Y Combinator 2025: trends from the AI cohorts
Actualizado: 2026-05-03
Each Y Combinator cohort is an x-ray of what Silicon Valley thinks will matter in the next eighteen months. Winter 2025 (W25) graduated a few weeks ago and Summer 2025 (S25) is already selecting. Reviewing which companies entered, what business models they propose, and which verticals weigh most helps understand where smart capital is moving.
This post isn’t a list, it’s a pattern read. I care about what the set tells us about the market and what founders operating outside the valley can learn.
Key takeaways
- The vertical agent has replaced the generalist assistant as the dominant pattern: more than 70 % of W25 companies included “AI” in their description.
- Agents that win graft onto existing tools (CRM, ERP, email) rather than asking users to change how they work.
- Devtools remain a profitable niche, now oriented toward concrete developer tasks with model assistance.
- Outcome-based pricing (per task completed) appears as an emerging model that aligns incentives and removes ROI-proving friction.
- The median time from founding to YC acceptance for W25 was barely four months.
The vertical agent bias is near total
W25 consolidated a trend visible since 2024: the vertical agent has replaced the generalist assistant. Instead of “ChatGPT for everything,” we see startups with agents for claims handling, contract auditing, M&A due diligence, second-opinion medical support, or bookkeeping reconciliation. The fraction of companies with “AI” in their description exceeded 70 % of the cohort.
The thesis is consistent with what revenue data shows: customers pay to save hours on a concrete process with structured inputs, not for a generalist assistant they already have free alternatives for. An accounting SMB saving twenty hours a month on reconciliation happily pays; that same SMB won’t pay for a generic copilot.
A second pattern is deep integration with the existing workflow. Agents that win don’t ask users to change how they work, they graft onto current tools (CRM, ERP, email) and operate from there. Adoption cost determines retention, and integration is far more decisive than the quality of the underlying model. AI startup funding trends confirm investors are also rewarding this thesis.
Devtools remain a profitable niche
Another visible slice of the cohort continues betting on developer tools. The angle has changed: in 2022 they were frameworks and libraries; now they are products solving concrete tasks developers already do, done faster with model assistance.
Examples of the pattern:
- Automated code review integrated in the pull request.
- Test generation from the diff.
- Cross-framework migration with verified transformations.
- Observability with natural-language interpretation.
All share one characteristic: they monetise per active user and have a low entry point (free trial on personal repos, team expansion once value is proven).
The interesting part is business durability. Cursor, Copilot and the large players capture raw volume, but niches like language-specific review, infrastructure-as-code agents, or legacy-code migration tooling still have room for focused startups.
Business models: outcome-based pricing appears
The most striking shift in recent months is the appearance of outcome-based pricing. Instead of per-seat or usage pricing, several cohort companies propose charging per task completed: per ticket resolved, per email answered, per document processed with customer acceptance.
The model is logical: if an agent replaces human work, the customer compares against the cost of that human, not the cost of a licence. Pricing by outcome aligns incentives and removes the friction of proving ROI.
The practical difficulty is attribution. Determining when a ticket is “resolved” or a document “correctly processed” requires a verification layer many startups build with heuristics, human confirmation and audit sampling. Those getting that layer right scale at high margins; those who don’t end up negotiating murky contracts.
What European founders can copy
Four transferable lessons:
- Niche discipline. W25 companies with the most traction had very specific descriptions: not “AI for lawyers” but “AI to review commercial lease contracts in California.” A hyper-segmented niche is easier to sell, easier to evaluate and easier to dominate before expanding.
- Iteration speed. The median time from founding to YC acceptance for W25 was barely four months. Teams already had a working prototype, several paying customers and a clear hypothesis.
- Retention obsession. The number YC asks for most is the retention curve, not net user growth. A product with a thousand monthly users and 90 % scalable retention is infinitely more valuable than one with ten thousand users and 20 %.
- Narrative clarity. Publicly shared W25 pitches share a simple structure: identified customer, quantified pain, concrete solution, traction proof, ask. No technical jargon, no distant futures.
My read
W25 signals the ecosystem is consolidating business around vertical agents with deep integration and outcome-based pricing. The large models remain infrastructure: value captured by startups comes from the product layer wrapping them and the domain knowledge making them useful in a concrete flow.
What I don’t see in the cohort, and should worry some founders, are startups betting purely on the underlying model. Building an LLM from scratch is no longer a startup thesis: it’s a hyperscaler thesis. YC has clearly pivoted to the application layer.
For European founders the takeaway is twofold. There’s huge space in local verticals Americans won’t attack (specific European regulation, language, country-specific tax processes), but execution must match the speed of American teams. The European niche only protects if the team moves fast.
