Pearl Research Labs’ founders held a live AMA in the project’s Discord last week, setting out a roadmap that runs through a floating-point hard fork, a possible token burn, and the prospect of AI inference being priced in the network’s own coin.
Within hours, a trading venue had turned the session into a campaign.
Margin Trade, the onchain perpetuals platform built by contributors to Solayer Labs, pinned its post repeating one line three times: trade $PRL spot & perps on Margin Trade.
The venue’s interest is straightforward. It has listed PRL perpetuals since mainnet and added PRL spot on 30 June, making it, by its own account, the only place to trade both. PRL has gained 56.6% over the past week to $0.5789, a market capitalization of $147.7 million against a fully diluted valuation of $1.216 billion, according to CoinGecko. Roughly 12% of the 2.1 billion supply is circulating.
Mining that doubles as inference
Pearl is a layer 1 that replaces hashing with matrix multiplication. Miners run noisy GEMM operations on Nvidia GPUs – the same arithmetic behind AI training and inference – commit the results, and wrap the work in zero-knowledge proofs for onchain verification. The approach comes out of a 2025 paper by Komargodski, Schen and Weinstein on proofs of useful work from arbitrary matrix multiplication, which claims overhead of roughly 1+o(1).
Co-founder and chief executive Omri Weinstein holds a Princeton doctorate in complexity theory and has held academic posts at Columbia and the Hebrew University, with stints at Nvidia and Vast Data. Mainnet went live on 27 April. Supply is capped at 2.1 billion, exactly 100 times Bitcoin’s 21 million, on a declining reward schedule and roughly two-minute blocks.
A fork that thins the field
Pearl has not published a recording of Tuesday’s session. The fullest account is a set of notes posted to X by an attendee, with Margin Trade circulating the headlines; the following is second-hand until Pearl releases a transcript.
FP8 support is written and published to GitHub, according to those notes, but needs a network hard fork to activate. FP4 is targeted for the first quarter of 2027. Together the two are framed as opening useful work to frontier models that cannot participate today. Pearl currently supports inference across Nvidia’s range, Blackwell included, with other accelerators next.
The fork would also clear out the low end. Older GPUs will find it very difficult to stay on the network, the founders said, with 40-series cards including the RTX 4090 expected to drop off before long. The notes describe that moment as the likely inflection point at which useful work begins to scale in earnest.
The absorption question
The part the market cares about is how any of this reaches the token. Pearl Labs is weighing options for the inference marketplace it expects to host: Weinstein suggested inference purchases may need to be paired with or denominated in PRL, and said a burn was under consideration as a way of absorbing emissions.
He also put numbers to it. If 15% of GPUs were to run two-for-one “Pearlified” inference and use the dollar margin from that work to buy PRL on the open market, he argued, the buying would absorb the network’s entire issuance – even at 10 to 30 times the current fully diluted valuation.
The comparisons offered were Cypherpunk, the largest Zcash treasury company, which holds roughly 18% of that network’s hashrate and mines expressly to accumulate, and the listed Bitcoin miners of 2021 and 2022, which financed operations with debt and equity so they would not have to sell production. Pearl’s miners, on this argument, are funded by inference revenue and are likewise not forced sellers.
One partner on the record
Pearl’s clearest commercial proof point remains Together AI. In May the inference provider announced Gemma-4-31B-it-pearl, a serverless endpoint offered at a discount of more than 25%, with the gap covered by PRL emissions. Pearl “changes the unit economics of AI, by allowing every GPU cycle powering AI training and inference to simultaneously produce a native proof-of-work digital asset,” Weinstein said at the time. Together said it planned to widen the Pearl range and eventually pass emissions to customers.
A second name surfaced at the AMA. Attestable’s zero-knowledge system, paired with Pearl’s kernel, was described as one of the few ways to prove that a provider is running the model it says it is running – a pointed reference to inference providers quietly substituting cheaper models to widen margins. Neither company has issued a release, and the partnership rests on the AMA account alone.
Weinstein also said Pearl was in close contact with two further major inference providers, and that several neoclouds are mining the network without having announced it. Internal tests of the floating-point integration reportedly show close to zero overhead.
Why the venue is shouting
Margin Trade reached mainnet in early June, live by the 3rd and formally announced on the 9th, a cross-margin, non-custodial venue settling positions, funding and liquidations onchain. It was built by Solayer Labs contributors alongside former Citadel and Kraken traders, and launched with crypto perps, gold, silver and oil, a synthetic US equity index, and a PRL perpetual at up to 3x.
That wider product still exists, but it is no longer the story the company tells. Since 9 September the feed has been almost entirely Pearl, including a #LongPRL campaign running to 19 September that pays 200, 120, 80, 60 and 40 PRL to the five best quote-tweets, each required to carry a Margin Trade deposit address.
The venue has a direct commercial interest in the narrative it is pushing: it earns fees on PRL flow, sells leverage on the token, and is paying PRL to recruit longs. None of which makes the underlying thesis wrong, but it is worth weighing against a source with nothing to sell.
What is still missing
Pearl has one named enterprise customer paying for inference. The share of proof-of-useful-work mining that corresponds to inference somebody actually bought, rather than work performed speculatively to earn emissions, has not been disclosed – and that figure is what the word “useful” ultimately rests on.
The FP8 fork has no date and requires network consent. Mining economics have already turned: daily revenue for an RTX 5090 fell from about $33.80 to $17.19 within weeks of launch as difficulty climbed. Spot liquidity is thin, at roughly $893,000 across listed venues in 24 hours for a token that moved 56.6% in a week. And the distance between a $147.7 million market capitalization and a $1.216 billion fully diluted valuation is nine-tenths of the supply still to come, which is precisely why the absorption question matters more than the roadmap.
