Artificial intelligence (AI) startups may be the brightest stars in the venture capital universe these days, but the glow is beginning to blind some investors. Even as early-stage funds continue to pour record amounts into the sector, several venture capitalists (VCs) are quietly stepping back, wary of inflated valuations, weak monetisation plans, and untested business models.
“Many AI companies are putting forward high valuations, yet in several cases, their strategies for monetisation and long-term sustainability remain unclear,” Milan Sharma, founder and managing director of 35North Ventures, told FE. While he agreed that AI has the potential to redefine the future, he said survival will depend on fundamentals like strong technical talent, ethical governance, and a clear roadmap for scaling.
The caution stems from a growing sense of deja vu in global tech circles, with parallels being drawn to the dotcom bubble of the early 2000s. Andrej Karpathy, a founding member of OpenAI, recently noted in a podcast that artificial general intelligence (AGI) remains at least a decade away, despite rapid progress in large language models over the last three years. He warned that exaggerated claims about AI’s capabilities could end up damaging the credibility of the field.
The numbers, however, tell a different story of exuberance. According to CB Insights, AI startups cornered more than 50% of global venture capital investments in 2025.
In India, the enthusiasm is palpable too, with both new and established funds vying for a slice of the AI pie. But behind the scenes, investors admit that identifying the true winners is proving tricky.
“This is because we have only begun our AI journey. And this holds true for both startup founders and the VCs backing them,” Bhaskar Majumdar, managing partner of Unicorn India Ventures, said. According to data from Tracxn, AI startups in India have raised $474.57 million so far this year. Among the notable fundraises are those by UnifyApps, Composio, Spyne, Netrasemi and Beacon. The biggest deal came from Uniphore, which secured $260 million in Series F funding from a consortium that included Nvidia, AMD, Snowflake and Databricks.
Majumdar said his firm is being especially cautious about betting on front-end AI startups. “When it comes to defining winners in AI applications, it is very difficult to predict outcomes because the sector is evolving so fast. The real opportunity, we believe, lies in the infrastructure layer – the chip industry, AI data centres, and micro power generation, which will support this high power-intensive sector.” Unicorn India Ventures has invested in startups such as Netrasemi and Kluisz that play in this space.
For many investors, the challenge is not about spotting the next unicorn but about separating substance from hype. Somdutta Singh, founder and CEO of Assiduus Global, said the distinction between meaningful AI integration and superficial adoption is often blurred. “A company saying it uses AI can mean anything from basic automation to truly transformative use. The balance lies in applying AI for what it can genuinely do today, while staying aware of how much remains to be achieved,” she said.
VCs are increasingly digging deeper before writing cheques, asking whether AI is truly the core enabler of a startup’s value proposition or merely a marketing add-on. Singh said that authentic innovation always reveals itself in the quality of data, the robustness of models, and the depth of the problem being solved. “The best investments will come from recognising that difference early on and supporting founders who build with authenticity, not hype,” she added.
Abhishek Prasad, managing partner of Cornerstone Ventures, echoed this view, saying his firm prefers to back AI startups that are creating tangible impact rather than chasing headlines. “Our focus has always been on solutions and applications. Unless one has very deep pockets, investing in foundational models is risky. Even in the agentic layer, there’s a risk of commoditisation. The smarter play lies in application-led models that build domain-specific moats and create meaningful outcomes,” he said.