Gonka and Pearl are the two most serious attempts to replace proof-of-work's wasted energy with AI computation, and they represent opposite bets on what "useful" means. Pearl (PRL), launched by Pearl Research Labs in April 2026, verifies that miners performed enormous volumes of GPU matrix multiplication; Gonka verifies that GPUs served paying inference customers, and makes that paid service the consensus workload itself. The difference — work that is provably real versus work that is provably wanted — runs through every row of the comparison below.

One disambiguation before we start: this is Pearl Network by Pearl Research Labs, not the defunct Oyster Pearl or Perle Labs, both of which have used the same PRL ticker. Some aggregators list the wrong coin.

At a glance

Is Pearl decentralized AI, or verified matrix multiplication?

Pearl deserves a precise description, because its foundations are genuinely rigorous. The chain implements a real theoretical advance: the April 2025 paper "Proofs of Useful Work from Arbitrary Matrix Multiplication" by Komargodski, Schen and Weinstein — Weinstein is Pearl's founder — which shows how to turn large matrix multiplication into a mining puzzle with only ~1+o(1) overhead versus just doing the math. The base layer is a conservative fork of Bitcoin's btcd with post-quantum XMSS signatures, and mining results are wrapped in BLAKE3 commitments and zero-knowledge proofs. Miners are provably doing the matmuls.

The open question is whether anyone wants them. Today's proofs cover exact integer matrix multiplication, with an upgrade path to the low-precision formats AI actually uses (BF16, FP8) stated but not scheduled. And the chain itself contains no mechanism — no pricing, routing, escrow or settlement — for a customer to submit an AI workload and pay for the result. What Pearl verifies is that computation happened, not that it was commissioned.

An empirical study posted to arXiv in June 2026 — "The Usefulness Gap in Proof-of-Useful-Work" — pressed on exactly this point. Measuring 8,012 workers producing ~24 EH/s (about 320,000 RTX-3090-class GPU-equivalents drawing an estimated ~112 MW), it concluded the network "produces zero useful AI computation": the dominant mining software contains no inference code, runtime profiles match pure matrix math rather than memory-bound transformer inference, and the study got pools to accept shares built from random matrices. Caveats apply — it is a single-author preprint, days old at this writing, and Pearl Research Labs has not published a response — so treat it as a claim, not a verdict. But even Hashrate Index's broadly sympathetic analysis lands in the same place: computation is only useful if someone pays for the result, which makes most current Pearl mining "AI-shaped proof-of-work, not useful work."

How does Gonka enforce usefulness instead of asserting it?

Gonka closes the gap architecturally rather than aspirationally: AI inference is the chain's on-chain product, and the consensus mechanism is built out of it. Every request carries per-token pricing, the maximum cost is escrowed before a GPU touches it, and settlement pays providers for tokens actually generated — with the start-inference and finish-inference messages visible in the live transaction feed. The GPUs that win consensus power are, demonstrably and inspectably, the ones serving paying users.

The proof-of-compute side is engineered to bind to that same workload. Each epoch opens with a short Sprint in which participants generate transformer-based proofs seeded by a fresh block — the proof task exercises the same model-serving stack that handles customer requests, so capacity cannot be faked with hardware that would be useless for the actual job. Then every participant audits everyone else by re-generating a statistical sample of claimed proofs, and the fraud statistics go on chain; acceptance requires a compute-weighted majority. The full pipeline is in How Gonka's Proof of Compute Works.

The energy story follows directly. Pearl's estimated ~112 MW was measured (by the study above) producing no AI output — proof generation is the network's full-time job. Gonka confines proof generation to the Sprint at epoch boundaries and dedicates the long middle of every epoch to paid inference, so the electricity mostly buys something a customer asked for.

Where can a developer actually buy compute?

On Gonka, on the chain itself. Fees are the product of tokens processed, a per-model units-of-compute constant, and a unit price set each epoch as a weighted median of participant proposals; escrow guarantees the provider is paid and the user is refunded any unused maximum. Ordinary transactions carry no gas at all — the design, covered in Gonka Tokenomics, prices exactly one thing: model work. As of July 2026 the on-chain model registry lists MiniMax M2.7, Kimi K2.6 and GLM 5.2 — frontier open-weight models — and the inference page shows the volume being served.

