Telegram introduces privacy-focused AI summaries backed by Cocoon

Telegram has rolled out its first update of 2026, introducing AI-powered content summaries for channel posts and Instant View pages.

The new feature, aimed at enhancing user productivity while browsing long-form content, summarizes posts automatically using open-source models running on Cocoon, a decentralized AI platform designed to shield user data from exposure during inference. Telegram emphasizes that these summaries are computed without ever exposing the user's input or reading history to external entities, including the servers providing the AI compute.

Cocoon is a relatively new platform built on the TON (The Open Network) blockchain. It connects GPU owners who offer computational power with AI applications that require inference capabilities, all under a robust privacy and attestation framework. Telegram's integration with Cocoon marks one of the first large-scale deployments of the system.

Cocoon is composed of three main components: clients (used by applications like Telegram to send inference requests), proxies (which route requests to available compute workers), and workers (GPU-powered virtual machines that execute the model). Both proxies and workers operate inside Trusted Execution Environments (TEEs), specifically Intel TDX-protected virtual machines. This setup ensures that all data, from user prompts to AI-generated responses, remains encrypted and inaccessible to the hardware providers themselves.

Each inference request follows a multi-step, encrypted pipeline:

1. Telegram's backend (acting as a Cocoon client) sends encrypted prompts to a proxy.
2. The proxy verifies the worker's attested environment before forwarding the request.
3. The worker processes the request in a TEE, returning a response only to the proxy, that relays it back to Telegram.

All components validate each other's integrity via RA-TLS (Remote Attestation over TLS), with the root of trust anchored in an on-chain registry maintained on the TON blockchain. This registry tracks approved proxy IPs, TEE image hashes, and model identifiers. Currently, this infrastructure is managed centrally by the Cocoon team, with plans for DAO-based governance in the future.

Telegram, which claims over 800 million active users globally, is positioning Cocoon as a way to deliver AI features without compromising its long-standing privacy principles. Unlike traditional AI deployments, where user data is routed through centralized servers, Cocoon ensures that requests are verifiably encrypted and handled within isolated environments whose integrity is cryptographically attested.

The underlying design means GPU hosts cannot access or log any user data. Additionally, reputation scores for proxies and workers, based on metrics like response time and success rate, are stored transparently on-chain, helping clients route their requests to trustworthy compute nodes.

While Telegram has not disclosed which specific open-source models are used for summarization, the Cocoon backend supports model execution via vLLM, suggesting compatibility with large language models like LLaMA or Mistral.

As AI features become more embedded in messaging platforms, users should remain aware of how their data is processed. For users concerned with privacy in the age of AI, it is recommended to check in-app settings and disable these features where possible.