Companies that make money by providing physical infrastructure services for other companies to use — think Microsoft, with more than a million servers — have a big incentive to buy gear that’s as efficient as possible. And more of those companies are showing interest in leaving behind longstanding OEMs and looking to lesser known hardware builders that can be more enthusiastic about letting their customers buy a tailored product.
Previously, content-delivery network and security company CloudFlare had gone to Dell, Hewlett-Packard and ZT Systems for servers. But for its latest batch, CloudFlare went with Quanta. As a result, CloudFlare has been able to incorporate into the boxes (pictured) equipment that was just right for their workloads, including network-interface cards from Solarflare, Intel solid-state drives, and energy-sipping Intel 2620 Xeon E5 processors, said Matthew Prince, the company’s CEO and a co-founder (and one of our 2013 Cloud Trailblazers).
There could be some tradeoffs in going with these vendors: companies such as Quanta might not have as many service people available for support in every country, Prince said. Beyond that, there could be a layer of complexity around the distribution of the servers to all sites running a customer’s infrastructure. But at the CloudFlare office in San Francisco last week, it didn’t seem like much of an issue. Cardboard boxes containing servers were piled up along the walls but didn’t sop up all the floor space at company headquarters as they waited to be shipped to sites around the world.
ProfitBricks, an Infrastructure-as-a-Service (IaaS) provider, went to Supermicro for servers packed with AMD chips, said Pete Johnson, the company’s senior director of platform evangelism. But he thinks the company’s distinguishing factor is its use of InfiniBand technology — not only for faster networking, but also to allow customer to have virtual machines of whatever size they want.
That technology could help provide better service than the one-size-fits-all approach in play at Amazon Web Services, or then again IaaS providers could do so by buying more customized server gear through companies historically known as original-design manufacturers (ODMs), Johnson said.
Like hiring a contractor
Meanwhile, at least one other IaaS provider is rolling its own servers with the help of ODMs. Rackspace works with Quanta to build its own machines. NTT and Orange have joined the Open Compute Project, suggesting interest in streamlined server design. And while it’s not an IaaS provider, Riot Games worked with Hyve Solutions to build servers inspired by Open Compute Project designs.
Fast-growing IaaS provider DigitalOcean has worked with Dell and Supermicro to get servers optimized for its needs, but co-founder and CEO Ben Uretsky can see why Quanta and its ilk could be a compelling alternative in the future. “We started off with retail servers with the traditional guys, and we’re constantly looking for something else, something denser, something better performing, who can really listen to the requirements we have for specific components,” Uretsky said.
And CloudSigma, another IaaS vendor, uses servers with custom-tailored accoutrements from Supermicro and Dell’s Data Center Solutions group, although CEO Robert Jenkins said Quanta looks worthy of consideration. Jenkins sees Quanta as well versed in working on requests for specialized gear, and he is curious about whether the company would be able to turn around products in response to orders faster than other vendors. “If you can cut the lead time from six to eight weeks to two to three weeks, that would be a different ball game,” Jenkins said.
At the same time, Jenkins said, companies that directly purchase from server makers and then lease them out to companies that use the gear, such as CloudSigma, might be more reluctant to provide support for Quanta than brand names such as HP or IBM. That’s possible because after a IaaS provider uses a server for a few years and wants to get new server to put in its place, the value of the original white-box server is lower on the market than, say, a similarly used IBM server.
The thing is, as computing resources become more centralized in clouds, Jenkins expects more companies wielding large-scale purchasing power to be able to customize machines, which translates to less business for cookie-cutter servers and more for Quanta-style specializing.
We’ve long known that Google has made its own switches and servers, although the company declined to disclose information on its suppliers for this article.
And Amazon surely has the scale at which getting highly customized gear would make sense, too. Amazon is secretive about what it has actually done. But an AWS spokeswoman did write in an email that “we do have specialized server design teams, and we work with a variety of vendors and sourcing models to refine our server technology and ensure the best possible cost, performance and reliability.” The statement says nothing about which types of companies go to work when AWS buys more servers, but it does suggest that the cloud provider likes getting exactly what it wants.
What do the traditional OEM vendors think about all this? An executive at an IaaS company that gets customized gear told me that traditional server sellers have been willing to go toe to toe on price, but that practice might not be sustainable, particularly when fat margins are desirable. Nevertheless, Dell is apparently planning price cuts and other changes. HP is banking on its Moonshot server line to stay cool. IBM reportedly has looked at selling its server division. And as for Oracle, well, at least one analyst is still holding out for a turnaround.
This story was updated at 7:31 p.m. PT to correct information on ProfitBricks’ servers.