On Pearl, the honest answer is: not on Pearl. The network's one live paid-AI product is an exclusive Together AI endpoint announced May 15, 2026, serving Gemma-4-31B at more than 25% below standard rates, with the discount funded by PRL emissions captured during inference. That is a real product from a major AI cloud, and Together says it is the first of a planned portfolio. But it is one model through one vendor, priced and settled off-chain — the blockchain's role in the transaction is to be the subsidy, not the marketplace.

What do the two hardware fleets actually look like?

Gonka's fleet is an enumerable datacenter inventory. As of early July 2026, roughly 35 epoch-active participants run about 196 ML nodes with ~1,071 GPUs online, counted by model on chain: 376 B200s, 232 H100s, 176 H200s, 166 B300s and 56 A100s. You can check those numbers yourself on the participants page and the network tiles — they are chain state, not marketing.

Pearl's "~320,000 GPU-equivalents" is a different kind of number: a difficulty-normalized conversion of hashrate into RTX-3090-class units. The official miner is Hopper-only (H100 minimum, H200 recommended), pairs with a vLLM-based worker, and the official pool charges 20% fees; community builds extend mining to consumer cards, largely on rented cloud instances. The mining rush had measurable externalities — the June study found budget GPU rental prices rose 38% and cloud utilization jumped from 57% to 94% after miner release, displacing research workloads — and per-card economics compressed fast: hashrate.no estimates cited by Tom's Hardware put RTX 5090 revenue falling from about $33.80 to $17.19 per day in roughly six weeks as difficulty climbed. Gonka's income model is deliberately less exciting: escrowed fees for inference actually served, plus work-proportional minted rewards vesting over ~180 epochs. Revenue tracks delivered service, not a difficulty race.

What does Pearl get genuinely right?

Credit where it is due, because Pearl's strengths are real. Its cryptography is peer-visible research authored by its own founding team (IACR ePrint 2025/685) and featured by a16z crypto, not whitepaper hand-waving. The Together AI partnership is commercial validation from a serious AI cloud — something Gonka has not matched. Its monetary design is Bitcoin-grade minimalism: a fixed 2.1 billion cap, a smooth emission curve, post-quantum signatures, fully permissionless mining, and a claimed fair launch with no premine or VC allocation (a claim Hashrate Index calls "positioning worth verifying rather than gospel," but a clean story if it holds). And its bootstrap was explosive: roughly 8,000 workers within six weeks of mainnet. Pearl is also simply more visible as an asset — tracked on CoinGecko at roughly a $97M market cap (rank ~#257 as of July 5, 2026, though the PRL ticker collision makes any aggregator figure worth double-checking) and listed on exchanges like MEXC.

Where is Gonka younger?

Market presence is the honest gap. GNK currently trades over the counter via hex.exchange rather than on listed venues, has no CoinGecko rank to quote, and Gonka's 35 active participants look small next to Pearl's 8,000 workers. The comparison is not quite apples to apples — Gonka's number counts datacenter operators serving paying customers under a proof-and-collateral regime, Pearl's counts pool workers chasing emissions — but a younger, deliberately narrower network is what Gonka is. The trade-off it chose is that every claim it makes is checkable on its own chain.

The verdict: enforced vs asserted

Pearl has proven that miners will do verifiable matrix math at enormous scale for token rewards, and that is a legitimate research milestone. What it has not yet built is a reason for the math: no on-chain buyer, no native inference market, and — if the June study holds up — effectively no AI being computed. Gonka started from the other end: build the market first (per-token pricing, escrow, settlement), then derive consensus from the same GPUs that serve it. On one network, usefulness is a roadmap item; on the other, it is the block-by-block content of the chain.

You do not have to take our word for that — it is the whole point that you can look. Watch the inference lifecycle land in live transactions, check which models are serving, and see the epoch cycle turn in the blocks feed. For the background, start with What Is Gonka?