If you track the investments in the generative AI field by VCs like Andreessen Horowitz, Sequoia Capital, or Y Combinator, you will notice that most of the exits are because of them getting acquired by Meta, Google, or other big tech firms. Case in point: Startups such as MosaicML and Neeva.AI, which got acquired within two years of their formation. And yet, it becomes difficult for investors to assess which one would actually give them the exit they need.

Generative AI is moving at such a breakneck pace with new models, fresh approaches, and different startups that it becomes difficult to ascertain if one is offering anything different from the thousand others in the field. When it comes to investing in generative AI startups, the most important aspect for an investor is identifying the startups’ unique approach and how exactly they can offer the ROI.

“We have to ensure that there is differentiation and novelty in whatever the startup is generating and the enterprise application they are building has some moat and cannot be replicated overnight by any other 500 startups in the field,” said Som Pal Choudhury from Bharat Innovation Fund (BIF), in an exclusive interview with AIM.

Choudhury said that BIF has been focused on investing in deep tech startups since the beginning, and generative AI is no different from investments in AI, blockchain, or 5G networks. The firm hasn’t invested in any generative AI startup at the moment as Choudhury explained that it is very essential to deeply assess the moat for these startups and whether it is worth investing in them since the big tech is already leading the way.

According to him, there are broadly five type of generative AI startups:

1.  Core Elements and LLMs: Companies developing core elements and large language models (LLMs).
2. Enterprise Application Focus: Startups creating specific enterprise applications using generative AI technologies.
3. Open Source Modifications: Companies using open-source models and modifying them with their proprietary data for enterprise applications. 
4. Orchestrating Multiple Models: Startups combining various generative AI models to create more specific enterprise applications.
5. Basic Wrappers: Companies that provide basic wrappers around existing models, without significant differentiation.

Eximius Ventures, a firm that has invested in two startups in the field, is employing a similar strategy. Pearl Agarwal, the founder of the firm told AIM that there are two types of startups in the field — horizontal platforms that are focusing on core innovations, such as OpenAI, and vertical platforms that are utilising technologies from these companies to create applications in diverse use cases. “In both these cases, the exit strategy is similar to other tech investments and is mostly about getting acquired by a tech-giant or going into IPO,” Agarwal said. 

There are very few startups, not just in India, but globally, that fall under the horizontal platform category that are driving infrastructural innovation in the field. It is tough for smaller VCs to invest in these companies as they require a huge capital and it is also difficult to assess if these would even work in the long run. “We are essentially peeling the layers of these generative AI startups to understand if a startup is actually doing something innovative or is just building a wrapper around ChatGPT or other similar models and just trying to sell it as a generative AI startup. Overnight we have seen all chatbot companies doing the same. We don’t want to invest in any of those companies,” emphasised Choudhury.

Unlike big VCs, BIF is still figuring out if any startup is worth investing in since they are just building wrappers around the already existing technology. On the other hand, Eximius Ventures decided to focus on the vertical platform companies that are leveraging these platforms for specific use cases instead of building their own.

India has never been ahead in the AI infrastructure race anyway and the same should be expected from startups coming out of the country right now in the generative AI field. Choudhury said it is hard to assess if a startup would generate revenue or not, no matter which of the five layers it is working in. “We’ll see a lot more applications-based and enterprise-focused applications come up, and that’s where I think Indian startups will contribute the most,” said Choudhury. 

The dichotomy

Open source has made it easier for new startups to form, but difficult for investors to assess if they are worth investing. By the time a product is built, newer and potentially superior models may already have flooded the market. Venkat Vallabhaneni, co-founder and managing partner at Inflexor Ventures told AIM that this dynamic environment actually serves as an essential litmus test for startup founders. “We value and seek out founders and companies that demonstrate a deep understanding of this highly evolutionary field. These are entrepreneurs who factor the rapid pace of development into their product design and strategies,” he said.

Similar to Choudhury, Vallabhaneni also said that some successful startups decouple their products from the foundational models, allowing their offerings to evolve and adapt in tandem with the market. But even then it becomes very hard to find the distinguishing factors of these startups.

Additionally, a part of an investment firm’s examination includes a deep dive into the technical aspects of the startup. “We need to understand the types of model they are using and assess whether they have a solid grasp of these models and their implications. While it may be challenging to build intellectual property (IP) on the model itself, startups can create their unique IP based on the implementation of the model and the specific use cases they are addressing,” said Vallabhaneni.

While IP may not always be the paramount concern, possessing enforceable IP is a definite advantage. It could prove to be an important asset that enhances the attractiveness of the company in potential acquisition scenarios. One of the ways that the investors believe that startups could stand out in the generative AI field is through their intellectual property data. Even if the startups are leveraging open source and wrapping their technology around other technologies, leveraging proprietary data and then solving for use cases would be essential to stand apart. 

“Our due diligence process for generative AI startups mirrors that of other deep tech startups, with an added focus on regulatory aspects”, said Vallabhaneni. It’s crucial that these startups are cognizant of the shifting regulatory landscape for generative AI and are flexible enough to adapt to any new developments.

Amid investors like Sequoia Capital, Andreessen Horowitz, and Y Combinator, who are investing in big tech-backed companies such as Inflection or Anthropic, other VCs find it easier to invest in more application-driven generative AI startups that are building out of open source and hope to get acquired by a big tech soon. Looking at the trend, it appears that it might take 2-3 years for that to happen.