Server Buying Cools, But It’s Cool – Don’t Panic
June 10, 2019 Timothy Prickett Morgan
When a market is comprised of hundreds of thousands of customers, things tend to level out and are a lot more predictable than when there are relatively few customers. Before the public clouds took off a decade ago and before the hyperscalers created such large infrastructures to support billions of users running their applications, server buying was a lot smaller and it was also more predictable. Things tended to grow slowly, methodically and they also took time to slow down because not everyone felt an economic decline or a transition to a new system architecture at the same time.
That is no longer so with the modern server business. Enterprises are offloading some of their compute needs to the public clouds, and in other cases they are employing services provided by the hyperscalers – email, collaboration, and so on – instead of hosting them in their own datacenters. The hyperscalers and cloud builders are at the front of the line ahead of any new server chip generation, and they tend to buy aggressively ahead of the formal launches by Intel and AMD, and if the most recent quarter is any test, they are slowing down server purchases as they await the right time to invest in future chips from those two companies.
This is just the way it is now for a big portion of the market. But the underlying IBM midrange still follows its own path and is thankfully continuing to grow even if Big Blue overall is seeing a slowdown in server revenues as the System z14 mainframe upgrade cycle is winding down after being in the market for nearly two years. The System z15 processors will be etched using the same refined 14 nanometer processes from Globalfoundries as will be used with the Power9’ (that’s “power nine prime”) that is expected to come out this year, possibly with memory and I/O enhancements aimed at Linux machines. There is no word on when the z15 processors will debut, but there is generally somewhere between two and a half to three years between mainframe chip generations, so IBM is anywhere from six to 12 months away from another System z upgrade bump. The Power10 processors are being etched by Samsung using a 7 nanometer process that will also be used on the future z16 processors for mainframes. Power10 chips are expected in late 2020 to early 2021, and if history is any guide, we can expect the z16 chip around then, too.
Suffice it to say, there are a lot of moving parts that drive server shipments and revenues each quarter, and the market is in many ways more segmented and more complex than it was one or two decades ago.
For instance, the midrange server market is growing again, but it is a very different one from the classic days when IBM, Digital Equipment, Hewlett Packard, and a handful of others selling proprietary minicomputers were coming up against Sun Microsystems Unix iron as well as the Unix gear from their own companies as well as those longtime rivals. In the first quarter, IDC says that the volume server market – meaning machines that cost less than $25,000 – grew revenues by 4.2 percent to $16.7 billion, a little slower than the overall pace of the market and comprising 84.3 percent of that market. But the midrange, which means machines that cost between $25,000 and $250,000 in the way that IDC dices and slices the server space, grew by 30.2 percent to $2.1 billion, while high end iron costing more than $250,000 slumped by 24.7 percent to $976 million. The server market is starting to look more like a pyramid again and less like an hourglass. That midrange expansion, we think, is due to strong sales of both Power and X86 iron in this class of machine as well as to hybrid CPU-GPU configurations aimed at machine learning and HPC simulation workloads that can cost anywhere from $50,000 to $125,000 with four to eight GPUs in the systems. It would take sales of fewer than 10,000 of these machines to drive that incremental 30.2 percent growth, just to give you a sense of it. We are not saying this is what happened, but a few thousand of these machines can really boost the midrange, and we do believe this is the dynamic in the market. SAP HANA is also driving sales of bigger iron by IBM, HPE, Dell, and Lenovo, who dominate sales of such iron.
IBM has been impacted, no doubt, by the System z sales cycle, but we think there is another factor at work that relates to Power Systems. Back in August 2014, when it was clear that American IT vendors were going to have a much tougher time selling in China because the Middle Kingdom wanted indigenous technologies for its mission critical platforms, IBM inked a partnership deal with Inspur, one of the largest sellers of systems to enterprises and hyperscalers in China. Inspur had been building Itanium-based systems for mission critical workloads like ERP systems and databases, but that was not going to work out with Intel pulling the plug on Itanium shortly thereafter. Inspur joined the OpenPower Foundation and started designing and building systems based on the Power8 processors; it has subsequently launched Power9 iron. All of the sales that in the past IBM got in China for Power Systems directly are now more or less going to Inspur, with IBM getting an untold cut of the action after the fact. This use of Power chips in mission critical iron is fueling Inspur’s growth in part. (It also sells a lot of iron to Alibaba, Tencent, and Baidu, we know.)
In the first quarter of 2019, here is how the major server OEMs and the collective of ODMs who manufacture directly for the hyperscalers stacked up against each other:
As you can see, in the year-ago period, IBM and Inspur had roughly the same revenues – we had to estimate Inspur’s sales in Q1 2018 because it did not make the top five then and we had to do the same estimation for IBM in Q1 2019 because Big Blue, shockingly, did not make the top five – but now Inspur is considerably larger. We suspect that IBM will eventually rebadge Inspur machines to sell in its own catalog outside of China and will move away from its partnership with Supermicro for entry Linux machines aimed at HPC and AI workloads.
The X86 platform has reached its peak for sure, with 99 percent of shipments and 93.4 percent of revenues, but it also looks like sales of non-X86 platform are stabilizing at around $1.8 billion per quarter, give or take depending on the point we are at in the System z cycle, since 2016. IBM tends to bring in around 70 percent of that non-X86 revenue, again give or take depending on System z.
That’s about as good of a position as you can hope for with a non-X86 platform at this point, but there is the hope that IBM can compete at the high end better – particularly with AI and HPC systems that have GPU acceleration as well as those aimed at SAP HANA. We think that the traditional IBM i and AIX midrange can grow a little – as IBM i has done in the past six quarters – and that is about all you can expect given the nature of the customer base. IBM’s sales of Power Systems that run Linux (those AI, HPC, and HANA machines are all Linux) are growing much faster and AIX is pretty much flat to down in any given quarter. There is also an outside chance that competition between Intel and AMD will drive down X86 server prices in the coming years, pulling that revenue curve down even as shipments continue to grow, and that memory and flash will also get cheaper, pulling both of those curves shown above down a little.
And then there is always the prospect of Arm perhaps taking a 5 percent or maybe even a 10 percent bit out of the server market if and when hyperscalers and cloud builders step up and push Arm iron. Amazon Web Services and Microsoft Azure are monkeying around with Arm iron internally and AWS is selling raw capacity on its homegrown “Graviton” Arm chip, but thus far Arm is just noise in the data with the exception of a few HPC systems and a bunch of proof of concept boxes. These Arm sales all go into the non-X86 bucket, and they weaken Intel’s hegemony in the datacenter.