Vertical Foundation Models: what YC wants and how to validate it
Original request by Diana Hu, Y Combinator. Source: YC Requests for Startups.
The request in plain English
YC wants research labs that build specialized foundation models for industries like manufacturing, energy, and biology. The models should work zero-shot or from a few examples, and match or beat general models while using far less compute.
What YC's request says
- Target domains named: manufacturing, energy, and biology.
- Example: Google's TimesFM-3, a 330M-parameter model that forecasts related time series in one forward pass without task-specific fine-tuning.
- Challenges YC names: high-quality verifiable data, data rights, and proving clear accuracy and compute gains over existing specialist models.
Why now
General models are trained mostly on text and code. Many industries run on other kinds of data: sensor streams, time series, molecular structures, process logs. A model built for that data can be far smaller and far better at the task than a general model asked to reason about it in words.
The TimesFM example shows the pattern. A relatively small model trained on the right data can do zero-shot forecasting that would otherwise need a custom model per problem. That is a product: one model that replaces many bespoke ones.
What decides who wins
Data access decides most of it. The labs that win will have partnerships or proprietary sources that others cannot easily copy, plus a clear legal right to train on them.
The second factor is proof. Industrial buyers already use specialist models built in-house. A new vertical model has to show measurable accuracy gains or cost savings on the customer's own data before anyone switches.
Three startup ideas to start from
Foundation model for industrial sensor data
Zero-shot anomaly detection and forecasting across machines, sold to manufacturers tired of building one model per line.
Grid and energy forecasting model
Load, generation, and price forecasting for utilities and energy traders, benchmarked against their current models.
Lab data model for biology
A model trained on experimental data that predicts outcomes before running the experiment.
How to validate it before you build
- Secure one data partnership with clear training rights before raising for compute.
- Benchmark against the customer's existing specialist model on their data, not on a public leaderboard.
- Measure inference cost per prediction, since lower compute is part of the pitch.
- Find three buyers who would replace an in-house model if you beat it by a stated margin.
The founder trap
Building an impressive model with no route to the data it needs to stay ahead. In vertical AI, the data moat matters more than the architecture.
IdeaCheck's take
Technically deep with real demand. Teams that already have domain data or a partner who does should move first; everyone else faces a long road to a defensible model.
Test your version
Check your idea against Hacker News
Building something in ai models? See what 28,000 Hacker News threads say before you commit.
Your first evaluation is free with Google sign-in.