On 29 October 2025 at Blockchain Life in Dubai, Pavel Durov introduced Cocoon (Confidential Compute Open Network) built on TON. Cocoon is a decentralized network for private AI computations: requests are encrypted and distributed across GPU nodes, and GPU owners receive rewards in TON. Telegram was named the first major client, which should provide initial demand and a concrete real-world use case.
Cocoon focuses on model inference. Developers connect their services and submit tasks, the network selects available GPU executors, encrypts the data, and returns the result. For compute providers, this is a way to monetize idle GPUs without deploying their own infrastructure. For companies, it enables AI features without renting data center capacity and without the risks tied to the policies of a single centralized provider.
Key points from the talk:
- For the TON ecosystem: a native AI compute use case on top of the blockchain, closely tied to Telegram products and its audience.
- For the market: a stronger trend toward privacy-preserving computation and reduced dependency on centralized clouds.
- For users: a chance to access AI tools where data is processed confidentially and access rules are transparent.
The launch is planned for November 2025. The team has opened onboarding both for developers (workload and model details) and for GPU owners (hardware parameters and availability).
If the launch delivers on the promised privacy and robustness, Cocoon may become a notable step toward decentralized AI in tandem with TON and Telegram, potentially boosting demand for TON and broadening the set of use cases across the ecosystem.
