The datacenter is a tough battleground, with vendors at every part of the stack pushing and pulling against each other to try to win business. But for Intel, which has vanquished system incumbents from the glass house and which has kept upstarts at bay as it expands its own footprint into storage and networking, ironically enough all of that competitive pressure ends up generating the same revenues and profits for the chipmaker.
And it doesn’t look like that will change anytime soon. But we hear grumbling, from the hyperscalers and cloud builders that Intel is most proud of serving, that they want alternatives when it comes to compute. They always want a better deal as much as Intel wants to maintain its impressive revenue and profit growth in the datacenter. While we think something has got to give, it doesn’t have to give just because people want it to. The technology and the money have to line up.
As far as Intel’s Data Center Group is concerned, it looks like next year will be another great one, despite a slight blip this year thanks to an unanticipated slowdown in spending by enterprise customers. This part of the market is hard to call, or more precisely, is as difficult to call as gross domestic product is for the major economies of the world. Intel’s top brass held its Investor Day briefing recently – concurrent with the SC15 supercomputing conference, in fact, which is why we are circling back now to analyze what Intel divulged – and Stacy Smith said that part of the problem is a slowdown in the global economy.
“Several parts of our business correlate to GDP,” Smith explained. “And while we saw strength in lots of places, every one of the places where you see a strong correlation to GDP, we saw some level of weakness: The business portion of the PC market, the enterprise portion of the datacenter, the industrial segments of the IoT market.” A year ago, when Intel was looking ahead to 2015, it expected that the emerging economies (especially China and India but also Easter Europe and Brazil) would see GDP grow at around 4.3 percent compared to 2014, but now the latest stats show it coming in at around 3.3 percent. China declined much more than anyone thought, said Smith. In the mature economies, which still represent a lot of IT sales, the forecast was for GDP to grow 2.1 percent across those economies (dominated by North America, Western Europe, and Japan), but it has shifted down to around 1.8 percent growth. These downshifts in growth add up over a global economy that generated $77.8 trillion last year, removing untold billions from the coffers of IT vendors.
Despite this unpredictable slowdown in the global economy, the growth in spending on compute in the consumer-facing hyperscale applications and public cloud builders has more than offset the slowdown in enterprise IT spending (via Intel channel partners for the most part), and both Intel’s revenues and profits have grown, as the chart above shows for the past three years. The Data Center Group business grew 18 percent in 2014, to $14.4 billion, and operating income was $7.3 billion. Intel has managed to bring just a hair under half of its Data Center Group revenues in as operating profits so far in 2015. Smith did not provide a forecast for 2015, since the fourth quarter is not yet over, but we expect Intel to have somewhere around $4.5 billion in sales for the Data Center Group and maybe $2.4 billion of that coming in as operating profit after the basic bills for that business are paid, yielding $16.2 billion in sales and $8.1 billion in operating profit for the 2015 year.
We hear that hyperscalers have slowed down their purchases a bit in recent quarters and it only takes one or two deals to push into 2016 – when we expect for new “Broadwell” Xeon E5 v4 processors to come out – for Intel to not make the numbers we are projecting. A lot depends on if Intel is already shipping Broadwell Xeons into cloud and hyperscale accounts as 2015 comes to an end, and we have not heard that it is or it isn’t. We did not hear of any HPC systems shipping early with the future Broadwell Xeon E5s at the SC15 conference, either, and if the chips were impending and Intel was happy with its 14 nanometer yields, we would have expected to see that. None of this means that there have not been early ship Broadwell Xeon E5 customers, or that there will not be in January or February. It just means Intel was not talking about it.
Smith’s revenue and profit chart for Data Center Group only shows three years, and it does not really give Data Center Group its due as a tremendous growth engine for Intel and, to be blunt, the part of Intel that is making up for the demise of the PC. (Good thing all of those smartphones and tablets are talking to datacenters powered by Xeon chips.) Here is what a longer trend since the belly of the Great Recession looks like:
As you can see, excepting a stall when the “Sandy Bridge” Xeon E5 chips were delayed in 2011 and 2012, revenues and profits have grown. This was also the time when RISC and Itanium processors supporting Unix took a big hit – particularly in China – that they have not recovered from as yet and, unless something radical happens with OpenPower (as IBM and its partners are hoping), probably will not. At least not on the scale of growth that Intel’s Data Center Group has seen in recent years, despite the enterprise spending hiccups.
Here is what Data Center Group’s CPU revenue looks like if you dice and slice it into the three main segments:
It is important to note that this is for processor sales only, and does not include motherboards, chipsets, network adapters, and flash memory. The latter two products are not even part of Data Center Group, although we would argue that Intel should segment its networking and storage sales in PC and datacenter buckets as it does for compute. (In the chart above, networking and storage refers to the use case for Intel Xeon and Atom processors, not for networking and storage products themselves like flash drives or Ethernet controllers.)
Intel had been expecting Data Center Group to post an overall 15 percent revenue growth rate for all of 2015, but cut that forecast earlier this year when the enterprise spending slowdown became apparent. Based on the latest data, Smith reckons that spending on processors for enterprise servers will be slightly down compared to last year, cloud spending will be up more than 40 percent, and the rest of the pie, which includes HPC, workstations, networking, and storage, will grow around 10 percent. The cloud business (which includes the hyperscalers as we talk about them here at The Next Platform), grew by more than 40 percent last year, too, by the way.
