Proof of Useful Work: Pearl and Nock

Bitcoin miners compute SHA-256 hashes non-stop. That work is essential to the network's security, but once a block is found the results are of no use to anyone. Enormous amounts of power and hardware go in, and the computation itself is thrown away

Proof of Useful Work (PoUW) asks a simple question: if we are going to burn power and GPUs anyway, why not spend them on work someone else also needs, such as AI inference or ZK proofs? The mining race stays, but the puzzle becomes useful computation, so the same work secures the network and leaves something useful behind

The two projects that stand out in the sector are Pearl ($PRL) and Nockchain ($NOCK). Watching Crypto Twitter, there is a sense that the two are being paired the way ZEC and XMR are in privacy. Put very simply, $PRL is the AI Bitcoin and $NOCK the AI Ethereum

Why mining with useful work was hard

Suppose someone says:

“If miners do AI work, why not just reward them with coins in proportion to the work?”

The idea itself is simple. But turning it into a blockchain's mining method makes the problem far more complicated

First, other nodes must be able to check quickly that the miner really did the work. If verifying means other nodes have to redo the same AI computation from scratch, it is no different from doing the work twice, and the point of useful computation is lost

Next, miners must not be able to resubmit results they computed earlier. If a result made yesterday could be pulled out today to collect a reward, no new computation would happen. So the system also has to guarantee that the submitted work was freshly done and tied to the block being built

Miners must also be stopped from choosing problems that suit them. If someone could pick problems they already know the answer to, or ones that are unusually easy, and take the same reward at a fraction of everyone else's cost, the competition wouldn't be fair

In the end, the hardest problem for Proof of Useful Work (PoUW) was not simply finding useful computation, but making that computation something nobody can easily cheat, that stays fair to compete in, and that can be used for the blockchain's consensus

Earlier attempts

There were several earlier attempts to use useful computation for mining. But turning the idea of doing useful computation into a mining structure that actually works was not easy

Project	Idea	Limit or difference
Primecoin (2013)	Used the search for mathematical structures built from prime numbers as mining work	Results could be checked quickly and the network actually ran, but there was no outside demand for what it computed
Gridcoin (2013)	Paid blockchain rewards for science computing done through BOINC	Rewarded useful work but relied on an outside system to verify it, and the network's consensus later moved to proof of stake (PoS)
Ofelimos (2022)	Proposed a PoUW design that combines optimisation problems with block production	An important earlier study that treated PoUW security in theory, but it was never deployed as a live blockchain network

Since then, Aleo, Qubic, Flux and others have tried different approaches. Each had its achievements, but there were limits to handling computation the outside world actually needs while making that very work into a secure and fair mining competition. The hard part was less finding useful computation than reshaping that computation to fit mining

Turning matrix multiplication into mining

Matrix multiplication is a core operation in AI training and inference. From a mining point of view it also has a useful property: the computation is expensive, but checking the result is cheap. With Freivalds' algorithm, proposed in 1977, you can check with high probability whether a result is correct using far less computation and without redoing the whole multiplication

But knowing the right answer is not the same as having computed it fresh for this block. With the same matrices, a result computed yesterday and submitted again today would still pass verification. To use matrix multiplication for mining, a mechanism was needed so that past results alone couldn't claim the reward

The 2025 paper “Proofs of Useful Work from Arbitrary Matrix Multiplication” proposed a way to do this. It perturbs the matrices with a value (a seed) that changes every time, and uses for mining not only the final answer but the intermediate values produced along the way. The design makes it hard for a stored past answer to stand in for new mining work. At the same time, the perturbation can be removed at relatively low cost, so the matrix product that was originally needed is still recovered

In other words, it is a structure in which the same operation can be used for mining while doing computation needed for AI and other work

Pearl is a project built on this research, and the paper's co-authors Omri Weinstein and Ilan Komargodski serve as CEO and CTO of Pearl Research Labs respectively

Pearl: turning AI computation into mining

What Pearl wants to build

Pearl wants to build a network that uses the computation AI already does as mining. It turns matrix multiplication, the operation GPUs perform most in AI training and inference, into PoW, so a single GPU computation earns both AI service revenue and $PRL mining rewards

Pearl calls this 2-for-1. Today an AI company runs GPUs to provide inference and is paid only for that. With Pearl, the same computation also secures the blockchain and mines $PRL. If mining rewards cover part of the cost of AI computation, GPU operators can offer inference more cheaply than before. That is Pearl's core economic model

It is also why Pearl is hard to see as just another GPU-mining coin. The goal is not to burn extra GPU power to make coins but to add mining as a second revenue line to AI work that would be done anyway

How much more does mining add?

