This article first appeared on GigaOM Pro, our premium research subscription service.
The ongoing imperative to lower power consumption in the data center drove another key acquisition last week. AMD, a company that has fought a decades-long losing war with Intel over building the best CPU, acquired low-power server maker SeaMicro in a move that says a lot about how low power is the new black and how the relationships and market dynamics between chip makers and server OEMs is transforming.
While the focus on low-power servers often centers on low-power chips, that wasn’t what AMD liked about SeaMicro. AMD wanted everything that connects to the CPU, and SeaMicro’s technology centers on the fabric that links together hundreds of low-power CPU cores, simultaneously reducing the number of components on the board while it turns off those cores not in use to reduce power consumption. SeaMicro has spent $50 million developing the technology.
It’s the technology that matters, not the OEM
Some analysts have written that the acquisition puts Intel at a disadvantage. For sure, it’s not an ideal scenario for Intel, but the biggest losers in the ecosystem by far are the OEMs — the server makers themselves. AMD CEO Rory Read was adamant on the call last week with analysts, saying, “We will not compete with our customers,” meaning AMD won’t use SeaMicro servers to go up against the likes of HP and IBM. In reality, however, it’s not that simple.
What AMD gets from buying SeaMicro is the technology to build a better server. AMD can license that technology to OEMs, which will eventually need to please their customers. And so the SeaMicro deal is one more step toward transferring the value, which an end customer like a Google or a Facebook needs, to a company like AMD and away from the OEMs.
The OEMs know this. When I spoke to SeaMicro CEO Andrew Feldman on Friday he said SeaMicro was pursued by buyers at both the chip level and the system level. But margins remain tight in the commodity server business, leading some companies like Oracle to abandon it altogether. Expensive acquisitions like the $334 million AMD paid for SeaMicro remain out of reach for OEMs, creating a vicious cycle where valuable technology slips away, further lessening the value OEMs have in the value chain. This has led to reports that Google and Facebook are going direct to Taiwanese design manufacturers to buy some of their servers.
Very sophisticated customers
What’s driving the rapidity of all of this change is the aforementioned operational costs associated with power consumption, which have sent companies like Facebook to places like Prineville, Ore., with its cheap power and mild climate (lower cooling costs). The other key driver is that as we look out five to ten years, AMD’s real customers aren’t only HP and IBM; they are the Googles, Facebooks and likely webscale companies with large cloud computing infrastructures we can’t yet imagine.
Speaking about the pace of innovation and disruption in the server market, Feldman noted, “Usually early adopters are a tiny portion of the market. What’s so interesting about the cloud, about the Googles, Facebooks and mega data centers is that they’re early adopters and a major portion of the market.”
He added, “If you can’t demonstrate value [webscale cloud providers will] cut you out. That creates challenges for every one of us. Everybody’s looking around and saying we have extraordinarily sophisticated customers who have no tolerance for the absence of differentiation.”
Demonstrating value to those customers has opened up the possibility of AMD’s doing innovative things, like even licensing an ARM core to build a non-x86 chip. Lisa Su, AMD’s SVP and general manager of Global Business, has said, “we’re not going to be religious on architecture,” and SeaMicro time and again has emphasized its fabric chip can work with any CPU architecture out there. Even more interesting is that as chipmakers often talk in terms of product road maps that are 2–4 years out, SeaMicro can turn around a new server design based on a new chip in 8–12 weeks. That’s a culture of rapid development and innovation, one I hope doesn’t get stifled at AMD.
The one saving grace about having massive buyers like Google, which has around a million servers, is that there is a lot of money to drive innovation. Google doesn’t just ask the question “Can we build a lower-power server?” It buys and tests servers that do just that. That mindset bodes well for AMD and its drive to acquire the building blocks that will allow it to quickly offer the chips and fabric that drive down power consumption for its customers — however you want to define “customer.”