AI Is Now Way Better at Predicting Startup Success Than VCs
News Summary
Researchers from the University of Oxford and Vela Research have developed VCBench, an open benchmark to assess AI's capability in predicting startup success. Their study reveals that large language models (LLMs) such as GPT-4o and DeepSeek-V3 significantly outperform leading venture capital firms and even Y Combinator in identifying future successful startups. VCBench utilized a dataset of 9,000 anonymized founder profiles paired with early-stage company data, with approximately 810 labeled as "successful" (defined by an exit or IPO). The models demonstrated superior precision compared to human benchmarks; for instance, DeepSeek-V3 achieved over six times the precision of the market index, and GPT-4o led with the highest F0.5 score. This research suggests that LLMs could become indispensable tools for early-stage investing, enabling the discovery of promising founders earlier and potentially fostering a more meritocratic environment in startup investments. VCBench has been released publicly, inviting further community validation of AI's potential in this domain.
Background
The venture capital (VC) industry has historically relied on experienced investors' networks, intuition, and pattern matching to identify promising startups. This traditional approach often carries inherent biases and favors "warm introductions," creating high barriers to entry for founders outside established circles. However, with the rapid advancement of artificial intelligence, particularly large language models (LLMs), their capabilities in data analysis and prediction have grown significantly. In 2025, AI technologies are already widely applied across various sectors like finance and healthcare, with high market expectations for AI's use in more complex, unstructured data environments. This research directly applies AI's predictive power to the highly challenging early-stage VC domain, seeking to overcome the limitations of traditional investment models.
In-Depth AI Insights
What does AI surpassing human investors in predicting startup success mean for the future of early-stage venture capital? - This signals a profound structural transformation within the VC industry. AI is no longer merely an assistive tool but is poised to directly participate in, or even lead, the critical function of "discovering" high-potential startups. - The traditionally human-centric and network-dependent deal sourcing function is at risk of commoditization. Early-stage funds that cannot offer significant value-add beyond capital, such as strategic guidance or industry connections, will see their competitive edge diminish significantly. - Investment decisions will become more data-driven and objective, helping to mitigate investment misses or biases (e.g., related to founder background, gender, or geography) often present in human evaluations. How should venture capital firms and founders adapt to this AI-driven industry transformation? - For VC firms, the core competency will shift from "discovery" to "empowerment." Successful VCs of the future will need to focus on providing deep operational support, market expansion, and talent acquisition as value-added services to their portfolio companies, rather than just capital. - We might see the emergence of new VC models, such as "AI-first VCs," utilizing proprietary AI models for large-scale, efficient screening, with human experts then conducting deeper due diligence and post-investment management. - For founders, the widespread adoption of AI means their business plans and team potential will be more easily discovered, regardless of their geographic location or network connections. They will need to focus more on product refinement and data performance, as these will become key metrics for AI evaluation. What are the potential negative implications or challenges that this shift might bring? - Despite claims of bias reduction, AI models' training data may contain historical biases, potentially causing AI to inadvertently replicate or even amplify these biases, especially when evaluating non-traditional or disruptive innovations. - Over-reliance on AI models could lead to a "black box problem," where investors struggle to fully comprehend the logic behind AI's decisions, thereby reducing their control over investment outcomes. - Industry concentration might increase further, with a few giant VC firms possessing the most advanced AI models potentially dominating the early-stage investment market, squeezing out smaller or emerging funds.