What I've learned as a VC from over 100 ML Pitches
The state of ML companies raising capital
Boom, a hot ML deal is in your inbox. You call everyone you know, or maybe you keep it for yourself if you're going to write a huge check and there's limited allocation. But wait, why is this ML deal so hot? It's got fancy advisors, a famous AI founder who put in $25K as an angel, and the founder went through YC. Ooooo, ah, oooo, but wait, what's the product!?
As I work with exceptionally talented AI/ML engineers and researchers, a common theme I encounter is extraordinarily gifted founders with no product in their pitch, let alone a GTM. It's not to say that these founders are incapable of carving out a market to go after; it's just that the quality of the tech has taken up so much mind share in our industry that sometimes we forget that spinning up an API to call your model might not be a long term business plan, especially in a croweded foundation model market. Yeah, you might hit the top of the hugging face leaderboard and see a massive bump in usage via your API, but then the trouble frequently starts.
Due to the standardization of ML infra and swappable models via API calls, your API gets swapped out, and its next model is up whenever someone beats your model on a benchmark; this leads to a considerable series A drop-off graduation rate, mainly impacted by transient userbases who swap API’s, resulting in terrible LTM metrics. Retention, CaC/LTV, WAU or DAU, and NRR all fall off - many famous multi-billion dollar AI/ML companies are specifically seeing unacceptable user retention levels, per public data.
The Answer? Build a product for a vertical, not a model for the world. Use your model to enable your product. A few significant players will win direct model-based model sales, and everyone else will fight over the crumbs; I encourage you not to be in the fight for the crumbs. An LLM or multi-modal foundation model is great when deployed for an industry with a workflow tool built around it; this creates the opportunity to develop and qualify a pipe, do meaningful user interviews for PLG, and get enterprise sales rolling. Don't build an LLM for lawyers; build a workflow tool that solves an existing problem that lawyers face and use a foundation model to enable your solution to be a cut above the pack.
If you're building and want to send a deck my way, don't hesitate to get in touch with me at decks@meridianstreet.co!