Pearl blockchain integrates AI inference with mining rewards through Proof-of-Useful-Work

Most blockchain mining burns electricity to solve puzzles that exist solely to prove you burned electricity. Pearl, a Layer-1 blockchain built by Pearl Research Labs, decided that was a waste of perfectly good GPUs and built something different: a network where the math miners do to earn block rewards is the same math that powers AI inference.

The project’s Proof-of-Useful-Work (PoUW) consensus mechanism uses matrix multiplications, the foundational operations behind neural network inference, as its proof-of-work function. Miners run these computations on high-powered NVIDIA GPUs, simultaneously securing the network and producing real AI outputs. Pearl calls it a “2-for-1” model, and the math checks out: the overhead sits at roughly 1+o(1), which in plain terms means the extra computational cost of doing both jobs at once is negligible.

How the dual-purpose engine works

When a miner on Pearl’s network processes a block, they’re performing matrix operations that can serve as part of an AI inference pipeline. The results are verifiable on-chain, meaning the network can confirm the work was done correctly without re-running the entire computation.

The network launched its mainnet on April 27, 2026, with a target block interval of approximately 194 seconds. Early mining infrastructure supports both pool and solo mining, all optimized specifically for NVIDIA GPUs.

Pearl’s native token, PRL, has a total supply cap of 2.1 billion. Block rewards consist of newly minted PRL plus transaction fees, with issuance following a smooth polynomial decline over time.

Together AI partnership and the discount inference play

Pearl’s most consequential partnership so far is with Together AI, which launched a discounted inference endpoint on the network priced more than 25% below standard market rates. The discount is subsidized by PRL emissions, meaning miners are effectively earning crypto rewards for providing compute that Together AI can sell to customers at a steep markdown.

On the technical development front, Pearl published an FP8 floating-point scheme on September 14, 2026, with FP4 schemes also under discussion. Running models in FP8 instead of FP16 or FP32 roughly doubles throughput on compatible hardware, which matters enormously when your consensus mechanism is literally running neural network operations.

Market performance and the road ahead

The PRL token has responded to these developments with considerable enthusiasm. In late September 2026, PRL surged to a new all-time high above $1.60 following a 67% rally in a single 24-hour period.

There’s a meaningful caveat worth noting, though. Empirical analysis of the network has shown that much of the computational power was being used for unpaid workloads rather than revenue-generating paid AI inference services. In other words, miners were doing useful work in the technical sense, performing real matrix multiplications, but much of that work wasn’t yet connected to paying customers.