# The Revival of Proof-of-Useful Work (PoUW) Source: https://eli5defi.substack.com/p/the-revival-of-proof-of-useful-work ## Summary Proof-of-Useful-Work (PoUW) designs aim to make the computation that secures a blockchain also produce valuable outputs, such as AI inference or ZK proofs, and the article surveys several projects attempting this. It cites Rafael Pass's paper, which models GPU allocation between pure mining, pure useful work, and "duplex" work where one computation provides both chain security and useful output, finding that PoUW does not make majority attacks cheaper once equilibrium prices adjust. Pearl (@prlnet) launched mainnet on April 27, 2026, using NoisyGEMM with Plonky2 proofs and claiming roughly 10% overhead for running real inference. Nockchain founder Logan Allen announced on June 10, 2026 that it will support multiple AI Compute Networks, starting with merge-mining compatibility with Pearl. The article also describes Ambient, Gonka, Quip Network and Qubic, and notes risks for each. ## Article For years, Proof of Work had one obvious flaw. It secured networks by burning compute on puzzles nobody needed. As AI infrastructure demand explodes and energy constraints tighten, a once-niche idea has returned to the center of the conversation: Instead of pointless hashing, what if the same computational effort that secures a blockchain could simultaneously deliver valuable outputs: AI inference AI training ZK proofs Optimization problems Verifiable compute The concept is elegant. The execution has been brutally hard. But it might change now with several protocols arises and give a new meaning for PoUW. Let’s dive in. For years, PoUW had two problems. Economic Incentives Verification The economic problem was incentive alignment. If the work is valuable outside the chain, miners may care more about external revenue than the token, which caused massive exodus via vampire attack like what happened to Monero x Qubic situation back in 2025. Eli5DeFi@Eli5defi While you're celebrating $ETH's all-time high, the crypto world just witnessed a $6 billion network being dominated by a $300 million upstart. What happened: ➢ @_Qubic_ offered miners up to 3× @monero’s payouts. ➢ Massive miner shift → Qubic hit 52.72% hashrate (3 GH/s). ➢ Chaos Labs @chaoslabs 1/ @_Qubic_ briefly seized 52.72% of @Monero's hashrate, crossing the threshold for network control Hashrate hit 3.01 GH/s as miners chased $3.13/day vs $0.64 on Monero. $QUBIC sales into stables drove $XMR down 28% in 30 days while $QUBIC rose 57%. https://t.co/YmkLt7YklC 6:58 AM · Aug 14, 2025 · 23.8K Views 66 Replies · 120 Reposts · 419 Likes However, the central technical barrier has always been verification. For work to secure a chain, it must be: Hard to produce Easy and cheap for the entire network to verify Tunable in difficulty Fresh every round Non-reusable across blocks Bitcoin’s SHA-256 puzzle satisfies all five properties trivially. Most useful computations do the opposite. Re-running everyone’s AI inference to verify it defeats the purpose. Modern GPUs produce non-deterministic results. Zero-knowledge proofs for frontier models have historically carried crushing overhead. This is why most early “mining for science” projects (Gridcoin, Curecoin, etc.) and even Primecoin (which hunted prime chains) never became serious consensus mechanisms. This resulted in the emergence of two architectures. Architecture A (true PoUW) makes the useful computation itself the consensus mechanism (which we will discuss today) Architecture B uses a normal blockchain (usually proof-of-stake) as a coordinator and marketplace for off-chain work. https://arxiv.org/abs/2606.06700 The recent paper from Rafael Pass above mentioned that PoUW is not automatically broken just because miners are paid for useful computation. He models a PoUW blockchain where GPU/compute resources can be allocated in three ways: Pure mining / Bitcoinia — classic useless PoW (secures chain, zero external value). Pure useful work/ Fortessia — e.g., ML inference sold to external customers (generates revenue but does not secure the chain). Duplex work/ Duplexia — the same computation produces both security (block rewards) and useful output, with some computational overhead. Equilibrium depends on two parameters: Duplex overhead (extra cost of doing both at once). Token-inference ratio (how widely adopted the native token is relative to the size of the external inference market). Crucially, Pass shows that PoUW does not make majority attacks cheaper once equilibrium prices adjust, the economic cost of a 51% attack remains tied to the block reward. In the Duplexia regime, the system can generate net-positive social value by creating additional useful computation that markets alone would not produce. But the paper also makes the standard for projects higher. Because, the best PoUW protocols are not the ones with the loudest AI narrative. They are the ones that minimize