The Pearl public chain uses GPU matrix computation to balance AI inference and mining. The PRL token has a fair launch, and inference service providers have a greater economic advantage than pure miners, potentially reshaping the future of proof of work.
Written by: @aliomerhorzum
Compiled by: AididiaoJP, Foresight News
Editor's note: Pearl is an L1 public blockchain project based on "Proof-of-Useful-Work" (PoUW), founded by Omri Weinstein, a mathematics PhD from Princeton University and a professor of computational complexity theory at Columbia University. Its core idea is to have miners use GPUs to perform matrix multiplication—the underlying operation of large language model inference—rather than traditional hash computation, thereby attempting to make mining compute simultaneously serve AI inference and blockchain network security. Pearl's mainnet launched in April 2026, with a maximum token supply of 2.1 billion PRL, adopting a gradually decreasing issuance curve to replace Bitcoin's four-year halving mechanism. The project claims a completely fair launch, with no pre-mine, no founder allocation, and no VC rounds. As of October 2026, PRL's circulating supply ratio is approximately 16%, with an FDV of approximately $2.8 to $4 billion. Pearl has partnered with Together AI to launch a Gemma inference API with a discount of more than 25%, and has joined NVIDIA's Inception accelerator program. Its core controversy lies in the fact that the on-chain consensus protocol cannot distinguish whether miners are computing real AI inference tasks or random matrices, an issue that has already sparked empirical academic research.
One of the most common misconceptions about Pearl is that its "Proof-of-Useful-Work" mechanism guarantees that every miner is performing useful AI inference. This is not the case. And I believe it should not be.
One GPU performs matrix multiplication for real inference requests, while another GPU performs random matrix multiplication for mining. To Pearl, both produce equally valid work. The protocol does not know whether anyone is paying for these computations. This is not a flaw; it is by design. The distinction between so-called useful mining and useless mining should ultimately be determined by economics, not by the protocol.
What exactly does the 5% overhead mean?
Pearl uses matrix multiplication as the basis of its proof of work, and matrix multiplication is one of the most core operations in modern AI inference. In this way, the same GPU can produce work that helps secure the monetary network while performing inference for customers. But making computation verifiable is not free.
Pearl needs to additionally perform randomization, cryptographic commitments, and other operations to construct and verify computation proofs. These operations consume resources, and this is the mining overhead.
When Pearl says its goal is to push overhead below 5%, it means: the additional cost of turning inference computation into mineable work should be sufficiently small compared with doing inference alone. It does not mean that only 5% of the computation is useful. The underlying matrix multiplication can serve both purposes at the same time.
But low overhead does not mean that an inference GPU can produce as much mining compute as a dedicated miner. These are two different things.
Why inference GPUs cannot outmine professional miners
Because a GPU serving real inference does not perform large-scale matrix multiplication at maximum throughput every second. The computational characteristics of inference workloads are completely different. During the prefill stage, the model processes the input prompt, and the GPU can execute a large amount of matrix multiplication; during the decode stage, the model generates tokens one by one, and performance is often limited by memory bandwidth rather than raw compute.
Dedicated miners can keep GPUs continuously performing matrix multiplication. Inference providers cannot do this because they must serve customers, meet latency requirements, and many workloads simply cannot saturate GPU compute.
So two questions must be viewed separately: how much slower does generating proofs for matrix multiplication make inference? And how much mineable matrix multiplication can an inference GPU actually complete? These are not the same thing.
Lowering overhead will make simultaneous inference and mining more attractive; improving GPU utilization, batching, and workload scheduling will increase the amount of mineable computation providers can produce. Both matter. And I believe the latter holds the biggest opportunity.
Why pure miners will eventually fall behind
Imagine two operators with comparable hardware. One is a dedicated miner that earns PRL solely by performing matrix multiplication to secure the network. The other is an inference provider that earns revenue from customers using its GPUs and also earns PRL from eligible computation within those requests.
Both are competing for mining rewards, but only the latter has a second revenue stream. This difference changes the economics.
Dedicated miners must cover hardware, electricity, and operating costs solely through mining rewards. Inference providers can cover part or even all of their costs through their core business, making mining incremental revenue. As competition intensifies and mining margins are compressed, operators relying only on mining rewards will come under greater pressure than competitors with the same hardware but also inference revenue. Miners with two revenue streams survive; miners with only one do not.
This does not mean pure miners will disappear overnight. They may have cheaper electricity, more efficient hardware, lower capital costs, or other advantages. But Pearl creates a structural economic advantage for productive compute that Bitcoin never had. As more inference providers join, this advantage will continue to amplify.
How much global compute does Pearl actually need?
This is the question I find most interesting. It is easy to assume that Pearl must first connect to most of the world's GPUs before useful inference can become the main source of network security. I do not think that is necessarily true.
The key is not just how many GPUs are connected to Pearl, but how much effective mineable matrix multiplication throughput they can contribute. A GPU in service may contribute far less mining compute than a dedicated mining rig because of differences in workload, batching, and utilization. But the AI inference industry is enormous and still growing. If a sufficiently large portion of that productive compute contributes even part of its available throughput to Pearl, the resulting mining compute could be considerable.
More importantly, unlike pure miners, these operators do not necessarily depend on mining rewards to survive. This shift does not require a protocol-level ban; it only requires productive compute to become sufficiently competitive in the mining market. Ultimately, economics can accomplish what protocol rules never needed to do.
The real opportunity: letting inference and mining reinforce each other
Mining overhead must come down, inference performance must remain competitive, and batching, routing, and GPU utilization must improve. Providers also need reliable infrastructure that allows them to participate in Pearl without harming their existing business.
These will not happen automatically. But I believe this is where Pearl's most important competitive advantage will emerge. Not in persuading cloud providers to become miners, but in making mining a natural extension of their existing inference business.
Consider the scale of what is happening. The world is spending trillions of dollars building infrastructure to produce intelligence, and millions of GPUs will compete to provide inference services as efficiently as possible. Every percentage point of additional revenue will matter. What Pearl can give these providers is something traditional mining networks cannot: earning mining rewards from computation they are already paid for.
As more inference providers join, the network will become more secure, mining competition will intensify, and the economics for pure miners will become increasingly difficult. At some point, the machines securing Pearl may overwhelmingly not have been built to mine Pearl in the first place.
Pearl does not need the world to build more machines to secure a currency. It needs the machines the world is already building.
We need useless miners until we no longer need them. I believe the shift from dedicated mining to economically productive computation will be one of the most important developments in the history of proof of work.