What is Pearl's FP8 Upgrade?
Pearl's FP8 upgrade hit testnet yesterday, and it's the first big test of its pitch to AI providers.
PEARL has spent the past week rallying despite the market’s beating.
A number of things have fed this counter-trend: Tier 1 exchange listings on Coinbase and Kraken last Friday, a large position disclosed today by well-respected fund Syncracy Capital and, most importantly in my eyes, Pearl itself taking a step toward proving whether its pitch to AI providers actually holds up. That came on Tuesday, when its upgrade to FP8, an 8-bit number format AI models already run in, went live on the project’s testnet.
That pitch is that AI providers can mine PEARL with work their customers already pay them for. So far, few do (or at least say they do publicly), likely because mining has meant changing how they run their models and accepting a slower service. FP8 is designed to make that a much smaller ask, which makes it the first big test of whether more will.
— (@)
What is Pearl?
As I wrote last month, the easiest way to understand Pearl is as a response to Bitcoin’s shrinking security budget.
Bitcoin pays miners in newly issued BTC and transaction fees to keep the network secure, and every halving cuts the new coins miners receive in half while machines and electricity cost the same. So miners are pivoting to a better business: AI, which CoinShares estimates can generate about three times the annual profit per megawatt. Evidence of this is that public miners sold a record 32,000-plus BTC in the first quarter of 2026 alone. Most recently, miner MARA disclosed last week that it had sold about 40% of its bitcoin stack since the end of September amid its own pivot to AI.

Pearl, itself a Bitcoin fork, piggybacks on the AI buildout. When you ask an AI model a question, what happens behind the scenes is effectively a bunch of math, matrix multiplication to be exact. Pearl lets some of those multiplications double as entries in the race to mine the next block. So the GPUs being deployed to serve AI customers can mine Pearl with the same work, a setup commonly referred to as duplex work. The customer’s payment helps cover the GPUs and electricity, so the block reward comes as a bonus, and miners have less reason to sell it.
The catch is that duplex work isn’t required. You don’t need to be serving inference to mine Pearl. Someone can simply multiply random numbers and earn block rewards. So far, this “junk mining” appears to account for the vast majority of the network, with Blockworks estimating earlier this month that fewer than 1% of blocks matched a real AI model. Part of the reason is that most inference doesn’t qualify for mining as things stand.
What FP8 changes
FP8 changes which inference counts toward mining. Models can use different formats for the numbers they multiply, and right now, Pearl’s mainnet only counts multiplications done in one of them, INT7, which few model providers use. Case in point, OpenRouter doesn’t even list it among the formats providers can report.
Enjoying this article?
Subscribe to Bankless or sign in
So a provider running a model in another format has to convert it to INT7 to mine Pearl, and even then, mining slows the service by 10% to 15%, according to Pearl Research Labs CEO Omri Weinstein, who gave that estimate in September.
The upgrade moves mining to FP8, a format providers already use. On OpenRouter, it’s the most common format providers report, accounting for 58% of the endpoints that disclosed one in October. Once the upgrade reaches mainnet, a provider running a compatible FP8 model could, in principle, mine PEARL with multiplications it’s already doing. It would still need to run Pearl’s software, but it could skip the format conversion that INT7 currently calls for. Weinstein has since said the new design should cut the slowdown to between 1% and 5%, depending on the model, a figure that comes from the team itself.
Early outside testing has found very different numbers. In a report published today, Kairos Research, which holds PEARL, said it ran Pearl’s current INT7 build and its FP8 testnet build on a rented H100. Under INT7, it found the server delivered 30% to 65% fewer tokens than normal serving. On the FP8 build, it found mining still cost about 20% to 40% of the server’s throughput. Kairos called that real progress and cautioned that the results come from one early build on one GPU. Even so, that’s far above what Pearl Research Labs has claimed, though Kairos tested only one chip.
— (@)
Switching to FP8 comes with a hardware trade-off, though. Currently, a wide range of GPUs can mine Pearl, because INT7 math comes out the same on every chip. FP8 math doesn’t. Different chip families round FP8 results slightly differently, so the same multiplication might come out as, for example, 1.234 on one GPU and 1.235 on another, while everyone checking a block needs to see the exact same result. Pearl addresses this by standardizing the hardware. Once FP8 is live on mainnet, miners who want to participate will have to use one of two families of Nvidia chips, Hopper or Blackwell, as the proposal is currently written. Everything else, including older Nvidia cards and AMD and Intel GPUs, would be shut out. By Kairos’s estimate, that would cut about two thirds of today’s hashrate.
To be clear, none of this stops junk mining in its tracks. Weinstein himself says it will always be possible. The intention is to remove the economic incentive: a miner who only chases PEARL leaves inference revenue on the table. That logic has limits for now. Kairos reports that in the build it tested, a server could mine only while reading a prompt, not while writing its answer, which would leave a GPU that only mines earning more PEARL than one mining alongside paid inference. Pearl’s bet is that, over time, serving inference alongside mining becomes the more efficient way to run the hardware.
What would count as success?
Established providers saying they mine PEARL, and offering more endpoints discounted by mining rewards, would be a great start.
Beyond that, two numbers will show whether FP8 works. The first is what mining costs a server that’s also serving paid inference. Pearl says 1% to 5%, and Kairos’s early test on Hopper found far more. Kairos plans to run a standing benchmark on every new build, so we’ll see whether those numbers converge. If they do, providers have much more reason to turn mining on.
The second is harder to see: the share of blocks matching a real AI model. Pearl’s chain can’t tell inference from random math, so that share only exists as outside estimates like Blockworks’, which put it below 1% earlier this month. If FP8 works, those estimates should climb once the upgrade reaches mainnet, which has no announced date yet. If they don’t within a few months of mainnet, the format may not have been what held providers back.