duplex overhead, generate real external demand, and make sure useful work remains tied to chain security. That gives us a better lens for the current ecosystem as this paper gives the category its first rigorous economic defense and a framework for evaluating which projects are positioned for which outcomes. @prlnet launched mainnet on April 27, 2026 — the first PoUW system running at meaningful scale with real cryptographic guarantees. Technical mechanism: Miners perform noisy matrix multiplication (a structured variant the team calls NoisyGEMM) on GPUs. They add low-rank noise to input matrices, compute the product via tiled execution, and use BLAKE3 hashes of output tiles as the PoW lottery. Because the noise is low-rank, the clean original result can be recovered efficiently. Only on winning solutions is a compact Plonky2 zero-knowledge proof generated to attest correctness. Verification is fast on-chain. The same GPU cycles can run real inference (via a vLLM plugin for models like Llama 3.3 70B) with roughly 10% overhead. This is the closest any project has come to true duplex economics. Status & economics: Bitcoin-style supply (2.1 billion PRL) with smooth polynomially decaying emission. Block time ~194 seconds. Early partnership with Together AI for inference workloads (some subsidized elements noted). High fully-diluted valuation reported in the $1.5–2B range shortly after launch, with trading initially OTC. Positioning vs. Pass model: Pearl is the clearest real-world example of duplex work. If token adoption grows relative to the inference market, it is structurally positioned for the Duplexia regime — where block rewards subsidize cheaper inference while still securing the chain. Strengths: Real ZK proofs (not statistical), direct alignment with AI workloads, production mainnet, strong technical team. Risks: Sustainability of demand once subsidies taper; whether the token captures value alongside external AI revenue. Eli5DeFi@Eli5defi If your favourite CT says there’s nothing interesting to talk about because the market is dead, hit that unfollow button. They’re washed and can’t do proper research. There are a ton of hidden gems in this space, for example, @prlnet, which is flying under everyone’s radar. 8:59 AM · May 15, 2026 · 13.1K Views 23 Replies · 11 Reposts · 76 Likes @nockchain launched with a fair launch in May 2025. It takes a fundamentally different approach: it makes the generation of zero-knowledge proofs itself the useful work that secures the chain. Technical mechanism: Miners compete to produce ZK proofs of state transitions or specific computations using a deliberately minimal verifiable virtual machine called Nock (a tiny ~12-instruction ISA optimized for efficient ZK compilation). The chain only records the succinct verified proof. This subsidizes and commoditizes ZK proving capacity — a scarce resource the broader ecosystem needs. Positioning: Nockchain is explicitly trying to become the base settlement and proving layer for the entire PoUW category rather than competing as just another specialized chain. Each new useful-work idea can launch as a verified Compute Network instead of needing its own L1 and security budget. Strengths: Clean fair launch, strong ZK infrastructure, now actively trying to aggregate the category. Lower current valuation offers asymmetric upside if the vision succeeds. Risks: Execution risk on the multi-puzzle vision; potential community backlash to the aggressive “vampire attack” framing on Pearl. On June 10, 2026, Nockchain founder Logan Allen announced that the network will support multiple AI Compute Networks, each functioning as its own ZK-proven PoUW puzzle. These networks will both secure Nockchain and subsidize useful economic activity. The first step is merge-mining compatibility with Pearl to bootstrap real GPU compute. Logan@loganallc Nockchain’s AI Compute Network will initially be compatible with merge-mining Pearl to bootstrap compute. Yes, we’re vampire attacking Pearl. nockmilio @nockmilio alpha 12:06 PM · Jun 10, 2026 · 41.8K Views 30 Replies · 40 Reposts · 208 Likes @ambient_xyz (a16z CSX-led $7.2M seed, also backed by Delphi and Amber) is building a Solana-fork L1 where miners run verified inference on one fixed, network-owned large model (600B+ parameters in the DeepSeek lineage). Technical mechanism: Uses Proof of Logits. Every token generated by the model produces internal fingerprints (logits). Verifiers spot-check a tiny random fraction. The statistical difficulty of faking consistent logits provides security with extremely low overhead (~0.1% claimed). The deliberate design choice is a single model rather than a marketplace, to keep GPUs efficiently loaded without memory fragmentation. Status: As of mid-June 2026, still pre-mainnet (testnet described as the next major milestone in recent coverage). Positioning: Strong fit for the Fortessia or Duplexia regimes if it can achieve scale. The single-model approach