Who Pays for Proof-of-Work? Pearl Has an Answer

Like many on the timeline, I've become enraptured by Pearl over the last few weeks.

As I argued in This is Not Bitcoin's Cycle, Bitcoin is facing some uncomfortable questions, and alternatives are increasingly getting a hearing. Zcash has drawn renewed interest for its privacy features and efforts to address the quantum threat. Pearl is emerging as a response to a different problem: Bitcoin's increasingly strained mining economics.

That's the lens through which Pearl makes the most sense to me. Behind the talk of "AI Bitcoin" and matrix multiplications is a relatively simple idea: what if the work securing a blockchain could also earn money doing something else? And could that ultimately produce better money?

1. What Is Pearl, and Why Does It Matter?

Pearl (PRL) is a fork of Bitcoin that launched in April with an ambitious goal: letting the work that secures its network also serve paying AI customers.

That work is matrix multiplications, the core math behind AI inference (running a trained model to answer requests). It calls this Proof of Useful Work, or "duplex work," since completing these "matmuls" also enters miners into a lottery for block rewards and the privilege of adding the next block. The more calculations they complete, the more tickets they receive.

Miners don't have to serve AI customers. They can perform the same calculations on random inputs and still earn PRL. The innovation is that useful AI work can double as mining, meaning miners could in theory offset their mining costs.

To see what Pearl is responding to, consider the economics of Bitcoin mining.

Bitcoin miners have always needed to sell some of the coins they earn to cover operating costs. But as mining rewards shrink due to halvings, and transaction fees fail to make up the difference, that pressure can increase. In Q1 2026, publicly traded miners sold a record 32,000-plus BTC, partly to cover costs and finance new infrastructure.

Rising Bitcoin prices or greater mining efficiency can offset these pressures, but neither is guaranteed, particularly when miners need it most.

Providing AI computing infrastructure, on the other hand, offers the prospect of steadier revenue amid growing demand. That's why we've seen miners increasingly pivot toward AI, which CoinShares estimates can generate three times as much annual profit per megawatt as Bitcoin mining.

This has consequences for Bitcoin's security, which depends on the computing power miners contribute to the network. The more of it securing the network, the more expensive it generally becomes to compromise. Miners are incentivized to contribute that power by Bitcoin's security budget, the new coins and transaction fees they earn. If enough computing power leaves the network without being replaced, attacks could become cheaper, particularly as mining rewards continue shrinking.

Pearl tries to turn this competition into an advantage. An AI provider already earning money serving customers could mine PRL using those same GPUs. AI revenue helps pay for the machines; PRL rewards become extra income. That could reduce the need to sell newly mined coins while opening mining to a much larger pool of hardware than Bitcoin's specialized ASICs.

There's a potential harder-money argument here, too. Less selling pressure could help, while greater demand to hold PRL could raise the dollar value of its mining rewards and, with it, its security budget. But more GPUs don't automatically mean more economic security, or more buyers.

— (@)

That's the case for Pearl as potentially better money.

2. The Marketplace Is Secondary (For Now)

Money is only half the pitch. The other half, and the one many point to for PRL's value, is a future where PRL becomes a currency to pay for AI inference.

This future is spelled out in Pearl's whitepaper, with a compute marketplace that could use PRL for payments, though it has no announced launch date. For now, Pearl serves inference through prepaid credits, not PRL, with launch prices up to roughly 75% below a major provider's standard rates, potentially subsidized by mining rewards.

To me, competing on price alone doesn't make Pearl's inference particularly interesting. Other onchain inference markets already offer steep discounts, with Surplus advertising average savings of nearly 90%.

What could set Pearl apart is inference that's not just cheaper, but also private and verifiable, two qualities that could give crypto-enabled inference a meaningful advantage over traditional providers. Privacy is already a selling point for providers like Venice and NEAR AI. Verifiability is harder: cryptographically proving a provider actually ran the model it promised.

Pearl is working on it through a partnership with Attestable on a "proof of inference" layer, though it's still in development. If it works, and Pearl can pair it with privacy, it could offer all three qualities at once.

