Patsnap invested ‘7 digits’ to develop an LLM. Will it pay off?

Build or buy?

That’s the age-old dilemma that many companies face when it comes to implementing new technologies.

With the advent of generative AI and the large language models (LLMs) that serve as the foundation of this technology, companies will have to confront this question again. Their choices could have significant implications on the future of their businesses.

Even as its closest competitor turned to partnerships, Patsnap, the Singapore-based unicorn whose platform lets researchers look up and analyze trends in intellectual property (IP), chose the first option.

Image credit: Timmy Loen

Despite the costs involved, Patsnap decided to build a proprietary LLM, spending “millions” of dollars to put together a team of over 50 engineers to train its model based on its own data.

Earlier this year, the firm launched an AI assistant designed to speed up the IP and R&D workflows of customers by allowing them to search through its database more quickly. Originally named CoPilot, the tool has since been renamed to Hiro.

Hiro draws from the LLM developed by Patsnap, which has been trained on patent records, academic papers, and other proprietary data.

The company’s numbers for 2023 are not available. But according to financial data from Alternatives.pe, Patsnap logged S$106 million (US$79 million) in revenue in 2022. It also recorded a loss before tax of US$30 million during the same period, marking a 12% improvement from the previous year.

Yet in spite of its losses, the company is not charging customers more to use Hiro. This calls into question how sustainable this initiative is, especially given the costs involved.

Need or want?

Some experts have noted that companies “don’t need” their own LLM if their goal is “to create new products or cut costs by automating processes.”

However, Patsnap co-founder Guan Dian tells Tech in Asia that the company has good reason to have its own LLM, which is built on top of Meta’s open-source LLaMA.

See also: How PatSnap grew to become a billion-dollar company

Explaining why Patsnap opted to build its own LLM, Guan says that while the big players like OpenAI’s ChatGPT and Anthropic’s Claude are “amazing” at answering questions in a generic context, both come with downsides.

Not a new product

ROI hard to measure?