You would be hard pressed to find a more cut-throat market than for selling servers into datacenters. Unlike SMB shops that buy onesies or twosies every several years or so, datacenter customers do more frequent refreshes, are more savvy, and they buy in higher volumes so they expect better deals. They also want high quality machines and support services to go along with it.
The big tier one suppliers have broad product lines and lots of hand-holding, for which customers pay through higher system and support prices, and the original design manufacturers have expertise in high volume manufacturing, but their hyperscale customers usually know as much about the systems as they do (or more) and they support their own software stacks – including tweaked Linux or Windows Server operating systems (the latter in the case of Microsoft Azure). The hyperscalers take on a lot of the system design and software support burdens themselves because they can afford such indulgence.
There is a third option, and one that an increasing number of customers are doing. That is to buy systems from Supermicro, or buy parts from the company and build servers either for themselves or, for reseller partners, for others that they then add services to. Because Supermicro both builds its own servers and facilitates others that do building, the company’s true influence in the server and storage market does not get counted properly in the quarterly server and storage stats that come out from Gartner and IDC. But suffice it to say that the company’s influence in these areas – and a growing one in switching as it peddles complete systems to customers – is larger than you might think and its installed base is bigger than you might expect given its 800 customers. Some of those customers are direct buyers of Supermicro servers, storage, and switches, but many of them are resellers who buy components and build systems that they in turn sell to end user customers.
We are not sure of what the multiplicative effect is, but what we do know is that the combination of the two businesses has worked well for Supermico over the past 22 years and that company founder Charles Liang has no intention of getting out of the components business as he ramps up the production capacity of its server, storage, and switching lines.
“That is our commitment to our partners,” Liang tells The Next Platform. “We like new friends, but we treasure our old friends as well. We work with them very closely and we look forward to continuing this partnership for many decades to come. And our computers need those components anyway.”
“With the contract manufacturers, their bill of materials is slightly cheaper and their manpower costs are slightly cheaper. But customers don’t just care about bill of materials costs – they care about total cost of ownership, and this takes into account performance, performance per watt, and density as well as – and especially – quality and service.”
Only a few years ago, when Supermicro began a shift towards “solution selling” to key customers, its server and storage server business was only a fraction of the size of its components business, which includes a stunning variety of motherboards, various plug-in cards, and system enclosures. In its fiscal 2015 year ended in June, Supermicro’s systems business (which includes servers and storage servers, which are essentially servers that are set up with lots of disk and sometimes flash storage to host storage software) accounted for 313,000 machines and generated $1.21 billion, about 61 percent of the company’s total revenues. Supermicro shipped nearly 4.7 million subsystems and components to value-added resellers, systems integrators, OEMs, and other customers who in turn build systems, and these parts brought in another $778 million in the fiscal 2015 year, which works out to 39 percent of total revenues. Server and storage server revenues, in the aggregate, are up 62 percent in fiscal 2015 against shipment growth of 19.5 percent, but you have to be careful with these numbers because Supermicro is counting server enclosures, not server nodes, and an increasing number of its machines have multiple nodes per enclosure. The subsystems revenues grew only 8.3 percent in fiscal 2015 against shipment growth of 5.1 percent. Overall, Supermicro grew revenues by 35.7 percent to $1.99 billion and nearly doubled its net income to $101.9 million.
While sales of components and completed systems rise and fall based on the Intel product cycle and the timing of deals that Supermicro does with big partners such as IBM’s SoftLayer cloud, the trends are generally upwards. Having broken through $2 billion in sales this fiscal year, Liang has set a goal of breaking $3 billion in sales in fiscal 2017 – and the company is ramping up its production capacity again, as it did to hit $2 billion, to make that happen. There is a delicate balance between funding growth and staying profitable, and Supermicro has walked that fine line expertly, managing to turn a profit every year it has been in business while still growing the company in a very tough market where big players will lose money to win a big deal.
“Our growth has been very strong across many different applications and sectors, but pretty much is still limited by our capacity,” explains Liang, who has aspirations to grow his company even faster and take on the likes of Quanta and Dell Data Center Solutions. “And by capacity, I mean manufacturing capacity as well as sales and services capacity. When our capacity grows, we will focus on big datacenters as well. Today we already support some famous datacenters, but we do not approach others because our capacity is still limited. But in the next few years, we will continue to grow our capacity very aggressively and very soon we will be there to support all of the big datacenters.”
Supermicro was founded in Fremont, California back in 1993, and for many years it made all of its hardware at the facilities in this location. Four years ago, Supermicro built a 1 million square foot facility in Taipei, Taiwan to start designing and manufacturing equipment locally for the Asian market. By being more local, Supermicro doesn’t deal with time zone issues and it does not have to pay overseas shipping costs for gear, which actually was adding up. The Taiwan operations will be expanded to 1.5 million square feet very soon, according to Liang, and he hints that another factory in Asia (which we suspect will be located in China) will be opening up. A year and a half ago, Supermicro acquired an old printing facility from the San Jose Mercury News and will be starting up operations in August; this will be a 1 million to 1.5 million square foot facility eventually, says Liang. Two years ago, Supermicro set up a smaller facility set up in The Netherlands to do system integration for the European market, and this is about to be will similarly be doubled soon.
