The introduction of ChatGPT in November 2022 marked a pivotal moment, ushering us into the Age of AI, transforming the tech landscape forever. Since December, nearly every pitch deck I’ve encountered prominently features AI within its first two pages.
However, as with any emerging technology, venture capitalists, including myself, must swiftly adapt our strategies to distinguish high-potential startups from those that are merely overhyped or likely to encounter challenges that impede their growth to venture scale. Understanding this distinction necessitates fluency in the various layers of the generative AI value stack, enabling us to identify promising investment opportunities and devise thorough due diligence strategies to assess the risks and prospects associated with specific startups.
The generative AI landscape encompasses:
- Data
- Middleware
- Fine-tuned specialized models
- The cloud and infrastructure layer
- Foundational models
- The application layer
Within this tech stack, we have identified several areas ripe for investment, while others present significant competition challenges for seed-stage companies. Here’s our detailed analysis.
Areas We’re Focused On
Data
One of the foremost challenges in generative AI—and consequently one of its greatest opportunities—is ensuring the accuracy and reliability of information. Generative AI models currently rely on vast datasets, some as expansive as the internet itself, which include valuable insights alongside a plethora of unrelated data.
We anticipate a future where generative AI applications integrate more precise data, utilizing specialized models tailored to particular domains. These targeted models will leverage proprietary data, enhancing personalization and accuracy for application outputs. By combining foundational models with proprietary data and the appropriate middleware architecture, we expect these specialized models to drive the application layer that consumers and businesses engage with.
Middleware
Middleware encompasses the tools and infrastructure supporting the creation of new generative AI applications, marking the second pillar of our investment strategy in the sector.
We see substantial promise in infrastructure and tooling companies that ensure safety, accuracy, and privacy across model outputs; orchestrate inference across various models; and facilitate the integration of proprietary data into large language models (LLMs).
Fine-Tuned Specialized Models
Integrating proprietary data with foundational models through effective middleware architecture generates specialized models capable of powering applications for consumers and businesses. These applications can establish strong competitive advantages, not only by accessing proprietary data and specialized models but also through traditional benefits like distribution and user experience.
Areas We’re Avoiding
Cloud and Infrastructure
While hardware and software infrastructure—such as semiconductor and chip manufacturers, along with cloud data centers that support GPU computing—play crucial roles in generative AI, they are dominated by a few major players like Google, Microsoft, and Amazon. Consequently, seed-stage companies face significant obstacles in competing within these domains.
Foundational Models
Pioneers in generative AI, such as OpenAI, Cohere, and Stability, have built foundational LLMs that are now well-known within the tech industry. These models incorporate hundreds of billions of parameters and have required substantial time and investment to develop. The significant funding histories of these companies reflect the immense resources needed to establish their products, making it virtually impossible for seed-stage firms to gain traction in this arena.
Investment Themes within the Tech Stack
Vertical Applications
We believe that the most successful generative AI-powered applications will extend beyond foundational models, incorporating ensembles of specialized models tailored for specific workflows. These purpose-built AI tools will rely on proprietary data unique to their industries, enhancing output quality and personalization. However, value is diminished if the required data is largely public. Connections to proprietary data sources, such as customer warehouses or integrated applications, can significantly elevate a startup’s potential for lasting impact. Therefore, not only must your AI outputs be defensible, but the surrounding software and automation must also be robust.
We see considerable opportunities for this technology across various sectors, including legal, healthcare, finance, retail, logistics, manufacturing, and hospitality—particularly in areas like document analysis, human resources, process automation, generative design, and agent support.
ML Middleware
Foundational models rarely operate effectively “out of the box.” Developers often need to navigate complex steps involving model orchestration, operations, prompt engineering, and safety measures.
More advanced and flexible tooling will empower developers seeking to harness generative AI, democratizing access to foundational model stacks for a broader range of innovators. This transition will stimulate development and adoption in the space, leading to significant advancements in developer frameworks, data sources, evaluation methods, and “model operations.”
Diligence
When evaluating startups within these promising investment categories, a nuanced due diligence process is vital for accurately capturing each company's risks and opportunities in the generative AI realm. We have integrated our established best practices for assessing early-stage software companies into a comprehensive evaluation framework specific to AI, encompassing several key areas:
- Team: We prioritize dedicated AI experts over founders who are simply pursuing the latest trend.
- AI Stack and Architecture: The choice of models remains critical. The architecture that developers construct around foundational models is vital, influencing aspects such as prompt engineering, storage and retrieval mechanisms, and user experience (UX) design.
- Data: A thorough evaluation of the proprietary data available to the company is essential, as well as their model training methodologies.
- Training: It's crucial to examine how startups benchmark their models against foundational ones and assess the accuracy and improvement potential of their offerings.
- Unit Economics: New AI models introduce unique compute costs, necessitating a fresh unit economics analysis for each generative AI venture. In today's macroeconomic climate, sound unit economics are critically important.
- David versus Goliath: This framework assesses how incumbents with established distribution channels might integrate generative AI into their existing product offerings. We dedicate significant time to this defensibility analysis for each generative AI opportunity we explore.
Ranking