Running mining on top of AI inference on the same GPU takes extra work: adding and removing the noise and producing the mining proof

In benchmarks Pearl has published, adding PoUW to inference cost 5.08% on Llama 70B and 3.9% on DeepSeek V3.2. It puts a few percent of extra cost on top of existing AI computation

The protocol uses integer matrices today, but a new scheme that uses FP8-based AI inference for mining is in development

How $PRL is issued

$PRL issuance schedule

• Supply is capped at 2.1 billion $PRL, issued only through mining with no pre-mine and no developer allocation
• Mainnet launched on 27 April 2026, with a target block time of 3 minutes 14 seconds
• Instead of halvings, the reward shrinks a little every block, approaching 2.1 billion over a long period
• At a block height of about 122,100 on 2 October 2026, about 332 million $PRL had been issued, 15.8% of the cap
• About 307 million $PRL will be newly mined over the next 12 months, roughly 93% of what has been mined so far

The first test: Together AI

The first place Pearl's economic model was put into a real service was Together AI

In May 2026 Together AI launched the Gemma-4-31B-it-Pearl inference endpoint with Pearl. It was priced more than 25% below the regular service, and Together said the discount is subsidised by the value generated from $PRL mining. The same forward pass produces both the AI inference result and $PRL, so it was a real product test of Pearl's 2-for-1 model of lowering AI prices with mining rewards. The original Gemma endpoint, however, was removed from serverless on 27 August. Pearl has since said it plans to offer a model using a new floating-point PoUW together with Together

What to watch next is whether the work with Together AI leads to follow-up models and lasting use, and whether the same approach spreads to other providers

Nockchain: one economy for all verifiable computation

What Nockchain wants to build

If Pearl concentrates on one huge compute market, AI inference, Nockchain's goal is wider. It wants something close to a blockchain where verifiable computation is bought and sold. AI inference is computation, and so is generating ZK proofs. Later it wants to add privacy apps and the execution of all kinds of programs written in Nock. Its whitepaper defines Nockchain as "a distributed market for verifiable computation"

In the long run the picture is to turn miners from people who only compute hashes into compute suppliers doing the work the network needs. If Pearl is a PoW specialised for AI computation, Nockchain is closer to a general-purpose compute platform built on PoUW

It started by mining with ZK proofs

From the start, Nockchain used the work of producing STARK proofs of NockVM execution as mining, instead of ordinary hashes. A STARK is a kind of ZK proof: a short proof that lets anyone check a computation was done correctly without redoing it. The network generated more than a billion ZK proofs in its first year

How $NOCK is issued

$NOCK issuance schedule

• Supply is capped at 2³² $NOCK (about 4.29 billion)
• Issuance was steeply front-loaded: about 54.5% was already out on 1 October 2026
• New issuance over the next 12 months is about 330 million $NOCK, around 14% of what has been issued so far
• Since the Aletheia upgrade in May 2026, 80% of each block's mining reward goes to miners and 20% to a protocol fund. No end date is set for the 20% share; the only stated direction is a return to 100% for miners once a useful PoW upgrade arrives

Pearl's AI-PoW and merge mining

The Logos upgrade on 14 August 2026 added an AI puzzle (AI-PoW) to Nockchain's existing ZK mining. Rather than build its own scheme, Nockchain adopted a structure compatible with Pearl's PoUW, which made possible merge mining in which the same GPU matrix multiplication competes for block rewards in both $PRL and $NOCK

At first the target split of blocks between AI and ZK was 30:70, but after the Anthropos upgrade in September it flipped to 70:30. This is not a fixed allocation but a long-run share reached by adjusting each puzzle's difficulty separately, yet since Anthropos around 70% or more of blocks are already coming from the AI puzzle

Logan Allen described this from the start as a strategy to draw in Pearl's compute. One distinction matters here, though. Being able to share the same work with Pearl is a different story from Nockchain having AI demand of its own. Merge mining secured GPU supply quickly, but what matters from here is creating demand from outside users who actually pay to have Nockchain run their computation

Team and how it is organised

Nockchain started at Zorp, the company Logan Allen led. Backed by a $5 million seed round led by Delphi Ventures, Zorp handled the early development of NockVM and Nockchain

When Zorp shut down in July 2026, project assets such as the trademark, domains and GitHub passed to Nock Community Co. Logan Allen has not left Nockchain, though: he keeps building it through his new company, National Compute

