Cocoon from Telegram: A Decentralized AI Network That Pays GPU Owners in Crypto (decrypt.co)

Telegram announced Cocoon (the Confidential Compute Open Network), a decentralized AI compute marketplace built on the TON blockchain that pays GPU owners in Toncoin for running private AI inference. Slated to launch in November with Telegram as its first major customer, Cocoon lets individuals rent out GPU capacity while keeping inputs encrypted end-to-end so neither the network nor the GPU host can read user data. Developers can bid for lower-cost, market-priced compute for models, and Telegram plans to integrate Cocoon into its Mini Apps and AI features across its billion-plus user base. Applications for GPU providers and developers are open now. Technically, Cocoon leverages TON’s sharded, multi-chain architecture (designed for millions of transactions per second) to coordinate marketplace transactions, pricing, and workload distribution. AlphaTON Capital has committed to invest in high-memory GPUs across data centers to support larger architectures (examples cited include DeepSeek and Qwen). The project positions itself as a privacy-first alternative to centralized cloud providers like AWS and Azure, promising competitive pricing and user control but also introducing uncertainty around service reliability, consistency, and operational complexity typical of distributed compute markets. If successful, Cocoon could shift AI infrastructure toward a user-driven, crypto-incentivized model that decentralizes both compute and data control.