Inside Pearl, the chain that mines with AI math

Pearl is one of the more interesting things I've looked at in crypto this year.

It's a blockchain mined on GPUs, where the mining is supposed to double as AI inference. Instead of Bitcoin's hash puzzle, Pearl's miners compete by doing matrix multiplication, the math behind most of the work when an AI model answers a question. In principle, one computation can secure the chain and serve a customer at the same time. I wrote about that idea, and why it's so hard to pull off, in an earlier post.

Apart from the mining, it's built a lot like Bitcoin. Its node software is a fork of btcd, an open-source Bitcoin implementation written in Go.

And it isn't just a whitepaper. Pearl launched on April 27, 2026, and by October 9 its explorer showed more than 125,000 blocks. Lately it's been producing about 490 blocks a day, roughly one every three minutes, each paying the winner about 2,270 of Pearl's coin, PRL. That's more than a million new PRL a day.

The idea behind Pearl, called proof of useful work, has been around for more than a decade. The goal is to make the computation that secures a blockchain also produce something people want. Earlier attempts kept hitting the same wall, because genuinely useful work is hard to turn into a fair, verifiable lottery.

Pearl is the most serious attempt I've seen to get past that wall, and it's using the one workload the world currently can't get enough of: AI inference. Its design comes from an academic paper co-written by Omri Weinstein, now CEO of Pearl Research, the company behind Pearl.

The more I dug into Pearl, the more impressed I was by the engineering, and the more questions I had about the "useful" part. That's not because anything is broken. The things you'd need to see to judge it are hard to find: who's mining, who's paying, and how much of the mining served anyone.

And the one clue the chain does offer, the shape of the work that wins each block, has moved sharply away from Pearl's public AI models. Since September, none of the 443 winning blocks I sampled had the shape of work from those models.

Everything below comes from what Pearl and others have published: its code, documentation, papers, Discord server, and block explorers. I didn't run a miner or pay for inference.

1. How does Pearl turn AI math into mining? ↓The mechanism
2. Which AI work can mine, and what does the chain show? ↓The evidence
3. Who's mining, and who's paying? ↓The network
4. Why does random math win? ↓The incentives

How Pearl turns AI math into mining

Bitcoin miners hash a block's data over and over until someone finds a result below the network's target. Pearl keeps this lottery aspect, but changes what generates the tickets:

1. The miner takes a fingerprint, or hash, of the two grids of numbers it's about to multiply, along with the block it's working on.
2. That fingerprint decides some random noise, which gets mixed into the grids. Together with step 3, it's designed to stop a miner from picking numbers that make the work easier, or reusing work from one block on another.
3. As the GPU works through the multiplication, it hashes the partial results along the way. If one falls below the network's target, that's a winning ticket.
4. The noise is stripped back out cheaply, leaving the real product for whatever needed it.

Checking works differently, too. Instead of having other nodes redo the multiplication, which would mean sharing the inputs, the winner builds a zero-knowledge proof: a compact certificate showing the committed work was done and produced a qualifying ticket, without revealing the grids. (Pearl's are built with Plonky2.) Nodes just check the certificate, so a customer's prompt or a company's model weights never need to go on-chain.

Where the AI comes in

The mining is only truly "useful" if it happens inside real inference. Pearl publishes a plugin for vLLM, a popular open-source tool for serving language models, that swaps some of the model's matrix multiplications for Pearl's noisy version. The model gets the result it needs, and any winning tickets are submitted in the background. A model has to be converted into Pearl's format first, and Pearl publishes four converted models on Hugging Face.

This is the part I find most convincing. The mining isn't a separate program running next to the model. It happens inside the model's own math, and Pearl Research reports about 4 to 5 percent overhead in its own production measurements.

What the network does, and what it doesn't

When I started reading about Pearl, I pictured customers sending AI requests to a decentralized network, which handed them out to miners. That isn't how it works, and the rest of this post makes more sense once that picture is gone.

