Zoho’s Sridhar Vembu flags AI spending concerns as LLM industry reassesses pace

As large language model (LLM) companies reassess the pace and cost of AI development, Zoho Corp chief scientist and co-founder Sridhar Vembu said the industry is beginning to confront the economics of escalating token usage, data-centre investments and AI infrastructure.  

“See, part of it is enterprises, for very good reasons, are reluctant to expose their information to this. That's where ZiaChat comes in. You could expose it to ZiaChat, we take care of that, what we have. So that's where a lot of R&D is going in, so that safety here,” Vembu said. “They (LLM companies) themselves don't have any particular expertise in it. They are also scrambling to figure out all that. The second part, the token maxxing, that extreme expenditure on tokens, companies have realised it. So there is some, you know, sort of, we cannot afford to spend this much at the pace at which we are spending across the industry.”  

Vembu said the scale of investment in data centres also raises questions about whether customers will be willing to pay enough to generate adequate returns. “And all of this is, and also in the data centre investments, the trillions of dollars when they need return, you look at the revenue implication of that, how much revenue they have to make, that revenue has to come from customers. Are the customers willing to pay those prices to justify those data centre investments?”  

“I think all of this and that, some calculations are now going on, but as usual, there is a, you know, initial—you just drive as fast as you can, then you start to pace and realise, and that's the stage we are in,” he added.  

Zoho has been expanding its AI capabilities, from launching its proprietary large language model, Zia LLM, in July last year to subsequently introducing its Agents Suite and Agent Studio. The company has now launched Zia Chat, an agent embedded in Zoho’s business applications that can also integrate with other agentic ecosystems using the contextual data of businesses.  

Vembu described Zia Chat as Zoho’s answer to the growing AI needs of enterprises, with a focus on security and controlled access to business data.  

“We have those 50-60 apps, the broadest suite in the industry, we are bringing it all, exposing it all in a very safe way with a guardrail-protected way to this intelligence, and this (ZiaChat) intermediary sitting there. So, when you chat with it, it's getting the thing from LLM on what to do. It's checking it, cross-checking it, and then doing it, querying and doing the operation, all of that,” Vembu told Fortune India at the company’s corporate hub in Chennai.  

The company has also launched plug-and-play, templated yet customised models of vertical-specific SaaS offerings across seven sectors, including retail and BFSI.  

“We are not charging for the customisation itself. We only charge for the finished product and we want to accelerate this faster. You will see over the next year or two, we will take the whole horizontal suite, have hyper-verticalised for each industry segment,” Vembu said.  

He said the approach could accelerate project implementation while allowing Zoho to pass productivity gains on to customers.  

Vembu stepped down as CEO in January 2025 after leading the company since 2000 and now focuses on research and development as chief scientist. He said Zoho has filed several patents, with some of its technologies nearing commercialisation. In AI, the company is focusing on how AI can be more deterministic in nature. “In the LLM case, you cannot meet the specification of accuracy because fundamentally, the technology is not capable of that. So you ought to build this layer to guarantee that accuracy—this system is probabilistic, and this system is deterministic. These two oughts to work together. This is where we have made breakthroughs,” he said.  

Zoho is also working on ways to reduce memory consumption in AI systems, Vembu said.  

India is Zoho’s second-largest revenue market after the US and one of its fastest-growing, with revenue growing 47% year-on-year in FY26. However, as the software industry adjusts to the AI era, Zoho’s consolidated net profit margin declined from 34.8% in FY23 to 33.5% in FY24 and 27.8% in FY25. Its expenses rose from ₹4,899 crore in FY24 to ₹6,234 crore in FY25.  

Vembu said the rising cost of computing infrastructure is also changing the economics of acquiring new customers. Serving a new customer now costs Zoho more than serving an existing customer, he said, because of the increase in computing hardware costs.  

“This is impacting us because we are paying for those servers now ourselves, and we are not charging extra from the customer. We have not raised prices. We have actually thought about it, and we may have to, at some point, go to the customer,” Vembu said.  

With spending on memory and computing infrastructure rising, AI is adding to Zoho’s costs even as the company seeks to pass productivity gains to customers.  

“We have to economize on other dimensions. That is why we have to reduce hiring, not because we don't have demand for software engineers, but because we have to pay for all of this. We have to pay for AI, and we have to pay for memory,” he added.  

In its 30th year, Zoho—originally AdventNet, which acquired the Zoho.com domain in early 2000 and rebranded itself as Zoho Corp in 2009—is also evaluating a potential entry into sovereign data centres if hardware and chip costs decline.  

“We are not going to do it as long as it requires extraordinary expenditure of borrowed capital. I'll put it that way,” Vembu said.