Zorp's closure is less a change of leadership than a reorganisation that separates running the protocol from the business side

Demand and roadmap

The role of first customer falls to National Compute, the company Logan Allen set up. But with a company founded by the project's own founder acting as its early customer, how much independent paid demand arrives from outside Nockchain is still unproven. Logan Allen describes the next step as moving "from MAC per sec to verified tokens per sec": measuring not raw compute per second but verified tokens per second. Tokens here are not coins but the word pieces an AI model processes

Further out, the plan is a full ZKVM (an environment where any program's execution can be verified with ZK proofs) and NockApps that can verify and process many kinds of computation, with privacy, private DeFi and post-quantum security added on top. Privacy is planned as a privacy pool app built to the ShieldedCSV specification, not as a feature of L1 transactions

When	Milestone	Status
Aug 2026	AI Compute Network	Done
Q4 2026	Useful inference routing protocol	In progress
Q1 2027	Full Nock ZKVM, ZK verification opcode, forced data availability, bridge withdrawals	Planned
Q2 2027	Privacy pool app	Planned
Q3 2027	Native token standard, private DeFi foundations	Planned
Q4 2027	Post-quantum signatures, ZK compute markets	Planned

What Nockchain has actually proven so far is ZK-PoW and AI merge mining. The bigger thesis, a general-purpose ZKVM, privacy apps and a market for verifiable computation, mostly hangs on the 2027 roadmap

Same sector, different directions

	Pearl ($PRL)	Nockchain ($NOCK)
Market it targets	AI compute	AI, ZK, apps and privacy
How it mines AI work	Its own PoUW	AI-PoW based on Pearl, plus ZK-PoW
What it has shown so far	A commercial test with Together AI	GPU supply gathered quickly through merge mining
Next 12 months of issuance	About 93% of current supply	About 14% of current supply
Open problem	Whether AI demand becomes $PRL demand	Its own compute demand and $NOCK's value capture
Price / market cap / FDV	$1.24 / $412M / $2.6B	$0.020 / $48M / $88M

If $PRL is the thesis closer to reality, $NOCK is closer to a long-term thesis with a bigger picture

What does the community say?

On X, views comparing $PRL and $NOCK split on whether merge mining, which shares the same computation, is a strength or a dependence on another chain. Readings of the valuation gap differ too. What follows are claims by individual writers about the two projects

The Pearl side: simplicity focused on AI compute

Pearl's supporters value a simple structure focused on AI compute over adding more features. brezshares argues that this difference will make Pearl the "AI Bitcoin". Tulip King likewise takes the view that programmability can actually weaken a token's character as money. For them, offering more features doesn't necessarily make a better token

Their criticism of merge mining comes from the same view. brezshares argues that no AI work is done for $NOCK alone, and sees Nockchain's AI mining as an extra reward attached to Pearl's work. The argument is that sharing the same computation doesn't mean the two tokens should hold the same value

The Nock side: merge mining and wider room to expand

Logan Allen, on the other hand, calls $NOCK "private programmable AI Bitcoin" and described merge mining with Pearl as a "vampire attack" to draw in compute early on. From this view merge mining is not mere dependence but a strategy that leads GPUs already mining Pearl to go after $NOCK rewards as well. Nock's supporters value the chance to grow the network without gathering compute supply separately from scratch, and then to expand into AI services, a ZKVM and privacy apps

The question left for both: are real use cases growing?

The sceptics look less at which one is undervalued than at what is actually computed and verified. jkrdoc asked projects selling "verifiable AI", Pearl and Nockchain among them, what exactly gets verified, and noted that Pearl accepts arbitrary matrices rather than only real customer work. The point is that a correct computation and a computation someone needed are different things

What should we watch now?

What to watch in Pearl and Nockchain is less how many GPUs join mining than how many customers and services actually use that computation

For Pearl, the question is whether cases like Together AI, where mining rewards lower inference prices, extend to other models and providers. For Nockchain, it is whether the compute gathered through merge mining connects to its own services and ecosystem. Beyond a single launch or partnership announcement, real paid workloads need to grow and customers need to keep using the services

The meaning of that growth for the tokens matters too. Mining rewards can make a service more price-competitive, but a GPU operator's business doing well doesn't automatically make $PRL and $NOCK more valuable. What matters is whether, beyond drawing customers in with mining rewards, the added usage in turn supports the network and the token economy

Beyond that, whether PoUW can settle into a lasting industry depends on whether the compute gathered for mining turns into services people willingly pay for and keep using.