• The chain only knows about PRL transactions and blocks. It has no idea inference exists: no requests, no customers, no inference payments. For the mining work, it checks that each winning block carries a valid certificate.
• Inference is a separate business that runs off the chain. Anyone can run one. If they serve a Pearl-converted model with Pearl's plugin and run a Pearl node, their inference work can produce mining tickets too.
• Inference businesses collect their own payments. Together AI, a large inference provider that partnered with Pearl, bills its customers. So does Pearl Research, which sells inference through its own paid API and bills those customers in dollars. The Pearl Network plays no direct part in either.
• Block rewards are the only thing the network pays. Whoever wins a block, or their pool, gets the PRL, whether or not an inference customer was involved. A miner only has to find a winning ticket, not serve inference. What's new about Pearl is that it's possible to find that ticket while also serving inference, with the same math.

With all the references to Pearl, it's easy to get confused. The name gets used for three different things.

• Pearl Network is the open blockchain. Anyone can run a node or mine, and the chain only keeps track of PRL and blocks.
• Pearl Research is the company that built the network. It also sells AI inference through a paid API, with accounts and dollar billing, and that service is separate from the network. Its GPUs can mine on Pearl while they serve requests, but nothing requires them to, and there's no public way to check how much of that inference actually mines.
• PRL, the network's coin, often just gets called Pearl too. Much like bitcoin, you send it with a key you made yourself, any node will pass it along, and anyone can check the result on the chain. Sending it has nothing to do with AI.

AI inference only touches Pearl Network through mining, and even there it's optional.

What AI work earns mining tickets today

This is where Pearl is strongest. Turning AI math into a fair mining lottery is the hard technical problem, and Pearl has working code for it. But the network's current rules, and Pearl's own software, narrow which AI work can actually mine today.

• Integer math only. Mainnet only accepts proofs built on low-precision integer math, while a lot of newer serving runs in FP8 floating point on the latest GPUs. Pearl has published a design for FP8 mining, but it's still a draft and its main implementation hasn't been merged.
• Mixture of experts, switched off. Many of today's largest open models are mixture of experts (MoE) designs, which send each piece of input to a few specialized sub-networks instead of running the whole model. Five of the seven models in Pearl Research's own catalog are built this way, including the two its docs suggest customers start with. Since July, mainnet has rejected mining proofs from their expert layers, though their other layers can still mine. Of the seven catalog models, only Gemma 4 31B, a dense model, has a public mining-ready version. The company may run private ones for the others, but I can't tell from the outside.
• Only part of each model. Even in a model that can mine, not every multiplication does. In Pearl Research's version of Gemma, about 54% of the weights are in the mining format, mostly in the feed-forward layers.
• Big batches only. By default, the plugin only mines on multiplications covering at least 1,024 tokens at once. That happens when a model reads a long prompt or serves many requests together. Writing replies word by word for a handful of users doesn't mine.

The MoE switch-off has its own story. In July, miners found a missing check in the verifier that let a specially built miner claim far more computation than it actually did. Pearl Research's post-mortem on its Discord server says the flawed proofs mined about 4.8 million PRL and made up 28% of new blocks while the exploit ran, peaking at 73% on July 23, the day Pearl shut MoE mining off. The company plans to bring it back after a protocol revision, adversarial testing, and an outside review. The math in the paper can be solid, and the code running the network can still have a hole in it.

What the chain shows

Nobody publishes the share of Pearl's mining that served paying customers, and blocks don't say whether a customer was involved. But the chain does give one clue, and it's easy to check.

Every winning block's certificate records the size of the grids that won it. A real model's layers have specific widths: one layer of Pearl's Llama 3.3 70B multiplies a grid 8,192 numbers wide by one 57,344 wide. Random grids can be any size, and miners tend to pick big, round ones, like 131,072. pearlchain.live, an independent explorer, decodes these sizes. I pulled every 50th block since launch, 2,580 in all, and checked each one against every layer of Pearl's four public mining-ready models, including the sizes those layers take when a model is split across several GPUs.

At launch, nearly every sampled block had the shape of one of Pearl's Llama models. That faded fast, to a little over half in May, about 6% in June, and around 4% in July and August. The last sampled match came on August 14. In September and October, none of the 443 sampled blocks matched, and pearlchain.live labels about three-quarters of them "synthetic benchmark shapes." Only one block in the whole sample matched Gemma, on May 8, and its widths also fit Google's older Gemma 3.

Model-shaped doesn't mean customer work, either. In launch week, a member of Pearl Research's team told miners on Discord that they could "bring your own inference demand or run an off the shelf benchmark." About 95% of sampled launch-week blocks used the same batch size, 32,768 tokens. That's what a fixed benchmark would produce, though it doesn't prove one.

