Telegram launches Cocoon, a decentralized confidential AI network on TON Blockchain

Telegram has launched a decentralized confidential AI compute network on the TON blockchain called Cocoon. The project is presented as a more private and cost efficient alternative to centralized AI compute services from providers like Microsoft and Amazon, which, according to Telegram’s founder Pavel Durov, act as costly intermediaries that reduce privacy, and Cocoon is introduced as a way to avoid those limitations by letting applications run AI models without exposing data to external infrastructure while also allowing GPU owners to rent out unused hardware and earn toncoin.

Cocoon uses a three part model where clients submit paid requests, proxies select suitable workers, and workers execute the tasks on their GPUs. Payments and settlements happen on the TON blockchain, and developers say that requests and responses remain confidential to the client. The network is positioned as an alternative compute layer centered on private processing.

Telegram expects to bring more GPU supply and developer demand in the coming weeks, though current scale has not been shared. Some industry voices have questioned Cocoon’s dependence on a narrow set of Intel processors, which excludes AMD, ARM, and other architectures. Meanwhile, Durov emphasized that Cocoon is open to any developer or GPU provider interested in private AI compute, not only Telegram users, leaving open the question of whether it will eventually allow running users’ own local AI models on the network.