Is This A Server Slowdown, Or Increasing Efficiency?

If you happen to believe that spending on core IT infrastructure is a leading indicator of the robustness of national economies and the global one that is stitched, somewhat piecemeal like a patchwork quilt. From them, then the third quarter sales and shipments of servers is probably sounding a note of caution for you.

It certainly does for us here at The Next Platform. But it is important, particularly if we have in fact hit the peak of the X86 server market as we mused about three months ago, to not get carried away. A slowdown in spending amount some hyperscalers and cloud builders at the same time that enterprises are pulling back a bit is alarming, for sure, particularly as Intel is getting ready to launch its “Skylake” Xeon processors in the middle of next year and IBM is getting ready to follow suit with Power9 chips in the second half of 2017. AMD will be rolling out its “Zen” Opterons next year, too, a bunch of new ARM server chips are coming down the pike, and Nvidia will be debuting its “Volta” GPU accelerators. Various FPGAs will be coming to market or ramping from Intel and Xilinx as well. To say that 2017 is going to be an exciting year for compute is an understatement. With all of the custom processors also being developed for parallel processing and deep learning, to call it a Compute Cambrian Explosion is not hyperbolic at all. Even if Fujitsu has stopped development on Sparc64 chips and Oracle is rumored to be preparing to mothball Sparc chips or the Solaris Unix operating system, or both.

Let’s talk about the numbers for the third quarter and try to sort this out, and we will start with what the box counters at Gartner think happened, and then we will look at IDC data. Both count the server market a bit differently, and across the two we can get a sense of what is happening. (Gartner looks at spending from the end user perspective, and IDC looks at vendor revenues into their channels plus their direct sales. In a perfectly balanced channel, these would be the same number, but the channel doesn’t work that way as it acts as a buffer between customers and suppliers.)

According to Gartner, in the third quarter companies consumed $12.71 billion in server gear, down 5.8 percent from the year ago quarter, with shipments down 2.6 percent to $2.69 million units. While Hewlett-Packard Enterprise remained the top server seller in the world, selling off its Chinese server business to the H3C unit of Tsinghua University had an adverse impact on its direct shipments and therefore its direct revenues. As a consequence of the server downturn and the selling off of this business in China HPE’s shipments were down 19.5 percent to 493,268 in the quarter, and despite higher average selling prices for machines, HPE’s revenues were pulled down 11.8 percent to $3.25 billion. That gives HPE 18.3 percent shipment share and 25.5 percent revenue share.

Here is how Gartner believes the revenues stacked up globally among the top five:

And here is how they stacked up based on shipments:

There are a couple of interesting things to note. First, Dell is closing in on HPE in terms of shipments thanks to the H3C spinoff, with 452,383 machines shipped in the third quarter, but HPE still has a considerable revenue lead with Dell only converting those machines to $2.23 billion in revenues. Like Dell and HPE, Lenovo also shrank revenues faster than the market at large, but by less than its two bigger rivals and its revenues kissed $1 billion driven by 228,097 units shipped. HPE’s units shrank five times faster than the market, Dell by four times, and Lenovo by two times. All three are being impacted by the success of Inspur, Hauwei, and Sugon in China and, increasingly, in Latin America. Inspur and Hauwei broke into the top five shippers a few quarters ago and are now solidly going to stay there unless something really dramatic happens with IBM’s Power Systems business. IBM has not shipped more than a 100,000 units of Power-based servers in a year in a very long time, much less in a quarter. So unless someone like Google and Rackspace and a few HPC centers and big enterprises go nuts for Power9 machines, Huawei and Inspur are going to keep their shipment rankings. Cisco Systems is still growing revenues thanks to enterprise customers buying rich configurations, but it does not peddle as many nodes as the two big Chinese server makers. Cisco grew its revenues 5 percent to $929.4 million in the period, bucking the trend among the top tier server OEMs, much to its credit.

IBM, which no longer has a System x Xeon server business and which is at the tail end of the Power8 system and System z13 mainframe generations, is having trouble pushing gear. Revenues were down 33 percent to $889.7 million. Three years ago, in the third quarter of 2013, when IBM still had an X86 server business, it booked sales of $2.82 billion. A year later, as the selloff to Lenovo of its System x division was being done, sales slipped to $2.32 billion, and a year after that, in the third quarter of 2015, they fell a lot further to $1.33 billion. Over that time, IBM’s revenue share of the server space has fallen from 22.8 percent in 2013 to 7 percent in the most recent quarter. And this is from a Big Blue that aspires for the Power-based server business, including its own and that of its OpenPower partners, to have at least 10 to 20 percent shipment share of the aggregate global server market. IBM has a faction of a percent of global shipments today, and probably ships about two to three as many servers as the ARM collective does, just to give you perspective.

The interesting bit in the Gartner numbers is not the slow and steady decline in the Unix server business, which has been proceeding at pace for the past decade and a half after peaking in the dot-com boom, but that Gartner shows the X86 server business actually declined in both revenues and shipments, the first time that has happened since the Great Recession when everything went into the ditch in the server world. Unix system shipments were down 37.7 percent to 14,485 units and revenues fell by 38 percent to $658.2 million. But the powerhouse of the datacenter, the data closet, and under the desk, that X86 machine, had a 2.3 percent shipment decline to 2.67 million units, and revenues fell by 1.6 percent to $11.34 billion. Among X86 suppliers, only Huawei and Inspur had shipment growth in the third quarter, and only Cisco had revenue growth.