The reason why enterprise and cloud are diverging so much is dirt simple. The number of transactions that we are processing to do all of our aggregate spending as consumers and businesses is going to grow more or less along with GDP, and a big chunk of that growth will be masked by the move from private datacenters to public clouds for parts of the application stack. The amount of time we devote to consumer-facing hyperscale apps – search, email, chat, video, social media – seems to have no bounds as yet and is growing as these applications get more users and they do more on these sites.
It really is that simple. And, we think, at some point the growth runs out on “the cloud” as Intel uses the term when everyone in the world has access to such software through client devices and a reasonable Internet connection. It may take a while for that to happen, but nothing will grow forever, and certainly not at 40 percent plus rates. In the meantime, Data Center Group’s very growth is dependent on this cloud market, which is why you see the company designing special chips for so many of the hyperscalers and cloud builders and doing the same to the telecom firms who are sick of expensive appliance-style infrastructure that costs too much compared to raw software on cheap iron.
As you can see, Intel is a lot less dependent, at least in terms of CPU revenues, on the enterprise than it was only a few years ago:
“If you think about the Data Center Group for 2015, we are seeing the enterprise segment actually down and yet we are still able to grow in the low double digits,” Smith explained. “If you go back three years, we would not have been able to generate that level of growth in a year where the enterprise was down.” Looking ahead, Smith said that Intel was anticipating Data Center Group to grow in the mid-teens, as it said last year, with operating profit growth in the “low double digits,” which is another way of saying there is some pressure on profits. The networkers are fighting back, and we think the hyperscalers are playing with ARM and Power servers – perhaps only as a way to keep Intel on its pricing toes. Smith also cautioned that “it is not an annual steady march” and that “there is going to be pluses and minuses,” as happened with the enterprise in 2015.
We think enterprise spending on server-class CPUs could hit a big downdraft if cloud computing (meaning using someone else’s compute and storage utility) really takes off. All it would take is slightly better security and radically reduced data movement costs for this to happen, and then a big chunk of upgrades that might have gone into private datacenters using relatively expensive servers bought in tensies or hundredsies by tens of thousands of large enterprises will be replaced by very cheap machines bought in the millionsies by a handful of players. It is hard to say what this might do to Intel revenues and profits. The 22 nanometer processes used to make the “Haswell” Xeon families is the best price/performer Intel has ever had, as Intel explained at the same investor meeting and as we already reported on, but its 14 nanometer processes have not ramped as well and cost more than the 22 nanometer processes. So costs are expected to rise for the Data Center Group as it transitions to 14 nanometers, and you can see now why Intel did not launch Broadwell Xeon E5s in September as many expected and is putting it off as long as it can. Intel is not under huge competitive pressure in the datacenter – yet. But a need to keep prices high while boosting performance only modestly with Broadwell could leave a gap for Power and ARM chips – and maybe even AMD’s Zen Opterons – to get some traction.
Here’s the thing. The hyperscalers and HPC centers that are most likely to try something else are the same customers that Intel is depending on for growth. Cloud builders have to provide X86 iron for a lot of their customers, since they are running the customer code, not their own like hyperscale and HPC shops, and for a lot of enterprise customers who want to run in the cloud, that means running Windows Server, which is not supported on Power or ARM chips. Companies selling SaaS products can do whatever they want, just like Facebook and Google, in terms of their underlying architecture, but they are still pretty conservative about taking the risk of moving to a new architecture.
Data Center Group general manager Diane Bryant gave a little more detail on how the cloud business breaks down as well as a forecast for growth in the different segments for that business between 2015 and 2019:
And here is a chart that shows how CPU shipments break down across the cloud providers and hyperscalers:
The interesting bit there is that for the clouds that are supporting enterprise workloads, Intel reckons that half of its processor shipments are being used to support applications moved to the cloud from private datacenters and the other half are for new workloads. Of the top seven hyperscalers – Google, Amazon, Facebook, Microsoft, Baidu, Tencent, and Alibaba – five are in the early ship program and five have custom processors. (It is amazing to us that all seven do not.) Four say they will sample Intel FPGAs in the first quarter of next year, and all seven have moved up the stack to buy higher performance and higher priced processors. Two are playing around with Intel’s silicon photonics, which are a key to its composable systems as well as datacenter networking strategies.
The one thing Smith did make clear is that even if Internet of Things, non-volatile memory such as flash and 3D XPoint, and software and services are important and sometimes growing parts of the overall Intel business, they are not particularly profitable. Take a look:
Add these three together, said Smith, and each of the businesses is north of $2 billion and are on their way to $8 billion, and with the addition of the Altera FPGA business (possibly next year) will be approaching $10 billion. The memory business will take some profit hits in 2016, he warned, as competition increases in the NAND flash market and Intel starts up its memory factory in China.
Speaking of 3D XPoint, Bryant confirmed at the investor meeting that this new class of memory would debut with the “Skylake” Xeon E5 v5 processors, which we expect in 2017 and which we told you about back in May. This 3D XPoint memory, which will also be sold by development partner Micron Technology, will start sampling in SSDs from Intel in 2016 to be followed by DIMM-style memory sticks in 2017. We wonder if Intel will make this 3D XPoint memory applicable to servers other than its own Xeons, and if it doesn’t, will Micron?