Public cloud providers face rising capital expenditure pressure

Capital expenditure continued to accelerate for the three leading public cloud providers in Q3 2017. Amazon reported its property and equipment acquired under capital leases costs increased 62% in the last quarter, bringing the total for the first nine months of 2017 to US$6.9 billion, against US$3.7 billion in the same period in 2016. Microsoft’s capital expenditure, including assets acquired under leases, rose 29% to US$2.7 billion, though it fell 18% on a sequential basis. Google’s capital expenditure increased 39% to US$3.5 billion. Collectively, Amazon, Microsoft and Google reported capital expenditure of over US$8 billion for the quarter, up approximately 40% year on year. Not all costs are related to data center buildout, but it does account for the majority. The combination of this expenditure and ongoing operating costs, including energy, bandwidth and labor, highlight the scale of infrastructure investment. The total increase in costs is just in line with the overall growth of cloud infrastructure services revenue, which Canalys estimates grew 43%.

The level of data center investment is expected to continue rising for the foreseeable future. Geographic expansion remains a key priority. AWS currently has 16 regions, with a total of 44 availability zones. Each availability zone is hosted in a separate data center, which is either its own facility or residing in a colocation data center. It has announced a further six regions (Bahrain, China, France, Hong Kong, Sweden and a new US-East GovCloud for Federal customers) and 17 availability zones. Microsoft’s setup is different, with 36 regions around the world, each with its own data center. It announced plans for six additional regions: two each in Australia, France and South Africa. At the start of the year, Google announced an aggressive expansion plan, with 10 new regions in Australia, Brazil, Canada, Finland, Germany, India, the Netherlands, Singapore, the UK and the United States. This will bring Google’s footprint to 17 regions, with 52 availability zones. Further expansion into Africa, the Middle East, Latin America and Southeast Asia is anticipated, as demand grows for low-latency services that maintain data sovereignty.

An increasing proportion of the companies’ data center capital expenditure will have to be allocated to infrastructure refresh, as the number of servers under their control grows. Refresh cycles are generally in line with major processor releases, on average once every three years. The launch of Intel’s Skylake processors will catalyze a new wave of upgrades, which provide more memory to support compute-intensive workloads, such as 3D rendering, financial and scientific modeling, and data analytics. Amazon, Microsoft and Google, along with the rest of the super seven cloud providers, had early access to these processors, at least six months before commercial availability. Google was the first to have them running in its data centers, and commercially available at the beginning of June, with customers able to create new virtual machines with Skylake across all Compute Engine’s instance types. Microsoft only added Skylake processors to Azure at the end of October, with the launch of its Fv2 instances running on Intel’s Xeon Platinum 8168. This leaves AWS as the only top four public cloud provider without support for a Skylake-based instance, after IBM launched its version in July. AWS is expected to release its offering later in the year.

Competition to deliver the best-performing services and greatest geographic reach will mean capital expenditure is unlikely to slow any time soon. The providers cannot be seen to be lagging technically, or not have enough capacity. Microsoft customers have already experienced capacity issues in the UK this year, with demand exceeding server deployment. Customers reported that they had to either wait for processing power, memory and storage to become available or use another region, which creates data sovereignty issues. Overall, the big three are now buying more servers per quarter than either HPE or Dell EMC sell worldwide in the same period. But they will have to find a balance of maintaining infrastructure investment to meet customer requirements while managing costs. Consequently, they are now relentlessly pursuing capital and operating savings by optimizing data center designs and developing their own hardware configurations for ODMs to build. For example, more than 90% of the servers Microsoft deploys use its Open Cloud Server design, which it claims to be 40% cheaper, use 15% less energy and reduce servicing time by 50% compared with off-the-shelf servers. It has released its latest server design, “Project Olympus”, which uses Skylake processors, to the Open Compute Project community.

There are signs that memory and SSD supply restraints are easing and costs starting to stabilize and even come down. This will help the public cloud providers’ costs, though they already receive large discounts for buying in bulk. But investors should be scrutinizing their balance sheets more closely and asking how they can sustainably maintain current spending levels, not just for the next quarter, but for the next three to five years, and possibly beyond. Microsoft and Google will be less scrutinized in the short term, due to their positive cash flows and large reserves. The real challenge will come when use and capital expenditure exceed revenue growth, which is highly possible, especially as competition drives down prices, and new lower-priced services and discounts cannibalize existing higher-priced offerings. Signs of this trend were mentioned during Amazon’s recent earnings call, but it is also an industry-wide challenge. The emergence of Alibaba and Tencent, as they expand beyond China, as well as Huawei, will add further pressure.