Cocoon is live · pricing not published · last verified 3 Nov 2026 · viability monitor
Cocoon
Cocoon is Telegram's decentralized confidential AI inference network on TON, paying GPU owners in TON to serve models inside Intel TDX enclaves.
Cocoon is a genuinely different bet: instead of selling AI inference, it coordinates GPU owners on TON and pays them in TON for serving models inside Intel TDX enclaves, with RA-TLS attestation and reproducible builds so a buyer can verify what actually ran. If you are a TON or Telegram developer who needs attestable, private inference and you are comfortable paying in TON, it is the most interesting option in that niche. If you need a mature hosted inference API with predictable billing and broad model coverage, established GPU clouds remain the safer route. Treat Cocoon as early-stage and hardware-gated — the docs reachable today cover architecture, deployment and a Qwen/Qwen3-0.6B build
Verdict generatedVerified 2d ago · liveness 66/100 · cite: rightaichoice.com/tools/cocoon
- Telegram ecosystem developers wanting private AI features
- GPU owners with Intel TDX hardware seeking to earn TON
- Projects needing verifiable and confidential AI inference
- TON blockchain dApp developers
- Teams that need a mature hosted inference API with published limits
- Users without TON blockchain familiarity
- Operators without Intel TDX-capable GPU hardware
We scan live Reddit threads, YouTube comments, X posts, G2 reviews and other communities — and hand you an honest verdict in under a minute.
- Honest verdict, not marketing
- Real pros & cons from real users
- Attributed quotes with receipts
3 free scans · no card needed
Skip Cocoon if you need a drop-in hosted inference API with published model coverage and per-token billing today, or if you have no Intel TDX-capable GPU hardware and no TON wallet to settle payments.
Serving models requires a confidential-computing GPU server with Intel TDX, so the real entry cost is hardware you may not own yet, not a subscription line item.
We did not reach a pricing page for Cocoon in this run, so treat any cost estimate as unknown rather than cheap. What is documented is the settlement mechanism, not a rate card: developers reward GPU providers with TON through smart contracts, and GPU owners earn TON for serving models. Budget for hardware and integration engineering before you budget for per-inference fees.
In short
Cocoon — Cocoon is Telegram's decentralized confidential AI inference network on TON, paying GPU owners in TON to serve models inside Intel TDX enclaves. Best for Telegram ecosystem developers wanting private AI features, GPU owners with Intel TDX hardware seeking to earn TON, Projects needing verifiable and confidential AI inference. Pricing not published.
What is Cocoon?
Cocoon — Cocoon is Telegram's decentralized confidential AI inference network on TON, paying GPU owners in TON to serve models inside Intel TDX enclaves. Best for Telegram ecosystem developers wanting private AI features, GPU owners with Intel TDX hardware seeking to earn TON, Projects needing verifiable and confidential AI inference. Pricing not published.
Cocoon (Confidential Compute Open Network) is a decentralized AI inference network built on the TON blockchain and introduced by Telegram. It splits the work between three parties. GPU owners contribute confidential-computing GPU servers and earn TON for running models. App developers plug Cocoon's verifiable inference into their backends and pay GPU providers in TON. End users interact with AI without exposing their data to a central provider. Confidentiality comes from running models inside Intel TDX trusted execution environments, with remote attestation over TLS (RA-TLS) and persistent key derivation via SGX/TDX seal keys. Payments settle through TON smart contracts. Distribution integrity is handled with reproducible builds: anyone can rebuild the worker VM image (./scripts/build-image prod, ./scripts/prepare-worker-dist, then sha256sum against the published checksums) and compare model builds tagged with hash and commit — the docs use Qwen/Qwen3-0.6B as the worked example. The project is early-stage: Telegram's own site frames it as connecting GPU power, AI and "Telegram's vast ecosystem," Pavel Durov introduced it in a keynote at Blockchain Life 2025, and it ships a ready-to-run TDX worker image plus setup scripts for GPU providers alongside architecture, developer, deployment, TDX/image, RA-TLS, GPU passthrough, smart-contract and seal-key documentation. It is aimed at crypto-native teams and Telegram developers building private AI features, and at GPU owners who already have Intel TDX-capable hardware. It is not a general-purpose hosted AI assistant, and it is not a substitute for a mature cloud inference provider.
Key features
- Decentralized AI inference on the TON blockchain
- Confidential inference inside Intel TDX trusted execution environments
- GPU owners earn TON for serving models
- Remote attestation over TLS (RA-TLS) with proxy-cli and certificate generation
- Persistent key derivation via SGX/TDX seal keys
- GPU passthrough and confidential-computing validation
- TON smart-contract based payment system
- Reproducible worker VM image builds verified with sha256sum against published checksums
- Reproducible model image builds tagged with model name, hash and commit
- Ready-to-run TDX worker image and setup scripts for GPU providers
- Developer integration path for embedding verifiable inference into app backends
- Architecture documentation for developers and GPU providers
- Deployment, testing and TDX debugging documentation
- Qwen model serving example (Qwen/Qwen3-0.6B) in the build docs
- Keynote introduced by Pavel Durov at Blockchain Life 2025
Our long-form take written Jul 14, 2026 is more than 90 days old, so it is being re-reviewed against current data and is not shown.
