Datacenters-for-AI outfit CyrusOne has scored another $7.9 billion loan to build more bit barns to fill full of GPUs.

The sum comes on top of $1.8 billion in capital the datacenter operator sored back in May, bringing its total debt financing to roughly $9.7 billion. The bulk of the funding will support a variety of existing and future datacenter developments with a heavy emphasis on meeting growing demand for sustainable AI infrastructure, the outfit explained in a Monday statement.

Whether this investment will include entire GPU clusters or just powered shells where others can park their own accelerators isn't clear. Given the amount of cash involved, we'd assume the former, and have asked CyrusOne for more information.

Whatever the case, it's now common knowledge that training the large language models (LLMs) that power generative AI chatbots, image generators, and code assistants requires massive quantities of accelerators connected over high-speed networks. It's not uncommon to see clusters of 10,000 or more GPUs working in this way.

CyrusOne's emphasis on sustainability is also unsurprising, and not just because it has committed achieving carbon neutrality by 2030. AI datacenters have garnered considerable attention for the vast amounts amount of energy and water they require to operate. Emphasizing sustainability can therefore impact both a datacenter operator’s bottom line and burnish its environmental credentials. We fancy CyrusOne is likely less motivated by preserving the planet than maintaining margins by reducing energy consumption.

It helps that ,any of the most sophisticated accelerators, including Nvidia's top-specced Blackwell parts, are designed with liquid cooling in mind, which also happens to be quite a bit more efficient than traditional air cooling to operate, something we explored back in March alongside their launch.

AI datacenters are shaping up to be a tantalizing investment, if you can stomach the high capital costs required to get into the game.

Nvidia's flagship Blackwell accelerators will set you back between $30,000 and $40,000 apiece when they arrive later this year. Demand to access the cards, however, means they can be rented at high rates. Our sibling site The Next Platform, found that it costs roughly $1.5 billion to purchase, deploy, and network 16,000 Nvidia H100s today, over four years that investment should generate roughly $5.27 billion.

That assumes you either already have somewhere to put those GPUs or the means to build your own datacenter. With $7.9 billion in new capital in hand, CyrusOne and its investors appear to be making the bet this AI thing isn't just another bubble waiting to burst and demand for accelerators will continue to grow.

CyrusOne is not the first datacenter operator to fund expansion with heavy borrowing. As we've previously reported, CoreWeave and Lambda have both put their GPU farms up as collateral in exchange for financing.

Among the largest examples to date came last month when CoreWeave secured a $7.5 billion loan backed by its GPU collection. ®

https://www.theregister.com//2024/07/10/cyrusone_ai_datacenter/