The X86 server increased its share of the server revenue pie to 89.2 percent and the server shipment pie to 99.3 percent. We have said it before, and we will say it again: These kinds of numbers, for no technical reasons, cannot hold with an Intel that has 50 percent gross margins in the Data Center Group. That kind of market share and profit will bring ever more intense competition, and Intel can only do what it is doing – become its own best competition.

Would it not be hilarious if Intel commercializes ARM servers best someday? Stranger things have happened, and will happen. Particularly if any big part, or several parts, of the world slip into recession in 2017. It has been a decade, after all, and the Dow Jones is at a peak along with X86 server sales, and nation after nation is turning inwards and getting protectionist. Instead of a housing crisis, an identity crisis could cause the next recession, and Intel might not fare as well this time as it did when the “Nehalem” Xeons launched into the gaping maw of the Great Recession. We are all more sophisticated when it comes to infrastructure these days.

Over at IDC, there is consensus that the server market softened a bit in the third quarter. The other professional box counter says that factory revenues for servers across all vendors was off 7 percent to $12.5 billion, and that shipments fell by 4.6 percent to 2.38 million. The healthy “Haswell” Xeon server ramp in 2015 has made for some tough compares, and so has a slowdown among hyperscalers, cloud builders, and enterprises.

IDC tracks servers by market segment as well as by processor architecture, and in the third quarter, volume servers, which cost $25,000 or less a pop, had a revenue decline of 4.9 percent, to $10.3 billion. Midrange machines, which cost between $25,000 and $250,000, had a 4.1 percent decline to $1.1 billion. High end servers, which include IBM and other mainframes as well as big RISC, Itanium, and Xeon NUMA servers, declined by a staggering 25 percent to $1.1 billion.

By architecture, X86 machines dominated, with IDC saying that machines using mostly Xeon processors accounted for $11.2 billion in sales, down 3.1 percent, and shipments came to 2.36 million units, down 4.3 percent. Machines using other processor architectures, including Itanium, Power, Sparc, ARM, and a smattering of others, drove $1.3 billion in revenues, but fell 30 percent year on year. X86 iron, by IDC’s numbers, had a 99.2 percent shipment share and an 89.6 percent revenue share.

The vendor breakdown by IDC is similar to that of Gartner, with the exception that IDC breaks out ODM direct sales to companies like the big hyperscalers and cloud builders. Take a look:

Here is the interesting bit of the IDC data. The ODM server makers like Quanta Computer, WiWynn, Stack Velocity, Tyan, and a few others posted revenues of $1.3 billion in the quarter against $11.2 billion for OEMs, accounting for 10.3 percent of total server revenues worldwide. In the second quarter of this year, ODMs had $1.9 billion in sales against $9.7 billion for OEMs, or 13.9 percent of the revenue total. If you look at the trailing twelve months, OEMs have $5.3 billion in server sales, up 22.1 percent compared to the period running from the fourth quarter of 2014 through the third quarter of 2015. During that same time, all other server makers saw a 5.7 percent decline to an aggregate of $53.6 billion in the trailing twelve months.

But don’t get ahead of yourself. The ODM data above does not include sales of custom iron made by Dell, HPE, Lenovo, Inspur, Sugon, and other OEMs. So ODM does not directly equate to hyperscaler plus cloud builder. What we think this shows, by the way, is that the hyperscalers and cloud builders that use ODMs are increasing their spending (Google, Facebook, and Amazon Web Services are the big ones) while those that are buying from the custom server divisions of the OEMs (think Microsoft Azure and its strategy of working with Dell and HPE to build a lot of its iron) are apparently slowing down. A slowdown in spending by Microsoft ahead of the Skylake Xeon launch and as it creates its new Project Olympus server platform that is designed for these processors as well as GPU and FPGA accelerators, stands to reason. You would wait for all this new compute in a new server design, too, if you were trying to build out massive datacenters like the ones in Quincy, Washington that we recently visited.

There are a few other factors at work here. Just like there was a virtualization downdraft when VMware ESXi, Microsoft Hyper-V, and Red Hat KVM took off in the enterprise datacenter just as the Great Recession started, we think that cloud-style computing, with sophisticated provisioning, orchestration, and management, is causing a bit of a downdraft in compute capacity. It is not that the capacity is going down – it absolutely is not – it is just that it is not growing as fast as it otherwise might as companies get more efficient. AWS and Microsoft have both bragged about how they could slow down spending because they could make better use of the compute, storage, and networks they buy, and enterprises are learning tricks, too, as are service providers and telcos.

There is another factor in this pressure on server shipments and revenues. It is more profitable to sell higher level PaaS or SaaS stuff than raw IaaS. The cloud providers that have more than IaaS can basically charge for the underlying infrastructure and then a premium for the service that runs on top of that. Companies that build their own services will, as they did in their datacenters, overprovision their capacity because this is human nature, but the cloud providers will not do this above and beyond the overprovisioning they already do. The net effect is that they buy less infrastructure than they might otherwise and they squeeze more work – and more money – out of that infrastructure, too.

Finally, we think that we are witnessing the effect of a substantial shift of workloads from private datacenters to clouds. This provides a double whammy. OEMs lose the sale to the datacenter and the ODMs and the ODM-like units of the OEMs gain them, and the workloads, as a group across these enterprises, run on less shared infrastructure (more efficiently in the terms of the aggregate capacity used, not in the speed of the individual apps) and therefore the server count is that much lower than it might otherwise be.

This slowdown is not as simple as enterprises pulling back. This is infrastructure transforming, exactly as we all planned. And maybe, if we are not lucky, the beginning of a recession.