ai

Decentralized AI Compute Compared: Cocoon vs Bittensor vs Gonka

Daniil Okhlopkov

• 7 min read

Three protocols are competing to become the default infrastructure for decentralized AI compute: Cocoon (built on TON), Bittensor (Polkadot/Substrate), and Gonka (Cosmos SDK). Each takes a fundamentally different approach to the same problem: how do you trust a decentralized network to run your AI workloads?

I spent the past week digging into the architecture, tokenomics, and real-world performance of all three. Here's what I found.

What Problem Are They Solving?

Running AI inference is expensive. A single H100 GPU costs $2-3/hour on AWS. Training a frontier model costs millions. The centralized cloud providers (AWS, GCP, Azure) control pricing, and demand is outstripping supply.

Decentralized compute networks flip this: anyone with GPUs can contribute compute, and developers can access it at lower cost without vendor lock-in. The question is how each protocol handles trust, quality, and coordination.

How Does Bittensor Work?

Bittensor is the oldest and largest of the three ($3B market cap, rank #33). Its core innovation is the subnet model: 125+ specialized marketplaces running under one token (TAO, 21M fixed supply, Bitcoin-style halvings).

Each subnet is a separate competition. Miners compete to provide the best output for a specific task. Validators score their work off-chain. The Yuma Consensus mechanism aggregates scores (weighted by stake) and distributes TAO rewards.

The key insight: Bittensor doesn't just sell compute. It creates competitive markets for AI quality. The best model wins, not just the cheapest GPU.

Concrete Bittensor Subnet Examples

Vanta / SN8 (Trading): The most successful subnet by revenue. Miners submit long/short/flat trading signals across crypto, forex, and equities. Validators track real PnL, Sharpe ratio, and drawdowns over 90 days. The market itself is the judge, no AI evaluation needed. Annual reward pool: $30M+. Recently launched Vanta Trading, a decentralized prop trading challenge with $149-$349 evaluation fees and 100% profit splits. Data partnerships with Glassnode and LunarCrush feed the ecosystem. AI is not required (miners can trade manually), but competitive pressure pushes toward ML models.

Chutes / SN64 (Inference, $133M cap): Miners host open-source models like Llama and DeepSeek. Validators verify outputs by sending the same prompt to multiple miners and comparing results (deterministic models produce identical outputs). Scoring: latency + uptime + correctness. Has processed 9.1T tokens and claims 85% cost reduction vs AWS. This is the closest to a traditional inference API.

Templar / SN3 (Distributed Training, $135M cap): The most technically ambitious subnet. Successfully trained a 72B-parameter model across 70+ geographically distributed nodes using gradient verification. Validates training contributions through mathematical proof (SparseLoCo), not AI judgment.

OCR Subnet (Tutorial example): Validators generate synthetic invoices, corrupt them (blur, rotation, noise), and send them to miners. Miners run OCR and return text + bounding boxes. Score = edit distance (text accuracy) + IoU (position accuracy) + time penalty. Straightforward, deterministic quality measurement.

How Does Bittensor Validate Quality?

This is the critical question, and the honest answer is: it depends on the subnet.

Some subnets have elegant validation. Trading subnets use real market outcomes as ground truth. OCR uses edit distance against known documents. Inference subnets cross-verify deterministic model outputs.

But the text generation subnet (SN1) has a circular problem: validators generate a reference answer with an LLM, then score miners by comparing their LLM output to it. Both sides are calling GPT. You are essentially rewarding "sounds like GPT," not actual correctness.

An academic paper analyzing Bittensor found that stake predicts earnings 3-8x better than actual performance quality. The performance-to-reward correlation for miners is only 0.10-0.30, while stake-to-reward correlation reaches 0.50-0.80. Over half of subnets require under 1% of wallets to hold 51% of stake.

Translation: Bittensor is more a game of capital than a game of quality. The subnet architecture is genuinely innovative, but the incentive alignment has real problems.

What About Compute Overhead?

A common concern: if validators re-run inference to check quality, aren't you paying for compute twice?

In practice, it's not that bad. Chutes (SN64) routes each request to one miner (not all 192). Validators do periodic spot-checks by sending the same prompt to multiple miners, but this is sampling-based, roughly 5-10% overhead, not 2x. The OCR subnet works similarly: validators check a subset of results, not every single one.

The real overhead is economic, not computational. TAO emissions are inflationary, and most subnet value is speculative rather than backed by actual service revenue. Only a handful of subnets (Chutes, Taoshi/Vanta) generate meaningful external income.

How Does Gonka Work?

Gonka takes a completely different approach. Where Bittensor incentivizes AI quality, Gonka incentivizes useful computation.

The core innovation is Proof of Work 2.0: instead of wasting GPU cycles on hash puzzles (Bitcoin) or consensus overhead (PoS), Gonka claims ~100% of compute goes toward real AI inference.

The mechanism works in two phases:

  1. Sprint (epoch start): All hosts run a standardized transformer workload with randomized layers. The number of valid outputs determines your voting weight for the epoch. This proves GPU capability.
  2. Inference (rest of epoch): Hosts process real inference requests from developers. 1-10% are randomly selected for re-verification by other hosts.

Verification is probabilistic, not deterministic. If you get caught cheating, you lose your entire epoch of rewards (323,000 GNK per epoch). New hosts face 100% validation rate that decreases as reputation builds.

Current state: 3 models available (Qwen3-235B, GPT-OSS-120b, Qwen3-235B-Thinking), 6,000+ H100-equivalent GPUs, 448+ active hosts, $80M raised from Bitfury and Coatue. The developer experience is clean: OpenAI-compatible API, SDKs for Python/TypeScript/Go, "switch one line of code from OpenAI" positioning.

The limitation: Gonka is an inference API, not a GPU rental platform. You cannot run arbitrary code. You call an endpoint with a supported model and pay in GNK tokens. This is closer to using OpenAI than renting an AWS instance.

How Does Cocoon Work?

Cocoon (Confidential Compute Open Network) is the newest entrant, announced by Pavel Durov at Blockchain Life 2025 and live since November 30, 2025. It runs on TON blockchain and takes a radically different trust model.

Where Bittensor trusts economic incentives and Gonka trusts probabilistic spot-checks, Cocoon trusts hardware.

The architecture: Client > Proxy > Worker. Every worker runs inside an Intel TDX confidential VM with NVIDIA Confidential Computing. The connection uses RA-TLS (Remote Attestation over TLS), which means the GPU owner physically cannot see your prompts, model weights, or outputs. Everything is encrypted at the hardware level, verified by Intel and NVIDIA root keys.

Key technical details:

  • dm-verity ensures model integrity. Clients can cryptographically verify that the worker runs the exact model they requested, not a cheaper substitute.
  • On-chain Root Contract stores proxy IPs, allowed image hashes, verified model hashes, and payment contract code.
  • No separate token. All payments are in TON. GPU providers earn TON directly from inference fees. No inflationary emissions, no speculative token economics.
  • Hardware requirement: NVIDIA H100+ with Confidential Computing support. Consumer GPUs cannot participate.

Currently supported models: DeepSeek and Qwen (open-source), with more being added.

Cocoon vs Bittensor vs Gonka: Side-by-Side Comparison

Cocoon (TON)

  • Trust model: Hardware attestation (Intel TDX + NVIDIA CC)
  • Token: TON (no native token)
  • Revenue for providers: Real inference fees in TON
  • Privacy: End-to-end encrypted, hardware-enforced
  • GPU barrier: H100+ only
  • Models: Curated registry (DeepSeek, Qwen)
  • Demand engine: Telegram (950M+ MAU)
  • Validation overhead: Minimal (attestation at connection)
  • Blockchain: TON
  • Maturity: Live Nov 2025
  • Open-source: GitHub (Apache-2.0)

Bittensor (TAO)

  • Trust model: Economic (stake-weighted quality scores)
  • Token: TAO (21M supply, halvings)
  • Revenue for providers: Mostly TAO inflation + some fees
  • Privacy: None
  • GPU barrier: Varies by subnet
  • Models: Per-subnet, miners choose
  • Demand engine: Organic / speculative
  • Validation overhead: 5-10% (sampling)
  • Blockchain: Substrate (Polkadot)
  • Maturity: Mature, $3B mcap
  • Open-source: GitHub (MIT)

