Industry roundup #6
Hyperscalers, neoclouds and data center REITs: I made sense of it so you don’t have to
Two days ago, I was reading this news about CoreWeave leasing data center capacity from Applied Digital for $7B in a 15 year deal. Originally a blockchain mining company, CoreWeave pivoted to become an AI cloud provider to capitalize on the booming demand for GPU compute. Applied Digital used to be a data center hosting company, but it’s pivoting too, changing their business model to a data center REIT company. The deal has been described as a “lifeline” for Applied Digital, and the stock price of both companies popped more than 30% on the news.
My first reaction was: isn’t that CoreWeave’s core business, building and running data centers? Why do these “neoclouds” lease? And why do we even need a new word instead of just saying … “small hyperscalers”1 maybe?
So I went down the rabbit hole, and this is what I found. TLDR is:
Hyperscalers are like Costco: you find everything you want there, they are everywhere, they are an established brand.
Data center REITs are like malls: they own and operate the building, they collect rent but do not own or operate any store. If nobody leases the space, they would be an empty building full of air conditioning equipment.
Neoclouds are specialized stores, like an artisanal bakery or a handmade jewelry shop, and give life to the mall.
Let’s dig into it.
Hyperscalers are the staple names, the usual suspects: AWS, Azure, GCP, and the like. Hyperscalers offer every type of cloud service, from object storage to databases to serverless, with GPUs as just one SKU type. They have a worldwide footprint, hundreds of managed services, enterprise compliance, and sometimes custom chips (i.e. non-Nvidia) such as AWS Trainium2.
They build and own data centers but also lease. AWS, for example, is split almost 50/50 owned vs leased. The reasons for leasing? Accounting optics is one, you want some opex alongside capex to show that you are a prudent and balanced steward of shareholder interests. Also, time to market plays a role: it is much faster to lease an existing site than to develop a greenfield build.
Neoclouds (CoreWeave, Lambda, Nebius, Together.ai) skip the 200-dish buffet and specialize in GPU-as-a-Service, focusing exclusively on AI workloads. The entire stack is tuned for the heavy, synchronous traffic of AI model training and large-batch inference. Their prices are often much cheaper compared to hyperscalers. Neoclouds are some of the biggest customers of Nvidia, and Nvidia has equity stakes in some of them. A bit sus and ponzi-esque, if you ask me. They buy and own the stuff inside the data center, but they lease the data center floor area itself. The reasons for leasing? Mainly liquidity. Their debt is lower quality, so it is better for them to invest in GPUs (their core asset) than to have illiquid real estate debt on their balance sheet.
This is where data center REITs (Applied Digital, Digital Realty) come into the picture. They are in the permissioning, building, and powering business, creating 250MW cooled containers that anyone (hyperscalers and neoclouds) can lease for 10-15 years. They are not in the data center hosting business, they do not manage the operations and the infrastructure. They are just the landlords, offloading the operational burden (and risk) to tenants like CoreWeave, swapping unpredictable usage-based revenue flow for stable long-term leases.
It seems that the industry is converging towards this model: either full vertical integration (i.e. the ownership side of hyperscalers), or landlords and tenants (i.e. neucolouds/hyperscalers leasing data center space to REITs). Between Costco, a mall landlord, and a niche little store that is (1) highly exposed to supply chain constraints, (2) very capital intense, (3) extremely specialized, and (4) vulnerable to Costco’s bigger scale, I would bet on Costco and the landlords.
Do you like this post? Of course you do. Share it on Twitter/X, LinkedIn and HackerNews
Although “small” and “hyper” is a bit contradictory. Maybe just “scalers”.