Pearl Links AI Inference to Mining Through Proof-of-Useful-Work
The Layer-1 network uses matrix multiplication for mining and AI workloads, while a Together AI endpoint applies PRL emissions to discount inference costs by more than 25%.
Pearl is building a blockchain around a simple proposition: the computations used to secure a network can also support artificial intelligence.
The Layer-1 network uses Proof-of-Useful-Work, a consensus design based on matrix multiplication rather than conventional hashing. Matrix multiplication is a fundamental operation in neural-network workloads, allowing GPU operators to use the same broad class of calculations for mining and AI inference.
The project’s commercial model received a boost in May when Together AI introduced an inference endpoint for Gemma-4-31B-it-pearl. The service was offered at more than 25% below its standard price, with the discount offset by PRL emissions generated alongside the inference workload.
That model is designed to give GPU operators two potential sources of value: blockchain rewards and payments linked to AI computation. Pearl measured overhead of 5.08% on Llama 70B using four H200 GPUs and 3.9% on DeepSeek V3.2 using eight H200 GPUs.
But the central question is whether Pearl’s mining activity is performing customer-requested AI work or merely running calculations that resemble the mathematics used in AI.
A June 2026 preprint by Abhinaba Basu analyzed 8,012 workers and reported a network capacity of 24 EH/s, equivalent to about 320,000 GPU equivalents and an estimated 112 megawatts of power consumption. Basu concluded that the workload he measured produced “zero useful AI computation” and found no inference code in the dominant mining software.
Pearl’s protocol documentation says miners can select their own matrix inputs, including inputs from AI training or inference workloads. That flexibility is important for efficiency, but it also creates a gap between supporting AI in principle and proving that a particular mining calculation served a paying AI customer.
The result is a technology with a clear commercial test ahead. Pearl must show that its inference partnerships can generate enough demand to connect mining rewards to useful workloads at scale, rather than simply offering a new form of GPU-intensive proof-of-work.
