New startup trends.
New tech trends start small. Y Combinator, one of the most selective startup accelerators in the world, funds thousands of new companies a year and has each one describe, in their own words, what they’re building. That’s a real, early read on what people are actually building, often before it’s widely recognized.
Each trend also gets checked against independent sources: research papers on arXiv, and what people are actually building and sharing on GitHub and Hugging Face. A trend that only appears in one place is weaker than one that shows up in all of them.
Industrial: 9× as large a share of new companies in Spring 2025 as before. Consumer Health Services: New this period. Research: Knowledge Distillation, 121 papers.
Left is stronger. Right is earlier and less certain.
Passed the strongest multi-period test. Historical total, not a score.
Radar-tier signals of early growth. Whisper, an earlier and noisier tier, is counted separately below.
Related paper groups. More papers means attention, not proof.
Categories already gaining traction
Passed the strongest test across the latest four Y Combinator groups versus the four before. Not predictions.
See the strong signals →What may be starting
Early movement that has not passed the strongest test. Possible next trends, with more uncertainty.
What researchers are writing about
Academic papers, separate from startup data. Shows where technical attention is gathering.
When the strongest signals appeared
Strong signals by the Y Combinator group where they first appeared. Attention, not a forecast.
The data points to these open questions.
Open questions attached to the strongest signals. Hypotheses, not conclusions.
See what’s next →Is activity elsewhere moving in the same direction?
This is a validation check, not the forecast. We compare YC category changes with funding, open-source projects, models, launches, accelerators, and research.
Outside activity moving in the same direction supports the YC signal. It does not prove YC caused the trend or saw it first; that requires dated history from each source. Hugging Face specifically is a loose keyword match on repo names, not a topic search — generic category names can pick up unrelated repos.