This check has limits. A match doesn't prove a model ran, and a miss doesn't prove the work was synthetic, since a private model could have widths I didn't check. Winning blocks also aren't a perfect sample of the network's work. But they're the best public sample there is, and the trend is hard to miss.

Who's mining

So who's doing the winning work? Pearl's explorer doesn't say much about who's behind each block, and third-party node counts show addresses, not distinct miners or GPUs.

The best early look came from an independent researcher, Abhinaba Basu. In late May, he analyzed 8,012 online workers belonging to the 15 largest miners on AlphaPool, then a pool with about a fifth of the network's mining power. He classified all of them as having hardware capable of running AI models, but the most common mining program among them appeared to have no code for running AI models at all. It generated its grids from random seeds.

Today's picture differs in the details but not in direction. When I checked pearlchain.live on October 17, one pool, Kryptex, had won about half of the previous day's blocks, and the next, PearlHash, about a fifth. Kryptex's Pearl page reads like a crypto-mining service, not an AI company. It recommends third-party mining programs for Nvidia and AMD gaming cards and offers to pay miners in Bitcoin, USDT, or gift cards. AMD cards can't run Pearl's official inference miner at all. None of that proves Kryptex miners never serve AI, but it isn't what an inference network looks like. It also means about half of that day's blocks were attributed to a single pool, and from September 17 to October 17, just 35 addresses received block rewards.

Pearl Research is also in the mining business itself. Its Pearl Compute platform sells fixed-term mining plans for regular currency. The company runs cloud GPUs for buyers, pools their computing power, and distributes the PRL rewards to their plans each day. The marketing page says those GPUs "perform real AI workloads while simultaneously mining," but the terms describe a mining service and don't mention AI workloads. I couldn't find how much of the network Pearl Compute accounts for, or which addresses it mines to.

Who's paying

Pearl Research runs its own inference business. Customers prepay in dollars or get invoiced if they're larger, new accounts start with a $5 credit, and its public pricing page lists seven open models, from Google's Gemma to DeepSeek, at what it calls special launch discounts.

Its docs blur the line, saying models are "served on the Pearl network's distributed GPU capacity rather than a single provider's cluster." Really, requests go to Pearl Research's own API, and the docs don't say whose GPUs answer them, or whether they're mining. There's reason to think not every model does. Six of the seven catalog models have no public mining-ready version, and five are MoE. The docs' glossary mentions FP8 and INT8 serving, but doesn't say which models run at which precision, or with mining on. Mainnet doesn't accept FP8 for mining. Pearl Research's homepage doesn't mention mining or a blockchain at all.

Pearl Research's named inference partner is Together AI. In May, Together launched a Pearl-powered version of Google's Gemma at a discount of more than 25%, paid for by expected crypto rewards. That's a big part of the pitch: block rewards paying down the price of inference. Together removed the model from its serverless service on August 27, and I couldn't find a reason given. That doesn't seem to be the end of the partnership, though. In September, Pearl Research still described future Together endpoints.

Some of the demand Pearl Research did point to came from inside its own ecosystem. In May, it told Pearl Compute plan holders on Discord that they'd get free access to its Gemma and Llama 70B endpoints, and that tokens generated there would "contribute rewards" to their mining. The team also said, in May, that its Discord help bot ran on Gemma and fed the same rewards pool. In that sense, the inference was real, but no outside customer was paying for it.

If you own GPUs, the public instructions show you how to mine, not how to get paid for serving customers, and I couldn't find a way to sign up to serve Pearl Research's API traffic. The closest thing to an answer was unofficial. A community moderator on Pearl's Discord described Together as serving its own customers on its own machines, with handing jobs to outside miners as a future step. And in September, the developer of a popular mining program said there still wasn't a way to rent GPUs to an AI service while mining. So an independent miner today has three options: mine random grids, run a model on prompts they write themselves, or start an inference business and find their own customers.

Pearl Research doesn't publish customer counts, request volumes, or revenue. That's normal for a private company, but it leaves the demand side almost entirely out of view.

Why random math wins

If you look at the mining plugin and what it actually takes in, it's only the grids of numbers and the settings for multiplying them. It doesn't take a customer, a payment, a signed request, or even the name of a model. Researchers call this the usefulness gap. Pearl's own node software includes a miner that mines on random grids, and Basu showed a pool would accept that work as valid.