Models under the hood
as of 2026-10-30
Limitations
- Cocoon is early-stage and hardware-gated: serving models requires confidential-computing GPU servers with Intel TDX, which rules out most consumer hardware.
- The documentation we could reach describes architecture, deployment, TDX/image tooling and a Qwen/Qwen3-0.6B model build example, but we found no formal inference API reference, model catalog, or rate limits published on the pages we reached; the main cocoon.org site is a landing page that routes to Documentation, Downloads, GitHub and socials rather than a product console.
- Payments run through TON smart contracts, so both GPU providers and developers need TON familiarity and a wallet.
- The network's economics depend on developer inference demand matching GPU supply, which is unproven at this stage.
as of 2026-10-09
Researching Cocoon? Get your full AI stack in 60 seconds.
Free, no signup — tell us your goal and get tools matched to your budget & existing stack.
Is Cocoon still active?
Cocoon is live · pricing not published · last verified 3 Nov 2026 · viability monitor.
Viability score
How well maintained and how widely used is Cocoon? Built from what the vendor actually publishes (docs, changelog, tutorials, integrations, pricing), whether the site is live, and how much real users discuss it. How we calculate this
Last calculated: November 2026
How we score →What do users say about Cocoon?
What people actually say about Cocoon — is it worth it?
Our community sentiment research written Aug 2, 2026 is more than 90 days old, so it is being re-reviewed against current data and is not shown.
How we verified this
Every listed fact about Cocoon — pricing, features, links, status — is re-read from the vendor's own pages by our verification pipeline and each change is logged with a date; the last pass ran on 3 Nov 2026. Analysis is drafted with AI models and checked against our editorial standards before publishing; see how we verify.
Verification history
We have re-verified Cocoon 13 times since . Each pass re-reads the vendor's own pages and re-checks every listed field against that evidence; passes where nothing had changed are marked as such.
- — changed pricing page
- — changed pricing page
- — changed pricing page
- — changed who should skip it, setup time, our verdict, pricing fit, features, hidden costs, switching out, description, limitations, switching in, summary
- — changed pricing page, limitations
- — changed use cases, limitations, switching in, switching out, who should skip it, hidden costs, features, who it suits, our verdict, pricing fit, setup time, description, summary
Showing the 6 most recent of 13 verification passes.
Free to cite with attribution — this page re-verifies continuously.
How much does Cocoon cost in 2026?
Cocoon: Pricing not published. Pricing last verified 3 Nov 2026.
Where the pricing makes sense
The company stage and team size where Cocoon's pricing actually pencils out — and where peers do it cheaper.
We did not reach a pricing page for Cocoon in this run, so treat any cost estimate as unknown rather than cheap. What is documented is the settlement mechanism, not a rate card: developers reward GPU providers with TON through smart contracts, and GPU owners earn TON for serving models. Budget for hardware and integration engineering before you budget for per-inference fees.
Has Cocoon's pricing changed?
No plan price for Cocoon has moved since we began logging in Aug 2026; the pricing page is re-read on every verification pass.
Who should pick Cocoon?
Pick Cocoon if you are Telegram ecosystem developers wanting private AI features, GPU owners with Intel TDX hardware seeking to earn TON, Projects needing verifiable and confidential AI inference, TON blockchain dApp developers.
Look elsewhere if you are Teams that need a mature hosted inference API with published limits, Users without TON blockchain familiarity, Operators without Intel TDX-capable GPU hardware, Projects that cannot take on TEE setup and debugging work.
Skip it if Skip Cocoon if you need a drop-in hosted inference API with published model coverage and per-token billing today, or if you have no Intel TDX-capable GPU hardware and no TON wallet to settle payments.
GeneratedReal-world workflow fit
Concrete scenarios for the personas Cocoon actually fits — and what changes day-one when you adopt it.
You build a private AI feature inside a Telegram mini-app. Rather than sending user prompts to a central inference provider, you integrate Cocoon's verifiable inference into your backend so requests run inside Intel TDX enclaves, then pay the GPU providers in TON through the smart-contract payment system.
Outcome: Your users get AI answers while their prompts stay inside an attested enclave, and you can point to RA-TLS attestation and a reproducible worker build as evidence of what actually ran.
You download the latest worker release — a ready-to-run TDX image plus setup scripts — bring up your confidential-computing GPU server, validate GPU passthrough and confidential computing, and join the network to serve models.
Outcome: Your idle TDX-capable GPU capacity earns TON for the inference it serves, and you can independently verify the distribution by rebuilding from source and comparing sha256sum output.
Before trusting a third-party build, your team runs ./scripts/build-image prod and ./scripts/prepare-worker-dist, then compares images/prod checksums against the published release, and runs ./scripts/build-model to confirm a model tar carries the expected hash and commit.
Outcome: You have a documented chain from published checksum to running image, which is the kind of artefact most hosted inference vendors do not give you at all.
