Explaining AI Inference Data Centers and Comapnies

My position on the current movement of AI "Neocloud" companies in response to Oracle earnings, Nebius Contract, and related and correlated events.

The last 3 years of data center planning has been focused on cloud data centers as the base demand, colocation facilities (multi tenant and low latency), and then another model which is these 1GW ($10B+) data centers for AI machine learning (not latency sensitive). We are reaching, in the 2H of 2025, a time when these AI data centers are STARTING to run AI workloads. People don't realize that we have spent years trying to build these and now we are finally able to use the chips Nvidia and AMD have built. (AMD actually more used for inference).... This is accelerating quickly and the second phase of AI is starting - AI Inference. This is the phase where money actually starts being made from AI.

Important is that the hyperscalers who build cloud data centers do not build AI Inference infrastructure. This is built by colocation companies and neoclouds (i.e. Coreweave, Nebius, Applied Digital, Iren and private (Stream, Vantage, Aligned, QTS, etc). The cloud companies lease this capacity from these players to run HPC racks for their own use cases, and not necesarrily to provide to the public cloud (yet). Even more interesting is that Oracle does not own it's own data centers (Mostly) they lease all this capacity from the above players.The recent announcement with Nebius, Crwv, Vantage, Oracle, and others is just the beginning. AI Data Centers are built completely differently than what the hyperscalers typcially build and they are not capable of building this.

I believe we will continue to see these announcements from Neocloud companies. Oracle made it perfectly clear that they cannot meet their demand and that is why groups like NBIS, APLD, and CRWV are able to land massive deals. The only groups that hyperscalers are willing to work with are companies who have track record of years long building data centers - even if it is formerly bitcoin mining. They simply will not work with start up companies because they cannot reach the required ~100% uptime among other things.

You've got companies under $10B - $50B who hold the keys to the AI inference race, which is the phase that is actually profitable, and then you have multi trillion dollar cloud giants who are leasing this space because they have not planned to build "AI Inference Data Centers" they build cloud data centers. You have Nvidia partners like Iren ripe for a massive hyperscale deal and maybe only 10 others that will emerge. Its a group of small players, especially public companies. (My take on iren to follow soon as they are ripe for a large contract)

Power in the US: I don't think people understand how bad the power situation in the US is and how incredibly challenging it is to source power in 2025 and 2026. Even worse in tier 1 markets where inference needs to be built. Some utilties can't provide power for 5-10 years from today. No amount of money can buy you access to power that quickly to build new data centers. Companies, specifically Coreweave, who have built multi GW pipelines with access to immediate power in highly critical locations are positioned in a way where they are one of the only ones who can act.