A Need for More Than Speed

     By the time D’Ambrosia was convening his 400G effort, Google went public with details of the next traffic jam inside its big data centers — complexity. Networking more than a million servers into one warehouse-sized computer required thousands of routers and switches and those systems demanded far too much time from human operators.

A group of academics and entrepreneurs gathered at Stanford to find a way forward. One of them, Amin Vahdat, a computer science professor Google hired to sort out its networking issues, became a spokesperson.

As the representative of one of the world’s largest buyers of networking gear, Vahdat had clout. He was also a good communicator, explaining complex and often controversial new ideas in a calm, clear way. That came in handy because the cloud giants were essentially upending the technology and business model of big switch and router makers like Cisco.

Vahdat is ultra smart, soft spoken and lucky. While still an undergrad at Berkeley he was among the grad students that did some of the grunt work for a landmark paper that described what would become the core technology for cloud computing -- massive clusters of networked PCs. He kept pursuing the field as a professor at Duke where he won tenure in record time, then took a job as chair of the computer science department at the University of California, San Diego, where he led research in 2009 showing a way forward for networking a giant data center. That got Google’s attention. 

By 2013 he gave one of his first keynotes on the new direction at an event ironically on the Cisco campus in San Jose. The talks became part of an annual drumbeat.

“The network is our fundamental barrier to delivering new features,” he said in a 2014 talk. “The future of cloud computing is about delivering new capabilities we can’t deliver now, not delivering old capabilities cheaper,” he added.

The problem was that networking companies had never focused on problems at the scale of a Google or Amazon. Each vendor developed its own proprietary switch and router software, accelerated by its own custom chips. The approach worked fine for connecting a few dozen, even a few hundred computers in the back offices of typical corporations, but at the scale of the new cloud data centers it was a nightmare.

The solution, Vahdat said, was to drive all the networking tasks to a new class of software running on x86 PCs. For this idea of software-defined networking (SDN) to work lots of new code needed to get written fast. Folks in places like Google were just starting that work.

SDN was the best way to prepare for the data tsunami that was already inundating the cloud giants, Vahdat said in a 2015 talk. Google saw traffic in its data centers rise 50-fold in the last six years, he said.

Within a year, the SDN concept started attracting carriers like AT&T and Comcast whose networks were getting flooded by data traffic from PCs and smartphones. Before long cloud giants and carriers formed industry groups to write open-source SDN code and host events to popularize it.

“I think we are at a point where things will start happening rapidly,” Guru Parulkar, an academic and entrepreneur who led the main lab developing the new code, said in a 2016 talk.

 “We have been working on SDN ideas for eight years. So, we are still talking about 10-15 years total for it to become mainstream,” said Parulkar who these days works on an AI-powered life coach he hopes can embody digital wisdom.

Before SDN was halfway done, cloud computing flattened the industry of computer and network systems makers who served it. Expensive, proprietary technologies were out. Dan Pitt, a former network systems exec who became a spokesperson for the movement, declared victory for the open-source code called OpenFlow.

“The future of the network is Ethernet, x86, and OpenFlow — nothing is controlled by a single party, it’s all community based,” he said at a 2014 gathering.

Looking back, “OpenFlow served as a proof point for SDN…it did not become a staple of SDN implementations,” Dan told me recently.

Instead, Microsoft led a project called SoNIC. With help from Broadcom, Dell, Nvidia’s Mellanox crew and others, it developed open-source software that became the foundation for SDN code many cloud computing giants use today.

And one lucky entrepreneur saw an opening. Andy Bechtolsheim – the socks-and-sandals wearing co-founder of Sun Microsystems who is said to have written the first check to fund Google – created a networking company in 2004 called Arista. Aiming for the clouds, it became a major supplier of network hardware and SDN software to Google and its peers. Arista is now clocking annual sales of $10+ billion. Its established rivals like Cisco quickly followed suit, often buying SDN software startups to attract sales.

And there was lots of money to be made. Analysts estimate Amazon alone -- the largest of the so-called hyperscalers -- now runs at least five million (some say 20 million) servers across 900 data centers in 50 countries.

What’s more, Amazon and its rivals are now in the midst of an unprecedented expansion, spending in 2026 alone an estimated half a trillion dollars to build a new generation of data centers designed to create more powerful AI algorithms than those that power today’s ChatGPT and Google Gemini services.

Coming soon: How the cloud rained on EE Times’ parade