It's worse than the useful work merely being optional. Tickets come from the eligible multiplication a GPU does in its mining layers, not from tokens or answers. So a dense model like Gemma should earn noticeably more tickets per GPU-hour than an MoE model, and a busy server more than a quiet one. Follow that logic to the end, and the most ticket-efficient model is no model at all. Random grids can use ideal shapes, run the GPU flat out, and skip everything real inference carries with it: the layers that don't mine, the small batches, the time spent waiting for customers.

Pearl's answer is economic, and it depends on who's mining.

• A pure miner has no customers and earns only PRL, so random grids are the obvious choice. But it pays the full cost of the GPU, about $1.70 an H100-hour in the whitepaper's example.
• An inference provider already pays for its GPUs through its customers. Mining only costs it the overhead, about a tenth of that, and it's a bonus on work it's doing anyway.

As more useful work joins and rewards per GPU-hour fall, pure miners should be the first who can't cover their costs. An economics paper written while its author was consulting for Pearl Research models this and argues it can hold up.

The logic works on paper. What's hard is how far away that point is. Here's a rough sense of the scale today, using the whitepaper's illustrative cost.

• Pearl pays out about 1.1 million PRL a day, roughly $2 million at mid-October prices.
• At $1.70 an H100-hour, that pays for about 50,000 GPUs mining full-time.
• In this rough scenario, random-grid miners get squeezed out only when real inference alone pushes rewards below that cost. If a real model earns about half as many tickets per GPU-hour (my assumption, not a measurement), that's something like 100,000 data-center GPUs serving customers with mining turned on, around the clock.
• Many random-grid miners own gaming cards, where the main ongoing cost is electricity, so the real bar is likely higher still.

The block shapes suggest the network is nowhere close. And even if it got there, the incentive wouldn't be a rule. Synthetic mining would be unprofitable, not invalid, and a GPU sitting idle would often still be most profitably filled with random grids. Providers would also earn more serving dense models than MoE ones, so they could discount those more.

To be fair to Pearl, it hasn't hidden any of this. Its whitepaper predicted that early on, few companies with useful work would be set up to mine, and that expectations about the future would carry the token price in the meantime. It also describes a future marketplace that could verify on-chain that computation happened "on account of Alice asking Bob." And some synthetic work is unavoidable, since customer demand comes in bursts and blocks can't wait.

The proof doesn't tell customers much, either. It shows that some committed matrix work happened, not which model ran, that the whole answer was computed correctly, or that it reached anyone. Pearl Research has announced a partnership with a company called Attestable to build proofs of which model ran, but it's still a proposal. For now, you're trusting Pearl Research's service the way you'd trust any other API.

What would show the mining is useful

The evidence so far points one way, but none of it settles whether Pearl's mining is useful. Some of what would settle it, anyone can check. Some of it, only Pearl Research can show. And some of it, nobody can prove yet.

• What anyone can check. The shape check above is the cheap version. A stronger one would match the weight fingerprint each certificate publishes against a published model's weights. That would show those weights were used in the winning multiplication, though not that a full model ran or that a customer asked. A small first attempt found no match, and as far as I know, nobody has done it in full.
• What only Pearl Research can show. Which catalog models run with mining turned on, what share of blocks its own serving fleet wins, how much Pearl Compute mines and on what, and how much of its inference revenue is left after free credits and PRL-funded discounts. Those numbers would put a floor under how much of the network's work served real customers.
• What nobody can prove yet. Whether an independent customer wanted any particular piece of work. Nothing ties a mining proof to a request today. A protocol change, like the marketplace Pearl's whitepaper describes, could show who requested and paid for a computation, but a miner could still pay itself. Independent demand needs evidence beyond the proof.

Pearl is still one of the more interesting things I've looked at this year. It took an idea that's been stuck for more than a decade, made a credible case that work useful to AI can pass the tests a proof of work has to pass, and shipped it as a live chain with code anyone can read. That's not a small thing.

So here's my takeaway. AI math can secure a chain. Pearl has shown this. But for now, the work winning the blocks I sampled looks a lot more like random math than AI, and the rewards still favor random grids over real models. The open question is whether real inference can grow to do most of that work, and whether the incentives can be made to reward it.

Sources