AI agents don't take coffee breaks, so networks must scale up: Ciena's Gautam Billa

With AI agents in the mix now, Billa said the network connecting the computing resources is likely to become a bottleneck in India's AI push, as their wider adoption in businesses drives up the data usage.

Telecom networks have usually relied on an assumption that people will not speak, text or use data at the same time, but AI agents could force operators to rethink network capacity, according to Gautam Billa, chief technology officer (CTO) and vice president of Solutions Engineering for Asia Pacific, Japan, and India (APJI) at Ciena

With AI agents in the mix now, Billa said the network connecting the computing resources is likely to become a bottleneck in India's AI push, as their wider adoption in businesses drives up the data usage.

Ciena is a US-based networking company that makes optical and packet networking systems, along with automation software for telecom operators, cloud providers and enterprises to build high-capacity networks.

"There's a lot of work that is already happening, starting from basic infrastructure to power, to putting fiber in the ground. I believe the industry will be ready when this wave actually hits us," Billa told ET AI at the India Mobile Congress (IMC) 2026.

Agents change the traffic equation

The CTO said operators traditionally used statistical multiplexing, a way of sharing capacity based on the expectation that users will not all need it simultaneously. Operators can provision less onward bandwidth for a base station serving 1,000 subscribers because they do not all talk or text together, he explained.

In network and telecommunications, a stat mux (statistical multiplexer or statistical multiplexing) is a technique that dynamically shares a communication channel by allocating bandwidth based on real-time traffic demand rather than fixed time slots; onward bandwidth refers to the data capacity available for the next stage, hop, or forward direction of a network path.

Replacing those subscribers with 1,000 agents changes the calculation, according to Billa, because the machines can keep generating traffic without any breaks.

"Now, what happens if you replace these 1,000... let's say human beings with 1,000 AI agents? Those AI agents are working non-stop, they don't take coffee breaks," he said. "They don't talk to people, they just perform the work assigned to them. So that stat-muxing factor goes away. Now, I have to build networks that are more scalable."

Agents also need to exchange the background information or context people already carry into a task, Billa said, estimating that an agent uses four to five times the data a human would use.

"From a network characteristic perspective, it's a lot more continuous, it's not bursty traffic, which means the scale changes," he said.

More data, greater demands

A similar shift extends to enterprises, where AI can work with emails, PDFs, presentations, videos and photographs, instead of the structured information stored in databases, Billa said. Such multimodal inputs, which combine different forms of information, expand the pool of data businesses can turn into intelligence, increasing the traffic their networks carry, he added.

He pointed out that the consequences of inadequate connectivity become particularly clear when companies train AI models across distributed clusters of graphics processing units (GPUs). Those clusters must exchange data after each iteration, so a slow connection delays training and leaves expensive computing equipment idle.

"The maximum investment you made is in GPUs and you want them to work as optimally and efficiently as possible. That's the reason the network speed is extremely important," Billa said.

A Ciena survey from August 2026 highlighted the need to improve network performance, with 88% of Indian respondents saying optical networks need urgent upgrades to support enterprise AI services, including 38% who said major upgrades were needed immediately.

Most of the demand in India comes from about 10 major cities, including Mumbai, Delhi, Hyderabad and Bengaluru, Billa said. Customers need networks connecting these cities, as well as the data centres within each city; these networks also need to connect to undersea cable landing stations to link India to the rest of the world.

Large cloud providers initially leased capacity from service providers before moving to managed optical fibre networks (MOFN) and now some are starting to build their own networks, Billa said.

"Now they are starting to actually build their own networks. They are more well-versed with the environment in India and how things work and how networks can be run in terms of reliability and stability," he said.

Under managed optical fiber network (MOFN), a provider supplies and manages a dedicated fibre network whose full capacity the customer uses, he explained.

Service providers also see a business opportunity in these dedicated networks, with 98% of Indian respondents in the survey expecting MOFN services to generate revenue over the next three years by connecting AI computing clusters.

Fibre, power come first

Meeting the growing demand also requires more reliable physical infrastructure, Billa said, pointing to fibre cuts during road construction and the subsequent need for backup routes. Providers are considering deeper burial and ducts containing multiple fibre pairs so they can carry more traffic, he said.

Moving a petabit of capacity between Mumbai and Chennai could require about 20 to 35 fibre pairs, according to Billa, who believes scale increasingly shapes infrastructure planning.

Billa said the plans for at least three or four additional subsea landing stations would bring more capacity into India, requiring terrestrial networks to connect the cables to data centres.

Asked what India should prepare for over the next five years, Billa said technology was changing in five weeks and urged players to focus on infrastructure projects such as laying fibre, securing power and preparing data centre facilities.

"Those are the ones that take time. Once these things are up and running, we can roll out networks pretty quickly...those are the things that we need to focus on in the long term," he said.