While AWS, Google, Microsoft, and Oracle have broadcast enormous capital spending on projects to extend their cloud infrastructure, new facilities aren’t yet showing in their live networks. Further, it’s worth questioning just what these new datacenters will do once they come online. Let’s take a look.

To check out what the cloud vendors are doing, it’s helpful to review their current infrastructure. For context, hyperscaler networks are generally described in terms of worldwide regions and availability zones. Regions are geographic locations in which services are available. A country may have one or multiple regions within it, though a region typically spans just one country. Availability zones are networked datacenters within a region. Availability zones can be one or more physical datacenters, or datacenters within which virtual resources comprise availability zones.

Though the cloud vendors differ in the fine points, it’s possible to assemble a general picture of their infrastructure by counting regions, availability zones, and countries served, as in the chart below:

Source: Company reports

Hyperscalers can extend availability zone services with points of presence (PoPs), which may account for some of the differences in countries served.

Also, hyperscalers offer services geared to reducing latency at the network edge. AWS, for instance, offers Local Zones consisting of network, storage, and compute resources placed at specific locations within a region to improve application performance. Similarly, Azure Edge Zones are Azure installations in metros, industrial parks, or similar locations to reduce latency or meet data residency requirements. Google Distributed Cloud brings Google Cloud functions to local installations of hardware and software, and air-gapping provides added data security. Oracle’s Cloud@Customer brings the vendor’s public cloud service capabilities into customers’ on-premises datacenters.

Hyperscalers' Big Plans

The networks cited above should get sizable extensions when the public cloud providers realize their ambitions for new datacenters. Below is a capex list, plus a list of announced buildouts, their costs and locations:

Source: Company and media reports

There’s a big gap between the existing regions and availability zones of the hyperscalers and what they have planned, though many projects are expected to come online within the next couple of years. Until they do, it will be tough to gauge exactly how the hyperscalers are benefiting from their gargantuan buildouts.

AI Availability Zones

Another trend is building among cloud hyperscalers: specifically, that of offering datacenter resources dedicated to AI. AWS, for instance, has launched so-called AI Zones with HUMAIN in Saudi Arabia and with SK Group in South Korea. These are datacenters equipped specifically with power, cooling, compute, chips (including proprietary ones) and networking resources geared to AI.

Other cloud vendors have similar efforts underway. Microsoft Azure hosts services for AI and HPC within availability zones in the U.S., Europe, Australia, Japan, Southeast Asia, Canada, Brazil, South Africa, the Middle East, and India.

Within selected regions in the U.S., Europe, and the Asia-Pacific area, Google offers pods comprised of its Tensor Processing Units (TPUs) dedicated to AI model training and inference. Alibaba also offers AI-specific resources within regions in China and Singapore.

What's It All For?

The level of aggressive datacenter buildout among the hyperscalers prompts questions concerning what it’s all for. To what extent are the new datacenters geared to customer demand for AI infrastructure versus their own infrastructure needs?

All of the public cloud providers have stated that customer demand is fueling the majority of their increased datacenter buildouts. A substantial portion of the new infrastructure is also aimed at providing faster and more efficient platforms for cloud services, including ones that enable customers to develop AI applications.

Most of the hyperscalers also offer their own AI models, and they are using their datacenter resources to train those. And in the case of hyperscalers such as Meta that don’t offer public cloud services for enterprise use, a majority of datacenter firepower is going to training models, in Meta’s case, Llama.

From all of this, one thing is clear: The hyperscalers aren’t slowing down their datacenter buildouts, despite not showing ROI from all their investments yet. As more datacenters come online, impact on revenues should become evident. In the meantime, a gap between sales and capex will persist.