Don’t get the wrong idea, though. The manufacturing capacity that Supermicro has built up over time has absolutely helped it expand its sales into what it calls Internet Data Centers. Eight years ago, that business accounted for a few tens of millions of dollars per quarter, but three years ago it took off and has gradually climbed to be a very significant portion of Supermicro’s sales. In fiscal 2015, Internet data centers accounted for $349.3 million in product sales for Supermicro, or about 17.5 percent of sales.
The interesting thing is that Supermicro has not needed those big hyperscale deals to grow, and it has carved out a niche – really several niches – for itself by focusing on engineering and time to market.
“The overall market has been growing very slowly, only a couple of percent per year,” Liang explains. “Most of our growth comes from having better solutions, and this is exactly our advantage. The first big advantage we have is time to market. Whenever Intel has something new, Supermicro is pretty much the first to make a solution using it for the industry.”
Being first is not sufficient, but it consistently gets the attention of customers. For the past ten years, as it has ramped up its systems business, Supermicro has also focused on energy efficiency and compute density, and Liang boasts that its system designs can save customers anywhere between 3 percent and 10 percent on energy costs over other systems because of the engineering that Supermicro does. The designs that Supermicro comes up with tend to push the extremes of compute, memory, and I/O density with Intel Xeon processors (a decade ago Supermicro was an innovator with AMD Opteron processors), and the Liang says that it will offer the best “Knights Landing” Xeon Phi systems when Intel launches them later this year. Because of the design engineering, Liang says Supermicro can save customers anywhere from 5 percent to 50 percent on hardware costs. That latter number compares its X10 MicroBlade modular systems with standard rack machines from competitors like Hewlett-Packard, Dell, and Lenovo; we presume the 5 percent is for like-for-like rack machinery of similar configuration.
As an example of the kind of engineering that Supermicro beings to bear, the company has just introduced a 1U server based on “Haswell” Xeon E5 v3 processors from Intel that can have four Tesla GPU or Xeon Phi accelerator cards plugged into it. The system can take the top-bin, 18-core Xeon E5 chips, which burn at 160 watts, and all of the accelerators plug directly into the system, not using flexible cables that add complexity and cost to a system. Liang estimates that the prior server hosting three accelerators had about a 10 percent market share among similar accelerated systems sold, and that is a much larger market share than Supermicro has in the server space at large.
Supermicro has carved out similar niches in very dense storage servers, aimed at hyperconverged and object storage systems. The company has just introduced a 4U storage server that crams 90 3.5-inch drives into the unit, for a total of 720 TB of disk capacity; Supermicro is also first to market with NVM-Express ports on its servers. (NVM-Express substantially reduces the latency and boosts the bandwidth for SSDs attached to the servers.) Supermicro also pioneered the multi-node chassis design with its Twin family of machines many years ago, which are popular among cloud builders both public and private.
Supermicro is not exactly a tier one system supplier or an original design manufacturer – it is a bit of a hybrid, and that is what gives it another advantage.
“We are a superset of these, we are not just in between, we are in between and above,” jokes Liang. “With the contract manufacturers, their bill of materials is slightly cheaper and their manpower costs are slightly cheaper. But customers don’t just care about bill of materials costs – they care about total cost of ownership, and this takes into account performance, performance per watt, and density as well as – and especially – quality and service. The contact manufacturers who make for IBM/Lenovo or Hewlett-Packard or whoever, they are very good at manufacturing, but relatively weak when it comes to total solutions and services. IBM, for instance, has teams and can teach contract manufacturers who to design, optimize, and debug. But most large enterprises do not have this capability and if they had to develop this capability, it would be too expensive for them. But Supermicro is a very strong manufacturer and is also very strong in design and service. We understand how much quality means to corporate customers.”
The one thing that Supermicro is probably not going to do, by the way, is deliver machinery that is compliant with the Open Compute Project standards set up by Facebook and its corporate friends to try to foment a manufacturing base around open hardware designs.
“Lots of people have been asking me that question for the past four years. Our solutions outperform OCP designs, and that is why there is no reason to make another product line that is different from our own,” Liang says with a laugh. “It is easy to design something and to share with everyone. If it is better, then sure we would follow. But if it is not better than what we have, why waste the time and why waste the energy to give customers the second-best choice? For those customers who want the second best choice, they can call OCP.”
It doesn’t look like Supermicro is all that enthusiastic about the rise of ARM processors for servers or a possible resurgence of the Power architecture in the datacenter, either. But the company keeps a watchful eye for shifts, as it did with the Opterons back in the early 2000s. “We can design anything, but for now, our research focus is on Intel platforms because Intel has a broad product line. We have lots of engineering manpower, so we keep an open mind about any technology,” says Liang.