The usefulness gap: what $50 of cloud GPU proved about AI mining

The study (arXiv:2606.04819, June 2026) did not need a lab. By sampling worker submissions on AlphaPool and reconstructing what the miners were actually computing, it profiled 8,012 workers: 100% generated their input matrices with a random number generator (a uniform -64..64 range), and the miner binary contained no inference pipeline at all. On a network of ~24 EH/s — hundreds of thousands of GPUs — the measured amount of real AI work was zero.

This is not a bug report; it is a property of the design. Pearl's NoisyGEMM protocol verifies that a noisy matrix multiplication was computed correctly — the tile hashes, the BLAKE3 commitments, the Plonky2 compression all check the arithmetic. Nothing in the proof system asks where the matrices came from or whether anyone wanted the result. The original Weinstein paper (IACR ePrint 2025/685 — real, peer-reviewed math) says so itself: enforcing usefulness is an economic problem, not a cryptographic one.

One survivor in five attempts is the base rate every 'useful work' chain starts against. The graveyard pattern is consistent: either the work is verifiable but worthless (Primecoin) or valuable but not independently verifiable (Curecoin/Gridcoin), so a middleman ends up trusted. Pearl proposed that economics would close the loop — miners would run real inference because AI demand would pay better than random tiles. The study measured that hypothesis: so far, it has not happened.

• A technical response to the study — either disputing the measurement or shipping enforcement. Neither exists yet.
• Supply transparency: Upbit and Bithumb deferred listing over circulating-supply concerns; with ~175M of a 2.1B cap circulating (~50x FDV dilution), the venues are waiting on reconciliation.
• FP8/BF16 support: NoisyGEMM ships INT8 only today — most real AI workloads cannot plug in even if miners wanted them to.
• A second independent AI partner after Together AI — one integration validates plumbing, not miner behavior.

Meanwhile the mining math is plain: the same study estimates mining ROI between −54% and −72% at then-current prices — meaning fleet margins must be computed as ordinary proof-of-work on random matrices, not as an AI subsidy. That is how pow.family prices it. The tracked-set summary: the cryptography is real, the study is real, the silence is real — and the three signals above are the falsifiable things to watch.