Investing in AI startups is hard

Rapidly declining costs for computational intelligence has lead to a profusion of copycat companies competing for the same pool of customers

Sorting AI users into buckets

My mental model for sorting AI users into various buckets is basically the following:

1. Foundational model builders (Google, Meta, Microsoft/OpenAI, etc.)

2. Entrepreneurs building on top of LLMs (GPT wrappers, AI-powered SaaS startups, etc.)

3. End users (consumers and/or businesses)

For the foundational model builders: LLMs are a sales funnel for cloud computing contracts. There are still a lot of companies that do on prem computing, and being able to layer, say, OpenAI’s models on top of Microsoft Azure may be a great sales channel for Microsoft to acquire more cloud computing customers.

For entrepreneurs building on top of LLMs: Decline in cost of computational intelligence is a huge risk: for each category of AI startup, there is a profusion of similar, easily copyable companies all competing with each other for the same pool of customers.

For end users: Great times. Tons of options, declining costs, increased power and capability.

The problem for venture capitalists is that they mainly invest in the second bucket. (There is some VC investment in the first bucket, but the vast majority of AI-related investment, outside of the hyperscalers’ capex spend, seems to be going to companies building on top of LLMs.)

Where this leaves us

The AI Investment Landscape

1. Oversaturation and Commoditization

   1. The decreasing costs of AI compute and the availability of powerful pre-trained models like GPT4 and its ilk, mean that barriers to entry are low. This leads to an influx of startups vying for the same customers, driving down potential returns.

2. Vertical Niches and Proprietary Data

   1. Vertical Niches: Startups targeting specific industries can develop highly tailored solutions that generic AI models can’t match. For example, AI applications in healthcare, legal tech, or finance can leverage industry-specific knowledge and regulations.

   2. Proprietary Data: Access to unique, high-quality data sets can provide a competitive advantage. This data is often not available in the public domain, and can improve the performance and accuracy of AI models significantly.

The VC Perspective

1. Market Size Concerns

   1. VCs usually look for large addressable markets to ensure substantial returns. Vertical niches typically represent a small addressable market, and this market size often compares unfavorably to more generalizable, horizontal applications.

2. Balancing Risk and Reward

   1. Investing in niche startups requires a deeper understanding of the industry and its specific challenges. Many, though not all, VCs prefer to spread their risk across broader, more versatile applications.

Strategies for AI Startups

1. Develop Deep Expertise

   1. This requires developing deep expertise and strong relationships with industry stakeholders. This can create a moat that is hard for generalist AI companies, whether startup or incumbent hyperscaler, to cross.

2. Build Strong Partnerships

   1. Collaborate with established players in the industry to ensure access to proprietary data, which will help refine AI models for specific applications. These partnerships may also lead to acquistion possibilities further down the line.

3. Explore Non-Traditional Niches

   1. There are emerging areas such as agri-tech, climate tech, and education where AI can create significant impact. These may be overlooked, yet still have substantial growth potential.

Conclusions

AI tech is no doubt very cool, and AI is no doubt the future. It is hard, though, to get excited about it as an investment prospect: most of the AI startups are in a war with many competitors for the same small set of customers.

This is, in some sense, no different from other startup bubbles. But because the cost of compute of the underlying substrate—the large language models—is declining towards zero, entrepreneurs and venture capitalists excitedly see oportunites whose costs are declining at a rapid rate. But if costs are declining, then everyone wants to join the party.

If everyone joins the party, then there’s no alpha for most AI startups.

Startups that focus on vertical niches, and which have access to specialized, proprietary data not captured by LLMs may accrue value. But therein lies the problem: many VCs avoid vertical niches because they reasonably see a ceiling to market size.