Who makes labs for AI data centers

The model developers sit at the top of everything in this catalog. They do not make parts; they generate the demand that every layer beneath them is responding to, and their training and inference plans are what a supplier's order book is ultimately a bet on. How they get compute varies more than the category suggests, and the difference matters to the supply chain. Some build and operate their own facilities. Some contract long-term capacity from hyperscalers, which is how a large share of frontier training has been done. Some sign with the neoclouds in the sibling part, and some do all three at once. Each route puts the physical build in a different party's hands, which is why an announcement about a lab's compute plans may or may not correspond to a new building anywhere. This is also where the catalog's evidence is thinnest, and it should be read with that in mind. Most of these companies are private, disclose little, and describe compute commitments in dollar figures over multi-year horizons rather than in megawatts or delivered racks. What to look for is whether a stated commitment corresponds to a site with power secured, whether the capacity is owned or contracted, and how much of an announced figure is a purchase order rather than a framework. Every placement on this part is a demand relationship, not a supply one.