Premium: Jockeying for position

Oracle and Broadcom results seem to be weighing on the AI side of the market today (as well as Fermi's phantom anchor tenant backing out), right after Google TPU/Gemini had already been weighing upon NVIDIA and neoclouds for the past few weeks. And now, space-based data centers seem to be the next shiny thing to ramp up the FUD, as if that will just show up overnight to disrupt the next 3 years of terrestrial AI buildout plans. [The proof of concept from startup Starcloud is running micro-models on one whole H100 in a satellite.]

SemiAnalysis released its latest ClusterMAX (2.0) ratings last month, reviewing the GPUaaS providers across hyperscalers, neoclouds, and others (aggregators and marketplaces). The post is all public (no paywall), but at 46K words, it's also dense with technical details of the tools, tests, and workarounds, as well as commentary on each player's stack.

This new rating system aims to look through the eyes of an AI-ready customer who is ready to create a reserved GPU cluster on a GPUaaS platform. Their write-ups are trying to appeal both to those customers (detailing the steps they took and issues that were encountered) as well as the GPU and AI chip providers themselves (highlighting areas to improve).  

Let's isolate what is important to a third audience: investors – aka what matters to hyperscaler, neocloud, or AI chip companies from here. This will be of interest to NVIDIA, AMD, CoreWeave, Nebius, IREN, the hyperscaler clouds, and any crypto-to-AI plays.

• CoreWeave continues to set the bar for neoclouds as the sole Platinum-ranked provider. They excel at burn-in and ongoing health checks, and are continuing to improve their software stack.
• Nebius is another pure-play neocloud to watch, which has maintained its Gold rating. They too have a strong software stack, and have system build and architectural advantages. They started in the EU, but have been moving way more heavily into the US over the past year.
• Crypto-to-AI play IREN continues to have a lackluster ClusterMAX rating, yet has since secured a major new deal with Microsoft – as did emerging smaller neoclouds like Lambda and Nscale. They are more focused on bare-metal sales to clients that bring their own software stack – but have a lot of runway (3GW secured capacity) to exploit.