is high-conviction and contrarian. Strengths: Extremely low verification overhead, focused economics, excellent team and backers. Risks: Execution timeline; whether the one-model bet wins versus more flexible multi-model systems. @gonka_ai launched in August 2025 with significant backing from Bitfury/Hyperfusion (reported $50M+ investment). It quickly assembled substantial real compute — roughly 6,000+ H100-class GPUs within months. Technical mechanism: The proof-of-work puzzle (“Sprint”) is deliberately shaped to mirror the mathematics of LLM inference and transformer workloads. Mining power tracks real model-serving capacity. Inference runs off-chain; cryptographic artifacts and proofs land on-chain. Rewards flow directly to compute providers with minimal middleman extraction. Positioning: Explicitly positioned as the “Bitcoin of AI compute.” Strong emphasis on binding token security directly to verifiable inference output. Strengths: Real deployed hardware at scale, clean alignment between mining power and useful capacity. Risks: Verification details at frontier scale still maturing; crowded decentralized inference market. @quipnetwork (Postquant Labs) launched its public testnet in early April 2026 (13,000+ signups). It is the only project seriously integrating quantum hardware. Technical mechanism: Miners solve real combinatorial optimization problems (logistics, scheduling, finance, simulations) using a mix of classical CPUs/GPUs and quantum annealers (D-Wave). These problems are hard to solve but cheap to verify — a structural advantage for PoUW. The Compute-Consensus Layer uses this work to secure the chain while producing sellable outputs. A separate Asset Layer adds post-quantum security primitives. Positioning: Excellent verification properties and real-world problem class with potential paying customers. Unique hardware diversity. Strengths: Easy verification, quantum angle, post-quantum focus, open-sourced elements. Risks: Still in testnet (centrally managed for now); quantum advantage claims need independent validation. @_Qubic_ (founded 2019) is the oldest project in this category still operating at scale. It uses Useful Proof of Work (UPoW) to direct mining power toward AI training for its on-chain Aigarth project (ambitious decentralized AGI goals). Technical mechanism: Mining energy is channeled into building neural networks rather than hashes. The network emphasizes feeless transactions and high performance (C++ direct execution). Mechanism details have evolved over time. Positioning: Longest track record but also the most fluid definition of “useful.” Has faced criticism for shifting mechanisms and past controversial use of aggregated hashpower. Strengths: Live network with claimed large miner base, high performance claims, direct training focus (larger theoretical TAM). Risks: History of mechanism changes reduces credibility for some observers; tokenomics experiments add complexity. The technical foundations for PoUW have advanced dramatically. Efficient matmul ZK proofs, Proof of Logits, structured noisy computations, and minimal verifiable VMs have made Architecture A practical in ways that were impossible even two years ago. The Rafael Pass paper provides the first solid economic map showing how these systems can create net-positive value rather than just rearranging existing compute. Nockchain’s aggressive move to become a meta-layer for multiple PoUW puzzles adds a new competitive dynamic. The remaining questions are now mostly economic and adoption-driven: How large is the market willing to pay a decentralization premium for verifiable, trust-minimized compute? Can tokenomics be designed so that useful-work revenue and security reinforce rather than cannibalize each other? Which projects will actually reach the Duplexia regime where the blockchain subsidizes genuinely additional useful computation? Pearl currently leads in production readiness and cryptographic rigor. Nockchain is making the boldest infrastructure bet. Ambient, Gonka, and Quip each occupy distinct and defensible niches. PoUW is no longer just an idea or a collection of academic papers. Real systems are live or in advanced testing, and the economic theory is catching up. At the end of the day, real usage data, miner behavior, and paying demand will determine whether this becomes one of the most important developments at the intersection of crypto and AI, or another promising category that struggles with the last mile of sustainable economics. NFA. DYOR. Substack → Telegram → https://t.me/eli5definews Disclaimer: This information is for educational purposes only and does not constitute professional financial or tax advice. Some content may be developed in collaboration with third parties, and we may hold positions in the assets mentioned. We strongly recommend conducting independent research and consulting with a qualified professional before making any financial or tax-related decisions.