Proof of Useful Work
and the 2-for-1 GPU.

Matrix multiplication — MatMul — is the foundational math of deep learning. Every time a model is trained or run, it is, underneath, multiplying enormous grids of numbers together, billions of times over. It is also one of the most resource-intensive and expensive operations in computing, which is exactly why AI runs on costly specialized GPUs.

Traditional proof-of-work, the mechanism that secures Bitcoin, is built on a different kind of math: hashing. Miners race to find a number that produces a particular hash. The work is real — it costs real electricity — but the answer is worthless the moment it is found. The puzzle exists only to be hard. Critics have spent a decade pointing out that this burns the energy of a mid-sized country to produce nothing of independent value.

Proof of Useful Work (PoUW) asks the obvious question: why not make the puzzle something we actually need solved? If the work securing the network were also the work of training and running AI models, the same electricity would buy two things at once — a secure blockchain and useful computation.

The idea is old. Making it work, securely, is the hard part — and that is what recently changed.

The revenue loop. A GPU has three states: idle, doing low-value hashing, or doing real AI work. PoUW collapses the last two into one. Instead of letting hardware sit idle between jobs — or running pure-hash mining that produces nothing — the card contributes directly to training and inference while simultaneously minting cryptocurrency. The compute you were going to pay for anyway now also pays you back.

Energy efficiency. This answers the oldest criticism of proof-of-work head-on. The electricity burned and the wear on the hardware now yield a commercially valuable output — AI compute — rather than disposable hashes. The same joules do double duty.

Decentralized AI networks.Stacked across thousands of operators, MatMul PoUW becomes a way to bootstrap a decentralized compute marketplace. By mining, an operator is effectively renting their hardware into a distributed AI cloud that compensates them natively, in the network's token, for the output their GPUs produce.

The cryptography is sound, but deploying MatMul PoUW at scale runs into two practical limits.

Hardware thresholds.Efficiently computing the cryptographic proofs needs specialized, Hopper-class tensor-core instructions — H100 or H200 territory. That leaves older and consumer-grade cards out of the loop entirely. PoUW is, for now, a game for owners of the most expensive AI hardware on the market, which both concentrates participation and ties the network's growth to GPU supply.

Verification overhead. Proving that an arbitrary matrix result is genuinely correct — and not a miner gaming the system — requires the random self-reduction machinery and strict verification protocols described in the research. That adds protocol complexity and a verification burden the network has to carry, even if the per-computation overhead on the prover side stays near 1 + o(1).

Neither is fatal. Both are the kind of friction you would expect from a primitive that is roughly a year out of the lab and into its first mainnets.

For a decade, proof-of-work's great weakness was that all that electricity bought nothing but security. Proof of Useful Work, built on a near-free way to make matrix multiplication itself the proof, closes that gap. The same computation that trains a model can now secure a chain and pay its operator — at an overhead the research pins at 1 + o(1).

It is early. The hardware bar is high, the verification protocols are demanding, and the first mainnets like Pearl are still proving the model in production. But the direction is clear: as AI compute gets more expensive, the pressure to make every GPU cycle earn twice will only grow. PoUW is one of the cleaner answers anyone has found — and it runs on exactly the kind of regulated, power-secured infrastructure the rest of the AI economy already depends on.