- Arista CEO Jayshree Ullal says the network can pay for itself by spotting idle AI accelerators and putting them back to work
- Open Ethernet can match proprietary AI interconnects like Nvidia’s InfiniBand and NVLink without vendor lock-in, she said
- Ullal expects AI to shift from centralized training to distributed inference, pushing Arista to build networks for “centers of data”
Under Jayshree Ullal's leadership, Arista Networks has become one of Silicon Valley's defining technology success stories. Since joining the company in 2008, the former Cisco executive has helped transform a promising networking startup into a $200 billion-plus megacorp, recently admitted to the S&P 100 and now forecasting approximately $13.3 billion in annual revenue.
Ullal still likes to describe Arista as a startup. But in an exclusive FNTV interview, she outlined an ambition considerably more interesting than yet another set of impressive financial results.
The network, she argues, could literally pay for itself
Not simply by moving more data or connecting more GPUs, but by making the enormously expensive machinery inside AI factories more productive.
Consider AI accelerators costing millions of dollars. According to Ullal, if utilization falls from 90% to 30%, the financial consequences can be enormous. The investment is still there, but much of its computing capacity is sitting idle, twiddling its teraflops.
Idle XPUs are the devil's workshop (for a company budget)
What if the network could recognize that waste and do something about it automatically?
"The network literally pays for itself if the management plane, the observability plane, can say, that's spinning. Let me spin up a new agent and connect it to that XPU without being asked," Ullal told me.
In a hyperscale AI factory containing billions of dollars' worth of accelerators, even modest improvements in utilization could have substantial economic consequences.
It's a significant departure from the industry's traditional habit of measuring networking value in bits per second, port density and the price of a switch. Or its newer habit of throwing token math at every business plan.
From silent hero to economic engine
Ullal describes the network as AI infrastructure's "silent hero," although she acknowledges it can become a "silent zero" when things go wrong.
AI factories depend on keeping enormous numbers of interconnected processors busy. Delays, congestion or failures can leave expensive hardware waiting for work and create a nightmare for the CFO.
Arista combines its Extensible Operating System (EOS) with network-wide telemetry, observability and intelligent management.
Rather than requiring engineers to hunt through log files, Arista's software already supports automated fault detection, diagnosis and corrective action. Ullal says its architecture can identify idle accelerators and initiate actions to restore productive workloads without waiting for an engineer.
"The old world of syslogs couldn't do that," she said.
Arista also supports automatic fault containment, agent recovery and software upgrades without scheduled outages, although some customers still prefer maintenance windows.
The economic proposition is compelling. If networking software can recover wasted computing capacity, its contribution extends far beyond connectivity.
Ullal was describing technological capabilities, rather than presenting independently verified customer savings. Nevertheless, the argument suggests a fundamental change in how networking companies demonstrate their value.
Ethernet versus the proprietary AI stack
When ChatGPT exploded onto the scene in November 2022, Ullal admits she felt Arista was "outside looking in."
AI infrastructure appeared to be heading toward proprietary interconnects, particularly Nvidia's InfiniBand and NVLink technologies, with Ethernet relegated to the sidelines.
Arista responded by accelerating open networking standards, becoming a founding member of the Ultra Ethernet Consortium and participating in efforts to extend Ethernet into scale-up AI networking.
The company now addresses three requirements: scale-up connections between accelerators, scale-out connections between racks and clusters, and scale-across networks connecting geographically separated AI infrastructure.
The significance goes beyond switching hardware to the nature of networking in the 21st century.
As I suggested during our conversation, the emerging AI economy risks recreating the lock-in economics of the mainframe era, with a handful of enormous companies controlling everything from processors to applications.
"It's our founding principle to be open and standards based and avoid vendor lock-in," Ullal said.
She argues that interoperability need not come at the expense of performance or reliability. Arista's experience building networks for hyperscalers has demonstrated that these objectives can coexist.
Beyond the AI factory
The next opportunity may be larger still.
I put it to Ullal that America's AI investment remains heavily concentrated in centralized data centers, while other parts of the world are increasingly focused on industrial machine learning and AI operating in the physical economy.
She agreed (always nice).
Ullal expects today's emphasis on centralized AI training to evolve toward distributed inference, smaller models and specialized processors operating closer to users and machines.
Computing previously evolved from mainframes to client-server architectures and mobile devices. AI, she believes, could follow a comparable trajectory.
"We're only just in the first phase of training, and there's much more to come," she said.
This explains Arista's transition from building networks for data centers to building networks for centers of data.
Computing and information can reside in AI clusters, enterprise campuses, branch locations and edge infrastructure, all requiring consistent networking, observability and control.
It also raises questions about operational sovereignty: who controls this increasingly distributed infrastructure, and who makes the decisions governing its operation?
Ullal acknowledges the growing responsibility this places on networking suppliers, particularly around security, isolation and reliability.
Meanwhile, she remains determined that Arista should retain its startup culture, despite its extraordinary growth.
"Don't change what's working well," she said, describing perhaps her greatest remaining ambition.
But the most consequential argument in our conversation concerned the economics of AI itself.
The industry has spent years obsessing over how many GPUs it can manufacture, acquire and install. Far less attention has been paid to how effectively those processors are used once installed.
Ullal is suggesting that the network can become an active participant in solving that problem.
For an industry investing hundreds of billions in AI infrastructure, the next competitive advantage may not be owning more machines. It may be making the ones you already own do more work.
And if the network makes that possible, the humble Ethernet switch has come a very long way indeed.
Learn more about AI networking
The AI factory is leaving the data center — Stephen M. Saunders MBE. Why the bigger economic prize may lie in industrial machinery, distributed intelligence and operational sovereignty.
'AI tailwind' keeps Arista at No. 1 spot in data center switching — now Nvidia looms — Mitch Wagner. Arista leveraged the AI buildout to pass Cisco in data center switching. Now it's facing off against the AI colossus.
Arista's Ullal: Ethernet is the 'eventual winner' for AI networking — Mitch Wagner. Ullal's case for open Ethernet challenging InfiniBand in AI infrastructure.
Nvidia's SVP of networking explains the 5 types of networks needed for AI factories — Linda Hardesty. How Nvidia's networking architecture addresses the same AI factory challenges from a different direction.
Stephen M. Saunders MBE is a communications analyst and USPTO-registered inventor examining how digital infrastructure — 5G, cloud and AI — is reshaping industry, power and society, as well as underpinning the emerging, ubiquitous global digital economy. As anchor of FNTV and a longtime industry insider, he focuses less on growth narratives and more on execution, risk and how hyperscale technology is distorting markets, governance and society at scale.