast global audience. By utilizing the existing TON ecosystem, which already supports Telegram's digital collectibles and payments, Cocoon hopes to create a transparent free market for AI computing, where prices vary based on supply and demand. The launch is scheduled for November.
Are others exploring similar ideas?
The decentralized AI movement goes beyond individual projects. Various networks are building complementary solutions throughout the AI landscape.
Bittensor (TAO)
Bittensor (TAO) establishes a unique marketplace for machine intelligence. Unlike networks focused on computing power, Bittensor uses specialized subnets that compete to produce valuable AI outputs ranging from text generation to data analysis. The network rewards models based on the quality of their intelligence, fostering a decentralized system that prioritizes useful AI outputs over sheer computational power.
io.net and Akash Network
Compute Providers tackle the essential demand for processing power through different approaches. io.net specializes in bringing together distributed GPUs into a unified cluster optimized for parallel AI training, providing scalable alternatives to traditional cloud services. Meanwhile, Akash Network offers a general-purpose cloud marketplace, allowing users to deploy any application container, often at significantly lower prices than conventional services, which is practical for various AI tasks.
Render Network Render Network illustrates how established decentralized infrastructures can continue to grow, now including AI applications. Originally created for graphics rendering, its global network of GPU nodes now supports additional capacity for AI inference workloads, especially in media creation workflows. This shift in Render Network shows how decentralized solutions can adapt to meet the demand for AI applications.
What’s next?
The race to decentralize AI computing is already in progress. Projects like those from the Liberman brothers and Telegram’s Cocoon represent two strong fronts in the same struggle: one challenging the economic and infrastructure monopoly of Big Tech, and the other safeguarding user privacy from the risks posed by centralized AI models. While the Libermans are relying on a pure, Bitcoin-like proof-of-work strategy for AI to establish a global supercomputer, Durov is leveraging Telegram's extensive network to bring privacy-first AI to a wider audience. They are joined by a growing number of DePIN projects, all affirming the belief that the future of computation doesn’t need to be centralized. The game is changing, and new players are entering the field.