The usefulness gap: a study of what Pearl’s miners compute

The rules check the arithmetic, not its purpose

Basu’s firmest result is about the verifier. A miner chooses its own matrices, and Pearl’s proof check confirms that the slices of them it opens hold values between −64 and 64, that they match the miner’s Merkle commitments and that a tile recomputed from them meets the difficulty target. Nothing in it asks where the matrices came from. All 20 proofs he built from random matrices passed it at a reduced test difficulty, and his own open-source miner had 44 shares accepted by a public pool, AlphaPool: 3 from an NVIDIA RTX 3090 and 41 from an AMD MI300X. (The abstract also names CPUs and Apple Silicon, which ran his miner but found no share.)

This is less a flaw uncovered than one measured. The paper Pearl is built on lets miners pick their inputs on purpose and leaves usefulness to incentives; Basu calls the gap “a design property”, not a vulnerability. Four rule changes have activated since his measurements: Mixture-of-Experts (MoE) proofs were allowed on 12 Jun 2026 and rejected again on 23 Jul, a noise-rank rule followed on 6 Aug and salted noise seeds on 11 Aug. None of them, as Pearl’s node describes them, asks where the matrices come from.

What the network was running

The claim that Mainnet’s mining did no AI work rests on indirect evidence about one program. On 30 and 31 May 2026, Basu classified the 8,012 online workers of AlphaPool’s 15 largest miners by share count (the pool then had about 21% of the network’s hashrate) by their hashrate: all, by his reckoning, had GPUs able to serve a model. Whether they did, he judges from alpha-miner v1.6, the closed-source program he identifies as the dominant miner. It contained no strings pointing to inference software or model files. The mining runs he profiled, on an RTX 3090 and on an MI300X (the latter with his own miner), used 2.7 to 3.7 GB of GPU memory at full compute load, far less than the 16 to 40 GB he estimates serving a model would take.

He calls the string and memory evidence suggestive, not conclusive, since strings can be stripped, and the GPU classification approximate, since it was calibrated against the same miner. So the finding is narrower than the headline: the miner he calls dominant, in late May 2026, mined without serving a model. That does not show what every miner ran then, or what any runs now. Pearl’s repository has its own miner that mines inside vLLM, an inference server; the paper doesn’t measure how many miners use it.

Figures with a date on them

The most quoted numbers are estimates built on dated inputs. The network’s 24 EH/s, in late May 2026, becomes about 320,000 RTX 3090-class GPUs drawing an estimated 112 MW; the paper doesn’t show the conversion. Pearl’s hashrate is itself an estimate, derived from difficulty in Pearl’s own units, so a GPU count from it depends on the card assumed.

The economics changed between two versions, two days apart. Version 1 (3 Jun 2026) priced PRL at $0.21 and found mining unprofitable on every GPU tier it priced, by 54% to 72%. Version 2 (5 Jun) used $0.76, CoinGecko’s price that June, and found returns from −1% on an RTX 3090 to +67% on an RTX 3060 Ti: near break-even, with his estimates of the network’s daily reward and running cost about equal ($1.63 million against $1.54 million). Some of version 2’s figures still follow the old price, such as the loss on Basu’s own eight-hour run and the value of the 1.26 PRL he earned. Price, hashrate and the GPUs mining have all moved since.

The GPU-rental claim is the least settled

Basu also reports that budget GPU rental prices on vast.ai rose 38%, and their utilization from 57% to 94%, after the mining software’s public release in May 2026, crowding out research users. The price rise was measured; its cause wasn’t isolated. The paper ties the 38% to two different periods (the two weeks after the miner’s release, and the months since Pearl’s launch), a simple average of its own table’s rows gives about 49%, and the comparison with data-center GPUs, whose prices rose about 15%, rests on an assumption he says the data support but cannot prove. He also writes that he can’t separate Pearl’s effect from other AI demand.

What would close the gap

The paper weighs five ways to make useful work enforceable. The first, checking that matrices look like model weights, he calls a necessary first step, but random numbers drawn to resemble weights defeat a naive fixed-threshold check of their distribution, though he notes a more sophisticated test might tell them apart. The others need evidence from outside the proof: on-chain compute contracts, which he notes Pearl’s white paper leaves to future work; trusted hardware that attests to the workload; signed commitments to the data; or economic incentives for attributable work. Each trades some of proof of work’s permissionless simplicity for evidence about purpose.

We found no published response from Pearl Research Labs (searched 21 Oct 2026). The paper reads the design’s bet as economic, with usefulness left to incentives, and argues the incentive points the other way for now: by his estimate, which he doesn’t source, running a model alongside costs a miner 10 to 30% of its hashrate, so a bare miner earns more. Nothing public yet shows how many miners serve models now.

What the explorer can and can’t add

Pearl Pulse sees the same boundary from the chain side. Proof Lens checks a block’s certificate bytes and commitments and shows whether its proof is dense or MoE; neither says whose matrices were multiplied, or whether anyone used the result. The mining pages attribute blocks to payout addresses, not to workloads. An answer to Basu’s question needs operating evidence from miners and their customers, which no block record holds.