Gonka (GNK)

  • Trust model: Probabilistic spot-checks + reputation
  • Token: GNK (1B supply, halvings)
  • Revenue for providers: GNK mining rewards + inference fees
  • Privacy: None
  • GPU barrier: Multi-GPU, high VRAM
  • Models: 3 models (governance-voted)
  • Demand engine: Organic (2,200+ developers)
  • Validation overhead: 1-10% (spot-checks)
  • Blockchain: Cosmos SDK
  • Maturity: Mainnet Aug 2025, $570M mcap
  • Open-source: GitHub

Why Cocoon Has the Strongest Position

After researching all three, I think Cocoon has three structural advantages that the others cannot easily replicate:

1. Telegram as a guaranteed demand engine. Bittensor and Gonka need to attract developers from scratch. Cocoon starts with Telegram as its first customer, and Telegram has 950M+ monthly active users. That is not a speculative demand story. Durov confirmed that Telegram already routes translation and experimental AI features through Cocoon. The demand is real and growing.

2. No separate token means no inflation problem. Bittensor's TAO and Gonka's GNK both have inflationary emission schedules. GPU providers earn newly minted tokens, which means the system works as long as the token price goes up. If it doesn't, providers leave and the network shrinks. Cocoon avoids this entirely: providers earn TON from actual inference fees. Revenue is demand-driven, not emission-driven. This is a fundamentally healthier economic model.

3. Hardware-enforced privacy is a real differentiator. For enterprise use cases, knowing that even the GPU owner cannot see your data is a concrete selling point. Bittensor and Gonka have no privacy guarantees whatsoever. As AI inference increasingly handles sensitive data (medical records, financial data, personal conversations), hardware-level confidentiality becomes a requirement, not a nice-to-have.

The trade-offs are real. Cocoon requires H100+ GPUs (limiting decentralization to datacenter operators), proxies are currently centralized, and the model registry is curated rather than open. But these are early-stage limitations that can be addressed. The structural advantages are harder to replicate.

What Can TON Learn from Bittensor and Gonka?

Cocoon is live and working, but there are ideas worth borrowing:

  • From Bittensor: The subnet model. Right now Cocoon is a single-purpose inference network. If it evolves into a platform where anyone can create specialized AI marketplaces (trading signals, data labeling, model training) with TON as the coordination token, the network effects compound.
  • From Gonka: The "100% useful compute" narrative. Cocoon already uses TEE for trust, but positioning compute efficiency as a first-class feature strengthens the value proposition against centralized alternatives.
  • From both: Developer ecosystem matters. Gonka's OpenAI-compatible SDK is a good template. The lower the switching cost, the faster adoption grows.

Frequently Asked Questions

Can I mine TAO without owning GPUs?

Yes. You can delegate (stake) TAO to a validator and earn ~18-20% APY. Minimum: 0.1 TAO. You can also mine on trading subnets like Vanta (SN8) using algorithmic strategies without GPU requirements.

Is Bittensor really decentralized?

Partially. The top 12 validators hold 79% of stake with no lock-up period. Over half of subnets require under 1% of wallets to control 51% of stake. The architecture is decentralized, but the economics are concentrated.

How much does Cocoon inference cost?

Pricing is dynamic and market-driven. GPU providers set their rates, and developers pay per inference request in TON. No published price tables yet, but the economic pressure from competition should push prices below centralized alternatives.

Which protocol is best for privacy-sensitive workloads?

Cocoon, by a wide margin. It is the only protocol with hardware-enforced confidentiality (Intel TDX + NVIDIA Confidential Computing). Bittensor and Gonka have no privacy guarantees.

Can I deploy my own AI model on these networks?

On Bittensor, yes. You can register as a miner in an existing subnet or create your own subnet with custom incentive mechanisms. On Gonka and Cocoon, models are curated/governance-voted, so you cannot deploy arbitrary models yet.

I'm Dan Okhlopkov, Head of Analytics at TON Foundation. I write about on-chain data, AI agents, and the economics of decentralized infrastructure.

Telegram | X / Twitter | GitHub | Instagram | Threads