Useful is optional

GM. Two things from us today: a PoW chain whose own supporters concede its miners aren't currently doing useful work, and a launch from our side of the house.

On our end, we launched the Blockworks MCP today. It runs on the Unified API we released last week, which puts all of the Blockworks and Messari data behind a single API key: more than 40K assets across 300+ exchanges, standardized financials for 400+ DeFi protocols, plus research, news, fundraising, unlocks and issuer disclosures. Add Blockworks as an app in ChatGPT, or paste one URL into Claude, Cursor or any MCP client. Ask questions in plain English, get numbers back with sources attached, chart them in your own branding. 

Separately, Nockscan posted a stat over the weekend that clashes with both sides of the PRL vs NOCK debate. Nockchain miners, running an AI-PoW mode that exists to merge-mine Pearl, produced 55% Pearl blocks between September 14-21st. 

Pearl launched in April as a Bitcoin fork with matrix multiplication swapped in to replace standard SHA-256. The pitch from Pearl Research Labs is that a GPU serving an LLM can mine PRL as a secondary revenue stream at ~5% overhead. Five months in, the chain sits at ~50 EH/s, PRL trades around $1.24, up 2.5% on the week, and Entropy paid 555 HYPE for the PRL ticker on Hyperliquid on Friday. Both camps agree on the facts, which is the odd part. In June, a preprint measured the network at 24 EH/s, and found zero strings referencing vLLM, Llama or any inference framework in the dominant miner binary, confirmed that the protocol accepts random matrices by design, and earned 44 pool-accepted shares with no model loaded. 

Brian Breslow of No Limit Holdings published a thesis that calls Pearl the next Bitcoin, and his answer to the paper is that it describes the design. Pearl only verifies expensive matrix multiplication. A miner who feeds it garbage pays full GPU cost for PRL. Miners who feed it customer inference have a customer paying the power bill and absorbing the overhead alone. Raise difficulty enough and the garbage miner becomes the marginal producer and winds down their operation. Rafael Pass, a Pearl advisor, wrote the economics paper this argument leans on. It is a coherent argument, but comes with a built-in test that thus far is playing out poorly.

Difficulty has tripled since the paper's measurements, from ~11M to ~33M. Under Breslow's model, synthetic miners should be on their way out. The block share tilted toward Nockchain instead, whose miners have no inference customers and whose "AI work" mode does nothing for Nock itself. Hashrate.no lists a 2,294 PRL block reward and about 1.02M PRL a day in emissions, $1.26M at current prices. The one place where useful work exists is Together AI, which sells a Gemma endpoint at a 25%+ discount and calls it Pearl-powered. The preprint's rebuttal is that Together's own GPUs run the inference, and PRL revenue subsidizes the price. 

Behind the discourse lies Breslow’s thesis that public Bitcoin miners are turning into landlords. His cohort table puts AI and HPC at ~2% of revenue in 2024, ~19% in the latest reported quarter, and 50-60% by year-end. Core Scientific reported 83% of Q2 revenue from HPC colocation, while TeraWulf reported 71%. Those buildings fill with GPUs on ten-year leases, and Breslow's bet is that a GPU under an AI contract mines PRL on the side because switching it off saves nothing. Bitcoin's hashrate has drifted lower since its 1.3 ZH/s peak in late 2025, and that part of the story holds regardless of what Pearl becomes. It is why a Zcash-style "1% of Bitcoin" comp keeps showing up on CT. Pearl at 1% of BTC implies a ~$18B FDV. For scale, prlstats has Pearl drawing 0.49% of Bitcoin's power today against 0.14% of its FDV, so the chain already buys proportionally more security than the market pays for.

Zana Ventures says the network will ban consumer GPUs in Q1 2027, meaning the game theory was supposed to price the RTX 3060 miners out, but the team decided to evict them instead. A cost model prices a dedicated RTX 4070 Ti rig at $0.33 per PRL, a ~73% margin at today's price, while an H100 bought purely to mine loses ~41%. The cheapest profitable way to mine PRL with hardware you bought for it is the consumer card the ban targets. To top that off, five consensus changes shipped between genesis and mid-August, including an MoE soft fork that miners exploited to spoof hashrate before the team switched it off 1.5 months later. A chain that still hand-edits its own miner set is a science project with a ticker on Hyperliquid, and for now the ticker might be the better trade.

— Nick

Stablecoin supply in Q3 was flat at about $306B, but the flows tell the story. Ethereum shed $6.9B while Tron absorbed $4.4B, mostly USDT. The rest of the gains went to newcomers: Robinhood Chain went from $100M to $1.06B, HyperEVM and Solana added ~$850M each.

Brought to you by:

​Join industry leaders, global partners, and OKX’s most valued customers in Singapore for a full day of live product demonstrations, ambitious ideas and a first look at what’s ahead.

Meet the innovators behind OKX’s products and discover how the capabilities come together to help you trade, move money and build your business.

​Apply to join us.

Every distributed system inherits the physical limitations, or advantages, of the network it runs on.

DoubleZero has demonstrated supply-side adoption, and the first tenant ecosystem connected more than half its stake, the majority of which publishes to Edge.

Read the full report from Blockworks Research to find out the open questions that decide whether Edge becomes a material revenue primitive or stays a proof of concept.