The Cacophony Of Many Different Server Markets
September 13, 2021 Timothy Prickett Morgan
Considering how skittery the global economy is, how wonky the world’s supply chains are, and how capricious spending by the big public clouds and the hyperscalers can be (and how much the top eight server buyers in the world make up of the overall market), the fact that the server market only had a 2.5 percent decline in the second quarter is not a surprise.
It’s just the normal noise in the data, and perhaps, if IDC is right, a slight shift away from using two-socket servers with a modest number of cores to single-socket servers with a larger number of cores. For any given amount of capacity increase, depending on the pricing sweet spots in any product line, it could take more single-socket nodes to cover the same amount of compute, memory, and I/O bandwidth or it could be the same. So if this trend persists, we should see node counts on the rise and revenues dropping or holding flat.
“Broadly speaking, server market performance was muted in the second quarter as the market shifted slightly towards single socket server configurations,” Paul Maguranis, senior research analyst for Infrastructure Platforms and Technologies at IDC, explained in a statement accompanying the server stats compiled for the quarter ended in June. “While servers purchased directly from ODMs declined year over year, some past backlog recovery within the hyperscale datacenter community contributed to a large jump in this segment when compared to the first quarter of this year.”
AS/400 and IBM i shops are no stranger to single-socket servers, of course. They have been using them since the early days of compute, of course, like all IT organizations that had infrastructure before SMP and NUMA processor clustering took off in earnest in the 1990s. The funny bit is that these days, many IBM i shops have a single core of compute, whether or not the system they have a machine that is capable of hosting a single socket (like the Power S812, Power S814, or Power S914 machines) or a pair of sockets (like the Power S822, Power S824, Power S922, and Power S924. That seems crazy, but the back office workloads are not getting compute heaving like the application layers and interactions across machines or the analytics stack are.
This is, in fact, one of the reasons why the non-X86 server business, which is dominated by IBM iron these days with a smattering of X86-like Arm servers and a smidgen of other architectures like Itanium, does not grow at the same rate as the X86 server business. The former grows more or less with gross domestic product, the latter with the increasing sophistication of the application stack – and now machine learning is being shimmed into that stack – and the exponential growth of data and the need to chew on it to get some sort of insight to drive the business. That is not to say that there is not a big portion of the X86 server business that looks and smells like the IBM i business in terms of workload. For many enterprises, a four-socket or eight-socket X86 machine is their main frame, in the original sense of that word.
You will notice the general and slight uptick in sales for non-X86 servers despite the fact that IBM is at the tail end of both the Power9 and System z15 server product lines.
The reason is that Amazon Web Services is using its homegrown Graviton and Graviton2 Arm server processors in a significant percentage of its server installations, and we think Microsoft, Oracle, and Tencent are also deploying a fair number of Altra Arm server chips from Ampere Computing as well. It’s enough to fill in the gaps, and so is the significant – and growing – revenues that Inspur is getting from its Power server line, which is distinct from IBM’s own Power Systems.
In the June quarter, X86 server revenues fell by 2.2 percent to $21.4 billion, while non-X86 servers were off 4.5 percent to $2.3 billion. We don’t know how many non-X86 server machines shipped, but the X86 iron utterly dwarfs it. AS/400 and IBM i shops have always bought a lot of X86 servers, of course. The AS/400 was born seven years after the PC era started in earnest, and the client/server revolution had impacts on all back office system architected. It would be perfectly reasonable to assume that AS/400 and IBM i shops spend a fair amount of money on such gear, and that they are conservative about what they spend on X86 iron and that, on average, their relative spending matches the market overall.
Overall, across all architectures, OEM and ODM server makers shipped 3.23 million machines, up one-tenth of a point from the year ago quarter.
This was against a backdrop of $23.64 billion in sales across all architectures, down 2.5 percent. That said, shipments and revenues were up sequentially, which is always good.
In the chart above, we have cast the IDC shipment data all the way back to Q1 1999 and taken the quarterly revenue figures for those 23 years and adjusted them for inflation to show what that revenue would feel like in the past using 2020 dollars. After the Dot-Com Bust in the early 2000s, you can see the affect of the IBM mainframe and Power Systems product cycles as well as those big iron machines from Sun Microsystems, Hewlett-Packard, and a few others in Japan like Fujitsu, Hitachi, and NEC. Things got wonky during the Great Recession, which also saw the rise of the hyperscalers and cloud builders, which drove shipments but which did not drive revenues. The shift from OEMs to ODMs by these customers removed revenues from the market, and revenues did not really take off until 2017, when the hyperscalers and cloud builders started to radically expand their workloads and capacity and drive more than half of shipments and just shy of half of revenues these days.
There really are two or maybe three server markets, not one. One way or another, companies and consumers end up using all of this capacity.
Finally, here is the play-by-play for the past six quarters by vendor:
Because Cisco Systems did not make the top five, we have done our best to estimate its revenues, which are not given out in the public figures put out by IDC; this is shown in bold red. It is important to note that the ODM Direct category is a proxy for growth among the hyperscalers and cloud builders, but that many of the OEMs (notably Inspur and Lenovo as well as Dell and Hewlett Packard Enterprises to lesser degrees) also sell gear to hyperscalers and cloud builders, so it is not a perfect reckoning of the impact of the shipments and sales to the Super 8 that collectively dominate the consumption in the server market.