As a VC investor, my views on generative AI startup boom....

Author: Gui Shuguang

Source: Angel Tea House

Original Author: SABRINA WU / VIVEK RAMASWAMI

Image credit: Generated by Unbounded AI tools

In the past nine months, as VC investors, most of the new start-up companies/new ideas we have seen are related to artificial intelligence (AI), especially generative AI (Generative AI), which is not enough Surprised. We've seen hundreds of startup pitches in this space, but only invested in a handful of them. Apparently we’re not the only ones facing this situation, with $1.7 billion invested in GenAI startups in the first quarter of 2023 alone, and that number could increase fivefold in the second quarter.

We would like to share some of the hot topics and projects we are witnessing, the important characteristics that investors pay attention to, and the elements that distinguish "good from great" from a financial perspective. It’s still early days for this space and nothing is certain, but we hope the following is helpful to founders as they look to differentiate themselves in this increasingly competitive space.

Estimated VC investment in the generative AI subcategory (Source: Dealroom)

1. What kind of ideas do we usually see?

Early stage (Pre-Seed/Seed/A round)

In the very early stages, we see a large number of "generative-native" companies emerging. These companies are themselves built on top of the underlying model, either as an application serving the end user or as a "middleware" tool layer that sits between the model and the application.

Idea 1: Use models to generate text-based content that can create new or enhance existing text in email, knowledge bases, and other applications.

Idea 2: "X's co-pilot"; AI agents work alongside human operators to augment their ability to write code, draft presentations, and perform other tasks. We've seen a lot of co-pilot apps targeting specific vertical use cases, as well as some trying to achieve a more "personalized" co-pilot.

**Idea 3: LLM (Large Language Model) tool for managing embeddings and vector databases. **

Summary: To be a differentiated early-stage generative AI startup, having one or more moats is very important. Moats can range from unfair access to distribution, AI/ML talent, compute, data, models, or having different perspectives on the problem domain you are solving and how to create a more delightful user experience.

Early-growth period and growth period (B/C+ round)

For the companies we see in the B/C stage, they are usually born in the "pre-LLM" era and are now figuring out how to best integrate the capabilities of the base model into existing products. We call these companies "generative-enhanced" (generative-enhanced) companies, they don't necessarily need to reinvent their wheels, but make sure they don't lose out to LLM-native startups.

Creativity 1: Predictive Analytics; many large-scale SaaS companies are using AI to extract insights from their existing large data sets to more accurately predict revenue growth, customer churn rates, and other indicators.

Idea 2: Personalization and Recommendations; this is one of the fastest and most impactful ways we see growth-stage startups leveraging AI. Emergence of underlying models allows both B2B and B2C companies to provide more robust and accurate product recommendations to existing customers.

Idea 3: "Instant Auto-Complete"; In almost all growth stage companies with a text or writing component, we see LLM being used for "Instant Auto-Complete", similar to what users experience with ChatGPT.

Summary: If you haven’t already started trying to improve your business or re-architect it to be more “AI-friendly”, consider dedicating a small part of your product team to building new features.

Warning to startups entering this space: It is important to assess how much funding has been raised by generative AI companies, especially in specific subcategories. Take a look at the market landscape of more than 250 generative AI companies charted by Dealbook. Companies in model building, copywriting tools, and vector databases have raised hundreds of millions of dollars in financing. Of course, that doesn't mean another innovative startup can't be launched in this space, but it's important to note that…

2. What does "good" look like from a financial perspective?

Our understanding of what a "good" financial metric looks like for an intelligent application company is still in its early stages, but in the SaaS space, we believe that "best-in-class" growth rate is similar to the situation in the figure below. Remember, we are no longer chasing growth at all costs, so efficiency and burn rate are important factors.

