Skip to content

Pearl - Proof of Useful Work for AI Compute

Jun 10, 20263 min

AI compute markets have a simple coordination problem: GPU providers spend real energy producing useful work, while proof-of-work blockchains spend real energy producing security. Pearl Research tries to merge the two.

Pearl is a Bitcoin-style Layer 1 where mining comes from AI computation, not hash puzzles. Its chain is live, PRL has a fixed 2.1B supply, and blocks target 194 seconds.

In this edition, we look at Proof of Useful Work, how PRL rewards are earned, and whether the token design actually maps to AI compute demand.

What Pearl Is Building

Pearl is a Bitcoin-style L1 that replaces SHA-256 hashing with matrix multiplication as its proof of work function, using block rewards, transaction fees, and a longest-chain rule. Bitcoin miners race to find a hash below a target value while Pearl miners race to produce valid matrix multiply proofs against committed input matrices. The chain has a fixed 2.1B PRL supply and targets 194-second blocks.

Matrix multiplication is the core arithmetic behind AI inference and training. When a neural network generates a response, it repeatedly multiplies large arrays of numbers through its layers, and every forward pass is a sequence of matrix multiplies. Pearl’s whitepaper argues that a GPU running those operations for a live AI workload can simultaneously produce valid PoW submissions as a side effect of computation already underway.

A miner commits to specific input matrices before computing, which prevents using trivial or pre-computed inputs that would give an unfair advantage. The commitment forces the miner to run the actual computation, where the output is then verifiable by other nodes without those nodes re-running the full workload.

How PRL Rewards Are Earned

PRL rewards work like Bitcoin block rewards. Miners receive newly minted PRL and transaction fees when they produce valid blocks. A miner producing a valid block means committing to input matrices, running the computation, and generating a verifiable proof. Rewards are paid per block, not per inference request served.

The cost argument favors providers already running inference. A data center serving real AI requests pays lower incremental mining costs because the matrix multiplications are happening anyway. Pearl mining adds only the overhead of proof generation on top of the existing workload. A miner with no AI customers pays the full GPU cost to mine, making that miner more expensive than one with paying inference clients.

Pearl’s emission curve models Bitcoin’s scarcity design without copying the halving structure. The 2.1B supply decays on a smooth polynomial curve that allocates roughly half over four years, avoiding the abrupt reward cliffs that create short-term instability around Bitcoin halvings. Gradual decay keeps miner incentives more predictable than discrete step-downs, at the cost of slightly less aggressive near-term scarcity.

The Together AI Integration and the Open Loop

Together AI launched an exclusive Pearl-powered inference endpoint for Gemma-4-31B-it-Pearl in May 2026, priced at a 25%+ discount to standard rates and offset by expected PRL emissions. Together’s announcement noted that future Pearl-powered products may let inference customers receive part of those emissions directly. This is the first integration where real compute buyers receive a price reduction because their provider’s GPU work generates PRL.

Pearl does not yet have a marketplace where buyers pay PRL and receive inference. Rewards flow to miners for valid blocks, the mechanism does not distinguish between a miner serving real inference clients and one running matrix multiplications with no AI buyers. Pearl’s documentation frames the marketplace layer as a future direction, with compute contracts eventually settling natively onchain.

Together AI’s 25%+ discount is sustained by PRL emissions expected from its GPU workloads. If that discount holds as the emission curve decays and more providers integrate Pearl mining for similar economics, PRL drastically subsidizes compute. If adoption stays concentrated on one commercial partner while the broader mining base operates without real inference clients, PRL remains a speculative token.

Important Links