Proof of Useful Work: A Complete Guide to Useful Mining

Bitcoin consumes about 26 GW of power – that's ~150 TWh per year, comparable to Argentina's energy consumption. Every watt goes into computing SHA-256 hashes, whose sole purpose is to prove energy expenditure. In 15 years, Bitcoin has produced no useful results: no neural network response, no scientific calculation, no renders. All this energy is pure payment for security.

Ethereum recognized the problem and in September 2022 conducted “The Merge” – transitioning from Proof of Work to PoS. The network's energy consumption dropped by 99.95%. But PoS created a new problem: security now depends not on work, but on capital. Large stakers (Lido, Coinbase, Binance) control a significant share of the network. PoS sacrifices decentralization for energy efficiency – and decentralization was the blockchain's main promise.

PoUW offers a third path: retain GPU work (like in Bitcoin – security through computations), but direct this work towards real tasks (AI inference, scientific calculations, rendering). Security not through useless hashing, nor through locking up capital, but through useful work.

The concept of PoUW was formalized in the Ofelimos protocol, presented at the IACR Crypto 2022 conference – one of the world's leading cryptography forums. The idea: instead of meaningless hashing, miners solve real optimization problems. The result simultaneously confirms a block on the blockchain and creates value for the end-user.

The key challenge for PoUW is verifiability. In traditional PoW, checking the result is trivial: a hash is either less than the target or it isn't. Useful computations (a neural network response, a 3D scene render) are harder to verify. If the result cannot be quickly verified, an attacker could fake it, sending garbage instead of a real answer and claiming a reward.

Different projects solve this problem in fundamentally different ways:

• Approach 1: Mathematical proof (Gonka). Computation → PoC V2 cross-verification → BLS signature on the blockchain. 1–10% of tasks are verified by other nodes. If results don't match, a 20% collateral penalty. Security guarantees: mathematical, not subjective.
• Approach 2: Subjective evaluation (Bittensor). Computation → validators evaluate the “quality” of the response via Yuma Consensus. Problem: “quality” is subjective, and the system is vulnerable to validator collusion. Guarantees: economic (stake), not cryptographic.

The difference is critical: mathematical proof cannot be faked (regardless of how much money an attacker has). Subjective evaluation is vulnerable to a majority attack. This determines the different levels of trust in each approach.

Let's look at key projects implementing PoUW in 2026:

Gonka (PoW 2.0) – the purest implementation of PoUW. Each processed AI request simultaneously confirms a block. 99% of network resources are used for useful work, 1% for verification. The Qwen3-235B model (MoE, 22B active parameters) is served by clusters of H100/H200. $80M in investments from Coatue, Bitfury, Insight Partners.

Flux – historically one of the first PoUW projects, but in 2025, it abandoned GPU mining and switched to CPU nodes. Useful work = hosting containerized applications (Docker). Essentially, Flux became a decentralized cloud hosting provider, not an AI network. Market cap ~$23M.

Prime Intellect – focuses on distributed model training, not inference. Uses an approach similar to DiLoCo in Gonka, but as a primary product rather than an additional feature.

Bittensor is not formally PoUW in its purest form – Yuma Consensus is based on validators' subjective evaluation, not cryptographic proof. However, 126 subnets cover a wide range of AI tasks, and the project has the largest market capitalization in the segment.

The AI computing market is valued at over $150 billion and grows by 30%+ annually. Meanwhile, Bitcoin continues to burn ~150 TWh per year on empty hashes. PoUW resolves this contradiction: the same principle of “energy = security,” but the energy creates real value.

For Bitcoin miners with GPUs: after Ethereum's transition to PoS in 2022, millions of GPUs were left idle. Bitcoin mining on GPUs has long been unprofitable (requires ASICs). PoUW projects like Gonka give GPUs a second life – the same cards that once computed useless hashes now process AI requests and earn rewards.

For investors: PoUW is the convergence point of two major technological trends: crypto ($2+ trillion market) and AI ($150+ billion market). Gonka is the first project where PoUW is implemented in production with real AI requests, has passed CertiK audit, and attracted $80M from institutional investors.

The future of PoUW: currently, Gonka serves one model (Qwen3-235B). The roadmap includes multi-model inference: hosts will be able to serve different models (text, code, image) depending on their GPUs. DiLoCo adds distributed learning – Gonka will be able not only to run but also to train models. This transforms Gonka from an inference network into a full-fledged AI platform – open, decentralized, and based on mathematically verifiable Proof of Useful Work.