Zoho's $20-mn AI bet: why the SaaS giant built the homegrown LLM from scratch

“Zia LLM is built to serve enterprise needs from the ground up. Tailored for businesses, not consumers,” says Mani Vembu, Chief Executive Officer, Zoho Corp., affirming that their newly launched AI stack is fully homegrown, trained without using customer data, and deeply integrated into Zoho’s ecosystem.

With this statement, Zoho, the Chennai-headquartered technology company, made its boldest move yet in artificial intelligence unveiling its proprietary large language model (LLM), Zia LLM, built completely in-house using 256 GPUs by NVIDIA. The company announced this and a host of other AI tools, including speech recognition models in Hindi and English, a no-code agent builder (Zia Agent Studio), and a Model Context Protocol (MCP) server, at its annual user conference Zoholics India in Bengaluru.

At the heart of this launch lies a staggering $20 million investment over the past year to develop foundational AI capabilities and build Zoho’s own enterprise-grade backbone. “We didn’t just fine-tune an open model. This is our LLM, built from scratch, deeply tied to our product use cases,” said Ramprakash Ramamoorthy, Director of AI Research at Zoho.

The LLM comprises three custom models with 1.3B, 2.6B and 7B parameters, each optimised for enterprise applications like structured data extraction, summarisation, retrieval-augmented generation (RAG), and even code generation. Zoho’s approach focuses on “right-sizing” the model depending on user context—a resource-efficient strategy that lets it balance compute load with performance.

No customer data, no compromises on privacy

In an era where most foundational models feed off vast amounts of user data, Zoho stands out. Both Vembu and Ramamoorthy confirmed that no customer data was used during training. The models rely instead on a combination of open-source and publicly available annotated datasets. This also aligns with Zoho’s long-held position on privacy: customer data stays within Zoho servers and isn’t routed to third-party AI cloud providers.

Zia LLM is only one piece of the puzzle. Zoho also introduced over 25 domain-specific AI agents tailored to functions like HR, marketing, sales, and finance—each of them built to enhance Zoho’s 55+ application suite. For developers and partners, Zoho’s Zia Agent Studio provides a no-code platform to build, deploy, and monetise agents, effectively creating an AI version of an app store.

The new MCP server opens access to a large library of actions from Zoho apps, allowing AI agents—internal or third-party—to execute workflows while respecting user permissions. Over 15 Zoho tools are already exposed through the MCP server, with more to follow, and integration is also available with non-Zoho platforms via Zoho Flow.

No monetisation yet

What makes the move even more striking is that Zoho has no plans to charge customers for AI services—yet.

“We are consciously okay if we don’t make money in the short term, or even absorb the cost,” says Ramamoorthy. “What matters right now is adoption. Once customers experience the value, pricing will follow and we believe they’ll be happy to pay for it.”

Vembu echoes the sentiment, describing the AI investments as long-term R&D that will increase customer retention and lifetime value across Zoho’s product suite.

India, a crucial growth market

India continues to be a key growth engine for Zoho. In 2024 alone, the company reported 32% growth in the Indian market, with over 1,50,000 customers. Globally, Zoho is closing in on 1 million paying users. The company’s 10-year CAGR of 51% reflects consistent, profitable growth fuelled increasingly by deeper platform adoption and AI-led automation.

Interestingly, while Zia LLM is still being tested internally, full-scale customer rollout is expected by late 2025. The company believes its measured approach—focused on governance, accuracy, and contextual adaptability—will pay off in long-term user trust and product stickiness.

What’s next?

Zoho is now working on scaling model sizes, with new parameter sets expected by the end of next year. In parallel, it is expanding speech capabilities to other Indian and European languages and planning to release a reasoning language model (RLM) soon.

As the global LLM race intensifies, Zoho’s $20 million investment signals a strategic bet that deep vertical integration, privacy-first design, and long-term platform play will beat short-term flash. With Zia LLM, it may have just redrawn the map for enterprise AI in India.