$IREN: Addressing SemiAnalysis Reply Truths I appreciate @dylan522p for now telling the truth: "There's not a single cloud that doesn't have downtime, the question is how much, how often, etc." This is exactly what I'm saying in my post: IREN has had downtime but other clouds do too. What makes the @SemiAnalysis_ post on IREN a hit piece is that there is zero mention of this context or any data on the rest of the industry and IREN is singled out as having these issues. "Nebius downtime and reliability issues were much less." I have not claimed that Nebius doesn't have downtime, I'm simply pointing out Nebius have multi-day outrages too. The difference between my writing is that I do not exagerate claims beyond what I have data for. It's very hard to get industry data because many clouds do not publish all their incidents as customers have layers of failover so many incidents don't impact the final bottom line of customers. Many incidents stay private between the Cloud and Customer. TogetherAI "TogetherAI was downgraded in ClusterMAX because of reliability issues". $IREN is actually less than 1% of TogetherAI's compute with most of it coming from other Neoclouds and HS. "Together AI said that it was "partnering with independent cloud providers like Hypertec, and deploying our inference stack on hyperscalers like AWS, aggregating and accelerating a massive network of global GPU compute"" (2). TogetherAI has hit 1.15B in ARR and IREN is providing approximately 9m and 17m of compute in Q1 and Q2 respectively - splitting IREN's Q1 and Q2 revenue among FireworksAI, TogetherAI, HumeAI, and Fluidstack. Unless TogetherAI doesn't have region failover implemented, I do not see how IREN can be blamed for TogetherAI getting downgraded. One basic function of an AI Platform is have resiliency to handle less than 1% of compute having downtime. Multiple Sites "There's multiple sites with item that have had issues". At the time which the hit piece was being written, only 1 IREN site has running GPUs and it is Prince George. H1 in Childress for MSFT is up now but that's under heavy security and NDA so it will take several more months before @SemiAnalysis_ gets any info on that. I don't blame @dylan522p for this miss information as he keeps track of hundreds of sites so it's easy to get miss remember. Info Sourcing "Vipul is not even guju". My mistake here, I remember wrong and apologized in this post (3). Conclusion I've read other @SemiAnalysis_ papers and @dylan522p clearly knows the industry well but it's clear he didn't proofread this hit piece that his juniors wrote. There is a sharp drop in technical/industry knowledge between their real research and their hit pieces. It's almost as if they were written by completely separate organizations.