15% of GPUs” means 15% of all the GPUs that actually run paid AI inference. It's a big claim The idea is: Pearl can be mined via the work that inference already does. So if a provider is already being paid to do commercial inference, they can get more money for the same job, for little or no extra effort. Then that provider could offer their customers more for the same cost, or cheaper prices. Over time competitors might have to mine $PRL as well just to stay competitive But that only works if a few things are true at the same time. The extra compute cost has to be close to zero, the model’s answers can’t get worse, latency can’t jump, and it has to be easy to turn on This is seriously impressive team though, so the ambitious goals are far from impossible. Omri Weinstein is a Princeton CS PhD, Hebrew University / Columbia academic, former VAST Data chief scientist, and was a senior research scientist at NVIDIA. Ilan Komargodski is a heavyweight cryptographer and co-author of the paper the chain is built on. They’ve got NVIDIA Inception, a live Together AI endpoint that sells cheaper Gemma inference funded by Pearl mining, and they say a couple of big neoclouds are already mining. We've also seen co-marketing ads with Google out in the wild, the team speaking at closed Nvidia events, and more. What’s live today is mixed. The network is up and hashrate is large. Some useful work is showing up. A study earlier this year found most miners were just grinding random matrices, not serving models. The protocol can check that a multiply happened. It doesn’t force that multiply to be a real customer job. FP8 is supposed to fix that. The spec is out. It is not live yet. Current overhead on real models is still a few percent, not zero So the “turn on $PRL mining, make more money for free and outpace your competitors, competitor have to follow and mine $PRL too to stay competitive” story works best for smaller inference shops running open models. That’s really plausible if FP8 delivers. Google, Microsoft, Amazon, Meta, OpenAI and the rest are most of the fleet though, and they might not start mining $PRL on production GPUs just because a token prints on the side. It will need to profitable or they won't do it Bottom line: 15% is a large claim, but the people building this know the math and the GPU stack better than almost anyone in crypto. Together AI proves that companies will do it if the economics make sense. But it's still a long way to go getting from Together AI to 15% of the world’s paid inference GPUs