Use cases
- Deploy private AI features inside Telegram apps without exposing user data to a central provider
- Earn TON by contributing idle Intel TDX-capable GPU servers to serve AI models
- Build confidential AI features for dApps on the TON blockchain
- Integrate verifiable AI inference into decentralized identity or voting systems
- Provide attestable, privacy-preserving AI compute for research or educational tools
- Verify that a served model matches the published build by rebuilding and comparing checksums
Setup time & first value
How long it actually takes to get something useful out of Cocoon — broken out by persona, not the marketing-page minute.
GPU owners: budget hours, not minutes — you download the TDX worker release, follow the included setup guide, and validate GPU passthrough and confidential computing before your first model serves. Developers: integration time depends on your backend, since Cocoon is a network you plug into rather than an SDK you import and forget. Verifying the build with sha256sum is optional and adds its own
Switching to or from Cocoon
How to bring data in from common predecessors and how to get it back out — written for the switcher, not the buyer.
- →From a hosted inference API: re-point your backend at Cocoon's verifiable inference and settle with GPU providers in TON — expect to rewrite billing and retry logic, not just a base URL.
- →From self-hosted open-model serving: keep your model weights and add TDX worker deployment so inference runs inside an attested enclave instead of your own box.
- →From no inference infrastructure: start by reading the Architecture and Deployment docs, then build a worker image and confirm checksums against the published release.
- ↗To a hosted inference provider: the standard path if you need published rate limits and a support contract, since Cocoon currently documents architecture and tooling rather than commercial terms.
- ↗To self-hosted serving: if TON settlement and enclave attestation turn out to be more operational overhead than your team wants, move the same open models onto your own GPUs.
- ↗To another TEE-based inference service: attestation concepts (RA-TLS, seal keys) transfer, but the TON payment layer and reproducible-build scripts do not.
Resources & guides
- Documentationcocoon.org
Docs · Cocoon
Full product docs from cocoon.org
- Documentationcocoon.org
Architecture · Cocoon
Full product docs from cocoon.org
- Documentationcocoon.org
Developers · Cocoon
Full product docs from cocoon.org
- Documentationcocoon.org
Deployment · Cocoon
Full product docs from cocoon.org
- Documentationcocoon.org
Tdx · Cocoon
Full product docs from cocoon.org
- Documentationcocoon.org
Ra Tls · Cocoon
Full product docs from cocoon.org
- Documentationcocoon.org
Gpu · Cocoon
Full product docs from cocoon.org
- Documentationcocoon.org
Smart Contracts · Cocoon
Full product docs from cocoon.org
- Documentationcocoon.org
Seal Keys · Cocoon
Full product docs from cocoon.org
Tutorials & Learning
YouTube returned 6 videos for “Cocoon”, and we withheld 6: 6 could not be judged, because “Cocoon” is a single word that other videos use for other things. We are showing none, because we could not prove any of them are about Cocoon.
Official links
What are the best Cocoon alternatives?
The closest alternatives to Cocoon we list are Petals, DeepInfra and Recogni. Ranked by direct product-type match, not category overlap.
| Alternative | Pricing | Starts at | Last verified |
|---|---|---|---|
| Petals | Free | Free | — |
| DeepInfra | Paid | Pricing not published | — |
| Recogni | Contact Sales | Custom pricing | — |
Alternatives to Cocoon
View all
Tools that pair well with Cocoon
Common stack mates teams adopt alongside Cocoon, with the specific reason each pairing earns its keep.
Petals
Run Llama 3.1 405B and other huge language models at home over a BitTorrent-style decentralized inference network.
DeepInfra
Serverless AI inference on 100+ open models via an OpenAI-compatible API, billed per token with no GPU overhead.
Recogni
Rack-scale AI inference system using log-space math to serve multi-trillion-parameter models at 608 PFLOPS per rack.
Cocoon vs its closest rival
Cocoon's closest head-to-head is with Spider Cloud.
Our Cocoon vs Spider Cloud verdict written Jul 3, 2026 is more than 90 days old, so it is being re-reviewed against current data and is not shown.
Head-to-head comparisons
Cocoon vs Spider Cloud
Verdict being re-reviewed against current pricing and data.
Cocoon vs Temporal Ai
These aren't competitors — don't shortlist them against each other. If you're building a long-running AI agent or multi-step workflow that must survive crashes and retries, Temporal is the clear pick, with Python/Go/TypeScript SDKs, automatic Activity retries, Signals/Queries/Updates, and Serverless Workers in public preview on Google Cloud Run (2026-10-05). If you specifically want confidential inference settled on TON, with GPU owners paid in TON and models running inside Intel TDX enclaves, Cocoon is the only one of the two that does that — but it demands TON familiarity and TDX hardware.
GeneratedCocoon vs Presto Voice
Verdict being re-reviewed against current pricing and data.
Frequently asked questions
Categories
Best-of guides
Topics
For makers of Cocoon
- Get a Verified Listing for Cocoon ($39): a written decision within 1 working day.
- Free listing for another tool: Free reviews are decided in order, within 3 weeks.
Used Cocoon? Help shape our editorial sentiment research.