Neoclouds have been derided as sketchy for a while now. Throughout 2025 as Michael Burry and co stirred up talk of the so-called 'AI Bubble', neoclouds got hit hardest because they're possibly the most intrinsically leveraged type of business for betting for or against the existence of an AI bubble.

Neoclouds are basically like the old guard of cloud providers, except rather than serving databases and virtual machines, they're more focused on serving GPU inference, the means by which AI labs/product companies are able to serve agents/LLMs.

There was a moment for a second where even the industry itself wasn't quite sure that demand for compute was going to be as great as it is right now. Models were improving, but it felt like it'd be significant time before an LLM was truly able to run for a long time (an hour or so, back then). That changed with Opus 4.5 - the model (but at this point, considered an agent) - is considered by most to the first 'next generation' model. A step change, like with GPT 3.5, 4, o1, etc.

Opus kickstarted a 'run on the data centre' - a frenzied period of exponential demand that has most recently seen OpenAI's Codex revenue double in less than a week. GitHub temporarily paused new user signs up for Copilot, one of the biggest agentic coding tools in the world and dominant provider for enterprise customers. Anthropic pissed everyone off and took a meaningful PR hit in the AI race by testing whether they could get away with quietly removing Claude Code entirely from their $20/m subscription. Requests to Google's models inside their coding agent tool regularly take over 20 minutes to get a response while they admit they didn't buy enough compute to serve demand.

The only people that seem to be coming out on top right now are OpenAI. Why? Because they have all the compute.

Everyone is desperately signing huge compute deals left right and centre because they know the truth of the matter is: if you have no compute, then you have no chance. The labs know this best, but the wrappers and product companies know it's coming for them soon too. They get their inference served by the labs themselves right now, but if demand for Codex continues to grow at the rate it does, why would they make it cheap for you to use their models in your own product? Prices will skyrocket, as they already are: OpenAI doubled the cost of their frontier model for the latest release, GPT-5.5.

Enter, neoclouds.

Are you a big AI lab struggling to serve your own demand? Come rent them from us, we have loads!

OpenAI's models completely and utterly cost prohibitive? Come rent GPUs from us and serve your own inference using open source Chinese models - which by the way, are very, very good now!

Had a research breakthrough and need GPUs to do a training run? You get it.

Neoclouds are basically split down the middle - early bettors on AI becoming much more capable and demand for compute had yet to accelerate: CoreWeave, and famously Oracle fall into this camp.

Then there are the 'reformed bitcoin miners': IREN, CleanSpark, Core Scientific etc - businesses that already bought GPUs at industrial scale to mine bitcoin, back when that was relevant. Now they rent those GPUs to AI companies for much bigger margins - and they've gotten a head start on scaling out their product offering specifically for AI companies. Revenues continue to shoot up, but we're not even yet seeing the true effect of the recent explosion in demand for compute and how it's manifesting into pipeline for neoclouds - the doubling of Codex's revenue happened 5 days ago.

Neoclouds have been battered this year, a lot of it owing to a dogshit article by the 'tech tabloid' The Information. People who were betting that AI was just a useless toy because it 'hallucinates', or because they do real engineering or whatever else they convince themselves of in order to cope with their perceived loss of leverage. People who think AI just takes water and makes it cease to exist. Unserious people.

H2 2026 is when the neocloud model will be vindicated, and it'll become clear that compute was the real race all along.