Just today, Telegram founder Pavel Durov announced a brand-new decentralized concept called Cocoon.
So how does Cocoon work?
It’s built on the TON blockchain and uses Intel TDX TEE (Trusted Execution Environment) technology to ensure the privacy of all computation. This means your AI data can’t be intercepted or viewed by any intermediary — it’s essentially processed as if it were running on your own local device.
GPU providers install node software and contribute compute power to run AI tasks, such as model inference, in exchange for TON rewards.
Developers access the network via API to submit tasks. Telegram itself will be a major demand source — features like message translation, speech-to-text, and summary generation have already begun using Cocoon.
The entire network is open, and more GPU providers and developer workloads will be added in the future.
Now, the benefits.
For GPU owners, this is a real earning opportunity — especially for people with idle high-end GPUs. They can directly mine TON without worrying about the fees associated with centralized compute platforms.
Developers benefit by gaining access to low-cost computing resources, instead of paying the high premiums charged by Amazon or Microsoft.
And for users? Privacy is the biggest selling point — AI features are computed in a fully confidential environment, unlike traditional services that collect user data.
Telegram users will soon see new AI features, including more secure chat assistants. Overall, Cocoon challenges the monopoly of centralized computing, offering a fairer economic model and stronger privacy protection while connecting GPU providers with those who need compute.
The purpose of launching Cocoon is to break centralized barriers and let the massive amount of idle GPU power around the world actually do meaningful work.
This also overturns the traditional idea of GPU mining, where GPUs perform useless calculations. Cocoon instead uses GPU power to run lightweight but practical AI tasks, and the requesters pay the GPU providers directly.
For example, the Cocoon website shows a command like:
./scripts/cocoon-launch --instance 1 --worker-coefficient 2000 --model Qwen/Qwen3-0.6B worker.confChallenges
Although the vision is great, I personally see some challenges.
First, the official requirement is an NVIDIA GPU with CC support (H100+), which means consumer gaming GPUs basically don’t qualify. The entry bar is extremely high, making this more suitable for individuals or teams already running AI models.
Lower-tier GPUs won’t be able to handle meaningful models — some may not even run basic translation tasks, and might only be useful for simple math problems or lightweight inference.
I think if Telegram could release unified APIs for multiple AI models, and pair that with a broader GPU provider ecosystem, the platform would become far more compelling.
But for now, demand still seems limited, and it’s unclear whether GPU contributors will actually earn meaningful returns.
Comments NOTHING