That's when the PRL marketplace becomes interesting to me. But even superior inference doesn't guarantee lasting demand for PRL as a payment token. I'm skeptical of medium-of-exchange tokens in general. I don't think they've really worked, and I think a token needs to establish itself as money before payment utility matters.

For now, the mining thesis is more compelling to me: it depends on customers paying for inference, not necessarily paying in PRL.

3. The Risks

Pearl is young, and its design raises questions about its economics, technology and decentralization.

Many of these concerns have worthwhile counterarguments, particularly from Pearl investor Brian Breslow of No Limit Holdings, who has written extensively about the project. Below are the concerns that stand out to me most, alongside the responses I've found and one of my own ideas.

Inflation and "junk mining." Pearl is issuing a lot of new coins, with inflation still very high in its first year. But that's not a fixed annual rate. Pearl's issuance follows a smooth, polynomial decay curve rather than Bitcoin's four-year halvings. Like Bitcoin, about half its supply is mined in the first four years, so inflation falls sharply each year, to roughly 30% by its third year and about 11% by its fifth. Still, many miners aren't performing useful AI work at all. Researchers found that the dominant mining software they studied wasn't running inference, meaning miners were simply performing calculations to earn PRL, with operating costs they may need to cover by selling those coins.

• Counterargument: Brez argues that this is by design. Miners don't have to perform useful work, but those serving paying AI customers should eventually have a cost advantage over those who aren't. He also points out that not every newly mined coin gets sold. Still, it's unclear whether there's enough demand for both inference and PRL itself to support the economics.

The technology is still evolving. Pearl underwent three rapid forks to patch mining exploits this summer, quite the messy start. Its proposed FP8 upgrade could make mining alongside real AI inference more efficient, but would also exclude older Nvidia, AMD and Intel GPUs, shrinking the pool of people who could mine.

• Counterargument: Pearl argues that restricting eligible hardware is necessary to make the AI calculations reliably verifiable. It's a trade-off between making mining more useful and keeping it accessible to as many people as possible. Whether that trade-off is worthwhile remains to be seen.

— (@)

Mining power is already concentrated. Secondary reporting estimates that the two largest mining pools, Kryptex and PearlHash, recently found roughly 75–80% of blocks. Even if thousands of people own the GPUs, a handful of pool operators can still coordinate much of the network's mining power.

• Counterargument: Then again, miners aren't permanently tied to their pools and can redirect their computing power elsewhere. That limits a pool operator's control, but doesn't eliminate the risk. I haven't found a particularly satisfying answer to this concentration yet.

Could AI companies become the new mining giants? This is the longer-term concern I find most interesting. If companies already serving paying AI customers have the greatest mining advantage, won't Pearl eventually concentrate mining among large data centers? That seems at odds with the decentralized ethos of crypto.

• Counterargument: Brez sees bringing the AI industry's computing power into consensus as the point. But I wonder whether decentralized inference networks could offer another path. If their independent GPU operators began mining Pearl alongside serving paying customers, that could help distribute mining power across more participants rather than concentrating it among a few large providers. Of course, this depends on sufficient inference demand, compatible hardware and independent control over mining. Still, it's a possibility I haven't seen discussed.

Bitcoin's challengers have room to gain ground as its unresolved problems become harder to ignore. Pearl fits that broader picture, responding not to Bitcoin's privacy or quantum limitations, but to the economic incentives that keep its miners securing the network.

What excites me isn't simply that Pearl uses AI, or that PRL might one day pay for inference. It's that securing a blockchain could become a byproduct of work someone was already willing to pay for. That could reduce miners' need to sell, recruit more hardware and, if demand for PRL grows too, support a stronger monetary network.

None of that makes Pearl better money yet. But it offers a different answer to an increasingly important question: how do you sustain the economics of securing a proof-of-work network as its mining rewards shrink and the world finds more profitable uses for computing infrastructure?

Thanks to Brian Breslow for his feedback on this piece.