EPOCH

Abstract

The global demand for AI compute is growing at 10x annually while centralized cloud providers maintain artificial scarcity, extracting rents from both developers and hardware owners. Billions of dollars worth of GPU compute sits idle on consumer machines, gaming rigs, and enterprise hardware during off-peak hours.

EPOCH is a permissionless, decentralized compute network that connects idle hardware directly to AI workloads. Providers run the lightweight EPOCH node client, contribute cycles to training and inference tasks, and earn $EPOCH tokens proportional to their verified compute contribution. No intermediaries. No cloud provider markup. No KYC gates.

This paper outlines the protocol architecture, the Proof of Compute (PoC) consensus mechanism, the task distribution layer, and the economic model that aligns incentives between compute providers, task requesters, and the network itself.

The Problem

2.1 Compute Scarcity is Artificial

As of 2025, the top three cloud providers (AWS, GCP, Azure) control over 65% of the global cloud compute market. GPU instance pricing has increased 300% since 2022 while underlying hardware costs have remained relatively flat. The bottleneck is not physical hardware. It is access.

Meanwhile, an estimated 40 million consumer GPUs with ML-capable VRAM sit idle for 18+ hours per day. The aggregate compute potential of this idle hardware exceeds the total capacity of all major cloud providers combined.

2.2 Existing Solutions Fall Short

Current decentralized compute networks suffer from three critical weaknesses:

• Verification gap: Most networks cannot cryptographically prove that computation was actually performed correctly, relying on reputation or redundancy.
• Latency overhead: Task distribution is inefficient, routing workloads through centralized orchestrators that defeat the purpose of decentralization.
• Cold start friction: Onboarding requires complex setup, wallet configuration, and staking, creating a barrier that limits network growth.

2.3 The Opportunity

A network that solves verification, minimizes latency, and reduces onboarding to a single CLI command can capture the massive surplus of idle consumer and enterprise compute. EPOCH is that network.

EPOCH Protocol

EPOCH operates as a three-layer protocol stack connecting hardware providers to AI workloads through a trustless verification system.

3.1 Task Submission

Requesters submit AI workloads through the EPOCH SDK or REST API. Each task specifies compute requirements (GPU VRAM, model framework, estimated FLOPs), a maximum bid price in $EPOCH, and a deadline. Tasks are decomposed into atomic units called work packets that can be independently verified.

3.2 Node Matching

The orchestration layer matches work packets to available nodes using a multi-factor scoring algorithm:

3.3 Execution and Verification

Nodes execute work packets inside a sandboxed GPU runtime. Upon completion, the node produces a Proof of Compute (PoC) hash that is submitted to the settlement layer. The protocol uses a challenge-response verification system described in Section 05.

Network Architecture

4.1 Node Client

The EPOCH node client is a lightweight daemon (~45MB) that runs on Linux, macOS, and Windows. It performs hardware detection, manages the GPU runtime sandbox, communicates with the orchestration layer, and submits PoC proofs.

4.2 GPU Runtime Sandbox

All computation runs inside an isolated container environment. The sandbox prevents provider nodes from accessing task data outside their assigned work packet, ensures deterministic execution for verification, and supports CUDA, ROCm, and Metal backends.

4.3 Orchestration Mesh

Unlike centralized orchestrators, EPOCH uses a distributed hash table (DHT) based mesh where orchestration responsibilities are shared across high-uptime nodes. This eliminates single points of failure and ensures task routing remains functional even if individual orchestrator nodes go offline.

4.4 Supported Workloads

Proof of Compute (PoC)

EPOCH introduces Proof of Compute (PoC), a consensus mechanism designed specifically for verifying GPU computation on a decentralized network. Unlike Proof of Work which wastes compute on hash puzzles, PoC verifies that useful computation was performed correctly.

5.1 How PoC Works

1. Task Seeding: Each work packet includes a deterministic seed derived from the task hash. This ensures identical inputs produce identical outputs across any compliant hardware.
2. Checkpoint Hashing: During execution, the runtime captures intermediate state hashes at predefined intervals (every N forward passes for training, every K tokens for inference). These form a Merkle tree of computation checkpoints.
3. Challenge Protocol: After a node submits its result, a randomly selected validator node re-executes a subset of checkpoints (typically 3-5%) and compares hashes. If all challenged checkpoints match, the computation is verified.
4. Settlement: Verified results and their PoC proofs are anchored to Solana. The $EPOCH reward is released to the provider.

5.2 Validator Selection

Validators are selected using a weighted random function based on stake and historical accuracy. Validators earn a 5% fee on the tasks they verify. Malicious validators (those who approve fraudulent work) are slashed and removed from the validator set.

5.3 Dispute Resolution

If a challenge fails, the full work packet is re-executed by three independent validator nodes. Majority consensus determines the correct result. The offending node's stake is slashed and the task is re-routed. Persistent offenders are permanently blacklisted from the network.

Tokenomics

6.1 Token Overview

$EPOCH is the native utility token of the network, deployed on Solana. It serves as the payment medium between task requesters and compute providers, the staking asset for validators, and the governance token for protocol upgrades.

6.2 Supply Distribution

6.3 Reward Mechanics

Provider rewards are calculated per completed work packet using the formula:

6.4 Deflationary Pressure

5% of all task fees are burned permanently, creating sustained deflationary pressure as network usage grows. Additionally, slashed stakes from malicious nodes are burned rather than redistributed, ensuring bad actors reduce total supply.

Security Model

7.1 Sybil Resistance

Nodes must complete a hardware attestation challenge during registration that verifies physical GPU capabilities. This prevents a single entity from registering virtual GPUs to claim rewards without performing real computation. The attestation runs a standardized benchmark that must complete within hardware-specific time bounds.

7.2 Data Privacy

Task data is encrypted in transit and at rest. Providers only receive the specific work packet assigned to them, never the full dataset. The sandbox prevents data exfiltration, and all intermediate states are wiped after task completion and verification.

7.3 Network Partitioning

The DHT-based orchestration mesh is resilient to network partitions. If a segment of the network becomes unreachable, the remaining nodes continue operating independently and reconcile state when connectivity is restored. Tasks assigned to unreachable nodes are automatically re-routed after a configurable timeout (default: 120 seconds).

7.4 Economic Security

The cost of attacking the network (submitting fraudulent computation) always exceeds the potential reward. Validators must stake $EPOCH to participate, and the slashing penalty for approving invalid work exceeds the verification fee by 10x, making collusion economically irrational.

Roadmap

Conclusion

The demand for AI compute is outpacing centralized supply. The hardware already exists, distributed across millions of machines worldwide. What has been missing is a protocol that connects this idle capacity to real workloads with trustless verification, fair compensation, and zero friction onboarding.

EPOCH is that protocol. By combining a lightweight node client, the Proof of Compute consensus mechanism, and a Solana-based settlement layer, EPOCH creates a global compute mesh where anyone with a GPU can contribute to AI development and earn proportional rewards.

The era of centralized compute monopolies is ending. A new epoch begins.