Cheap, mass-produced semiconductors have transformed our world, with smartphones, laptops, sensors and tablets, all connected to big cloud computing systems.
Now the business power of that transformation, in particular the size and might of the biggest cloud systems, is in turn affecting the powerhouses of the information age.
Intel, the world’s largest maker of chips, used to sell standard forms of its so-called x-86 line of processors by the millions, both for small personal computers and cloud systems with hundreds of thousands of networked computer servers.
Next year, however, custom designs will be on half of the chips it sells to public clouds – an increasingly important part of the 18 million chips Intel’s data center business sells annually.
“We have never said no to a custom solution” in chip designs for cloud, said Diane Bryant, the head of Intel’s data center business. “We get orders from the tens of thousands to the hundreds of thousands” of special chips.
Public clouds are computer systems that anyone can access, and use the computing power as some kind of service. The biggest public clouds, like Amazon Web Services, Google Compute or Microsoft Azure, frequently sell computing power or data storage itself. Other big players, like Facebook or China’s Baidu, are doing several things for large populations of consumers.
Other public clouds, including outfits like Twitter and eBay, are considered second tier by Intel, but still reach hundreds of millions of people. That is why they want specialized chips. Companies like A.W.S. “are running a million servers, so floor space, power, cooling, people — you want to optimize everything,” Ms. Bryant said. “The name of the game is customization.”
Indeed. Amazon has gone so far as to change around the 50 megawatt power substations that power its bunkers of computers. Deep in Google, former company executives say, the company has changed the way Internet packets of data work, to optimize how quickly they flow through Google.
Facebook, which has by far the world’s largest public repository of pictures, might want a certain number of chips that can render images well, or rapidly process the steps for image recognition. That could affect the number of processing cores it wants on a chip.
“As the dependency on technology increases, you have more workloads going across greater amounts of infrastructure,” Ms. Bryant said. “The name of the game is customization.”
Then there is the infrastructure itself. When eBay installed a hyper-efficient cooling system, it asked Intel for chips that had a greater thermal tolerance (they can be worked harder that way.)
The demand may be clear, but the economics of this business seem odd: Intel has long profited by turning out tens of millions of the same thing. Why go over to custom work?
The reason is, again, computers. Intel’s chip fabrication plants are now so automated that leaving out an unwanted core, or changing other properties, is a matter of a few new commands to the machine. Customers are willing to pay a little more for the special run of chips, or pay an engineering fee for the special service.
Since 2012 Intel has had an internal program called “Just Say Yes,” dedicated to looking for targeted workloads requiring custom chips.
The difference appears to be showing up in Intel’s results. While most of Intel’s chips still go into PCs, about one-quarter of Intel’s revenue, and a much bigger share of its profits, come from semiconductors for data centers. In the first nine months of 2014, the average selling price of PC chips fell 4 percent. But the average price on data center chips was up 10 percent, compared with the same period in 2013.
There might be a longer-term worry in this new world of individual chip orders so big that they are almost commodities: Could the growing size and comparative advantage of the big public clouds mean that we’re headed for consolidation?
If that is so, Intel could then find itself relying on a very few customers, each of which has lots of pricing power over Intel.
So far that seems unlikely. According to Ms. Bryant, in 2009 Intel counted seven tier-one players: Microsoft, Google, Facebook, Amazon, Tencent, Alibaba, and Baidu. There were at the time some 20 tier-two players.
Today there are still the big seven. Among tier two, there are 200 public clouds. In other words, more companies are figuring out new kinds of services they can sell that justify the expenses of running a big cloud.
People also used to say that PCs would eventually have demand problems. They did, starting around 2011, but the pundits had been predicting the drop for decades. We may be in a similar position with public clouds, and closer to the early days than the finish.