Product Release Time: One of the advantages of smart applications is the ability to release products faster than ever before. We envision many smart app companies launching products in a “beta” state so they can start collecting user data and use it to create a “reinforcement learning from human feedback” (RLHF) loop. Historically, it can take a year after a product launch to reach $1 million in annual repeatable revenue (ARR), but we may see generative AI companies hit $1 million in ARR faster because customers can See return on investment (ROI) quickly. Many generative AI products also benefit from virality through product-led growth (PLG)/bottom-up sales (e.g. Jasper, Lensa, Harvey, Tome, etc.).

Customer Retention: While a generative AI company may attract new customers quickly, it may also have a higher churn rate. For a SaaS company, a good gross retention rate is around 85%-95%, and best-in-class is closer to 95%+. In terms of net retention, we think a good rate is 110%-120%+, best case is 120%+. A higher churn rate could be due to the model consistently producing wrong outputs, the emergence of other competing products, etc. A big factor in the PLG approach in the case of smart applications is that it's very easy for customers to try a new product or pay $10-20 a month, only to churn quickly.

Cost of Goods Sold (COGS) and Gross Margin: We expect many smart application companies to have new costs related to: 1) models; 2) training and fine-tuning; 3) facility management operations. We've heard that the cost of running queries on these LLM and vector database stores (via companies like Pinecone) has been high. In many cases, we've heard that customers may run queries on a model until they get the output they want, and since they pay per license, the number of queries run has a material impact on cost. As a result, we expect AI-driven companies to likely see lower gross margins.

3. What is the difference between "good" and "excellent"?

As with any other technology or industry, as VC investors, we still ultimately evaluate great teams, huge markets, and a keen understanding of customer pain points. These basic principles will not change:

**Customer-centric/solve real pain points: **In any new technological change, we will see many new companies just trying to "follow the trend" and create "cool" technology, but they don't really solve customers' problems pain points. The first question to understand is: you are solving a "hair on fire" problem, is generative AI a better way to help solve this problem, or is it an unnecessary technology?

Team: In this new era of LLM, the opportunity to build new products and start companies has been very democratized. As a result, we see many founding teams starting businesses in areas where they have little industry knowledge or expertise. The question to understand is: why is your team best suited to solve this problem?

Ability to Adapt and Execute Quickly: There is no doubt that this field is evolving rapidly. Now more than ever, it is important for teams to be agile and quickly adjust products and strategies as needed. At the same time, it's important to stick to the fundamentals and not just chase the hype. In other words: how will you react and understand when is the right time to make a potential adjustment to the company?

Reproducibility: While AI can help companies get off the ground faster, it also means there may be far more competitors in a category than there used to be. Just look at the publicly released maps of the generative AI market landscape and the money pouring into the category. Good founders and teams recognize where there are unique holes to fill and largely avoid subfields where they can quickly get lost in the clutter.

4 Conclusion

As VC investors, we are as excited and optimistic as anyone about the full impact AI will have. However, from the hundreds of project pitches we've seen over the past year, it's clear that there's a lot of hype in the category, and it's more important than ever for founders to differentiate and stand out , and finally prove the value of the product.

Some other notes:

**Valuation:**Although the overall VC market has declined relative to its peak in 2021, AI (especially generative AI) funding and valuations are still high. This reflects VC and founder interest in the space, but it's important to note that, like any other cycle (like the dot-com bubble and bust), only a small fraction of startups ultimately survive to exit, with valuations following In the next few years, it may drop by more than 90%.

Generative Native vs. Generative Augmented: As a generative native company, what can you build that a generative augmented company can’t? As a new startup entering a category, what is the meaningful difference between you and existing companies? Big tech companies like Microsoft, Google, and Amazon are already rapidly adopting LLM, so understanding where you can effectively compete with them is key.

BUDGET CONSTRAINTS: With the macro environment challenging and budgets tightening, it's important to understand the real necessity of your product. In previous bull markets, almost any SaaS product could generate several million dollars in revenue. In the current environment and the ongoing (albeit waning) recession risk, Target's chief information officers (CIOs) are looking at every corporate expense to see which ones can be cut. Will incorporating AI into your product help them or ultimately irrelevant?

-----(end of text)-----

Original address: