Y Combinator used to fund companies that changed the internet.
| Cohort | Company |
|---|---|
| Summer 2013 | DoorDash |
| Summer 2013 | Webflow |
| Summer 2013 | Lob |
| Summer 2013 | EasyPost |
| Summer 2013 | StatusPage (acquired by Atlassian) |
| Summer 2013 | Casetext (acquired for $650M) |
| Fall 2025 | "AI receptionist for hotels" |
| Fall 2025 | "Brainrot IDE" (Cursor + Stake.com iframe) |
| Fall 2025 | Three separate "Lovable for X" companies |
| Fall 2025 | "Sales call cheating AI" |
| Fall 2025 | "AI Excel analyst for Excel power users" |
I went through the entire Fall 2025 batch, here's what I found.
I analyzed all 112 companies in the F25 batch that had public descriptions:
54.46% have "AI" in their description (61 companies)
15.18% explicitly mention "Agent" or "Agentic" (17 companies)
12.50% follow the "[Product] for [Niche]" format (14 companies)
0% admit to wrapping LLMs (they're not stupid)
15.18% explicitly mention "Agent" or "Agentic" (17 companies)
12.50% follow the "[Product] for [Niche]" format (14 companies)
0% admit to wrapping LLMs (they're not stupid)
More than half the batch is AI wrappers.
Let's compare this to Summer 2013.
S13 had 52 companies. The failures were at least trying to build real things. SpoonRocket was doing actual food logistics. BloomThat had supply chains. Standard Treasury was building banking APIs from scratch.
Even the companies that died were attempting hard problems.
F25 has companies that can be cloned in 72 hours. "AI for auditors." "AI for hotel receptionists." "TikTok for language learning" (so TikTok with a filter?).
The bar didn't lower, it's just completely vanished.
Let me tell you about Clad Labs, they're building the "CHAD IDE."
From their description:
A coding editor that mocks and integrates all the vices and distractions devs succumb to during AI coding downtime. CHAD IDE's twist is that it lets you directly shit-post on X, scroll Instagram or TikTok, play Stake, swipe Tinder, etc., all from within your IDE while waiting on AI model prompts.
This got funded.
The problem they're solving: "Devs get distracted during AI inference gaps."
The solution: Put the distractions IN the IDE so you don't have to alt-tab.
This is like saying people eat too much junk food while working from home, so you built a desk with a built-in DoorDash button and a mini-fridge of Oreos. Now you save time not walking to the kitchen!
YC gave them $500,000 for this.
The same accelerator that funded Stripe btw.
Want to know how cheap it is to build an AI wrapper in 2025?
GPT-5 mini pricing:
- $0.25 per million input tokens
- $2 per million output tokens
For $100/month in API costs, you can handle:
- 180+ million tokens of processing
- 90,000+ user conversations at 2,000 tokens each
- 3,000 daily interactions
You can build "AI for X" for less than a Netflix subscription.
Now compare that to what it took to build the companies YC used to fund:
Stripe: Two years in beta testing before public launch. The Collison brothers spent that time building payment infrastructure from scratch, personally onboarding early users, refining security features.
Dropbox: Built desktop clients for Windows/Mac/Linux. Delta encoding. Conflict resolution. LAN sync. Custom infrastructure on top of S3. Years of deep technical work.
Twitch: Live video streaming at scale when CDN costs were insane. Built custom infrastructure. Real-time chat at massive scale. Pioneered an entirely new category.
These companies couldn't be cloned in a weekend.
AI wrappers can be built in 2-4 weeks. The ecosystem launches 77 new AI wrapper products every week. There are literally two-hour workshops teaching developers to "ship an AI wrapper with advanced prompting, agents & RAG" in a single session.
YC used to fund companies where the product was the moat.
Now they fund companies where being first to a micro-niche is the "moat."
That is just a land grab in a game of musical chairs.
YC didn't always fund 140 companies per batch.
First batch (S05): 8 companies
First 10 batches average: 17 companies
2012: ~70 companies per batch
2021: 377 companies (Summer 2021)
2025: 140-160 per batch, but now they run FOUR batches per year
First 10 batches average: 17 companies
2012: ~70 companies per batch
2021: 377 companies (Summer 2021)
2025: 140-160 per batch, but now they run FOUR batches per year
They went from funding ~40 companies a year to 500+.
You can't maintain quality at that scale. The math doesn't work.
But here's the thing that makes this even more absurd: the acceptance rate actually went DOWN.
2010: ~3-4% acceptance rate
2025: 0.6% acceptance rate (Summer 2025)
2025: 0.6% acceptance rate (Summer 2025)
They're funding more companies while being "more selective."
How?
Applications exploded from ~1,000 to ~33,000+. So they can claim a lower acceptance rate while completely abandoning quality control. The denominator grew faster than the batch size.
It's a statistical trick that lets them say "we're highly selective" while funding three different "Lovable for X" companies in the same batch.
Paul Graham stepped back from YC in 2014. Since then, he's written hundreds of essays about startups, founders, and Silicon Valley.
Want to know how many times he's criticized YC's direction?
Zero.
The only time he discussed batch size issues was in 2012. YC grew from 60 to 80 startups and the system broke because they were using an O(n²) algorithm for partner-founder matching. He wrote about it in 2022 as a lesson about operational scaling.
That's it.
No essays about quality control. No commentary on AI wrappers. No discussion of whether YC lost its way.
He's either:
- Diplomatically staying out of it
- Trusting current leadership blindly
- Unwilling to admit the thing he built became what it was supposed to disrupt
PG used to write things like:
"The best way to increase a startup's growth rate is to make the product so good people recommend it to their friends."
And:
"A good startup founder is relentlessly resourceful."
Not "founders who can wrap an API."
And:
"The best startup ideas look like bad ideas but are good."
Not "the best ideas are '[successful product] for [niche].'"
The YC that rejected SendGrid for looking like a "spam company" (then changed their process after realizing it was a mistake) is gone.
Now they fund anything with "AI agent" in the name.
Paul Graham era (2005-2014):
The founding team was Paul Graham (essayist, created Viaweb), Jessica Livingston (marketer, author), Trevor Blackwell (roboticist), and Robert Morris (MIT professor).
These were philosophers and academics. They asked: "Is this idea worth doing?"
They rejected thousands of derivative ideas. They had taste. They wrote essays that shaped how an entire generation thought about startups.
Garry Tan era (2023-present):
Garry Tan runs YC now. He's a YC alum (S08, Posterous), was a partner from 2011-2015, then co-founded Initialized Capital. He's backed Coinbase and Instacart early.
The current general partners are almost all operator-founders who built billion-dollar companies. Jon Xu (FutureAdvisor, acquired by BlackRock). Andrew Miklas (PagerDuty, IPO'd). Tom Blomfield (GoCardless and Monzo).
These are people who know how to scale companies and get exits.
The filter changed.
Old YC: "Is this real?"
New YC: "Could this raise a Series A?"
New YC: "Could this raise a Series A?"
Old YC was about building the future. New YC is about portfolio optimization.
When you're funding 140 companies per batch, you're not picking winners, that's just buying Powerball tickets.
Fund enough AI wrappers and maybe one becomes Cursor ($100M ARR in 12 months). The rest can die. That's fine. The winners will return the funds.
It's Sequoia's playbook. Not YC's anymore.
Let me show you some real examples from Fall 2025:
Fastshot: "Lovable for mobile apps"
Compyle: "Lovable for Software Engineers"
Waffer: "Lovable for Animated Graphics Videos"
Compyle: "Lovable for Software Engineers"
Waffer: "Lovable for Animated Graphics Videos"
Three companies. Same concept. Same batch.
Denki: "AI for Auditors at Public Companies"
Cranston AI: "Full-service AI for accounting and finance"
Crunched: "The first Excel AI analyst built by and for Excel power users"
Cranston AI: "Full-service AI for accounting and finance"
Crunched: "The first Excel AI analyst built by and for Excel power users"
More companies solving the same problem slightly differently.
Clicks: "The first AI back-office worker that works like a human."
Mantle: "Build internal agents to automate back office work with one prompt"
Zalos: "Computer Agents for Finance tasks like reconciliation, in your system!"
Mantle: "Build internal agents to automate back office work with one prompt"
Zalos: "Computer Agents for Finance tasks like reconciliation, in your system!"
Are you seeing the pattern?
These aren't companies. These are features, or at best, distribution plays.
"We're not building better AI, we're just the first to call law firms and offer them an AI receptionist."
That's just sales arbitrage with a ChatGPT and ElevenLabs wrapper.
In September 2024, a YC-backed startup called PearAI launched. They essentially cloned Continue.dev (another YC company) and initially put a ChatGPT-generated fake license on it.
Huge backlash on Twitter. Tech media covered it extensively. People asked how YC's vetting process let this happen.
Garry Tan defended them, but that just made it even worse.
TechCrunch did a follow-up analysis and found this wasn't isolated. YC regularly funds startups that duplicate other YC companies. It's not just AI code editors. It's a pattern.
When your batch size is 140 companies and 54% of them are AI wrappers, of course you're going to fund duplicates. There are only so many "[Product] for [Industry]" combinations out there.
Fast Company published a feature in February 2025: "What is Y Combinator now?"
The article noted 87% of the Fall 2024 batch were AI companies. Penn State professor Arpita Agnihotri said: "From the entrepreneur's perspective, the core base of Y Combinator has diluted. The excitement has definitely reduced."
WebProNews reported YC's AI pivot "sparks hype debate and ethical concerns."
On Hacker News, people are asking: "What was the last truly great product out of YC built by a true hacker/team?"
Another comment: "Founders look at YC batches, see that it is 99% AI companies and are then forced to also go in that direction if they want the benefits of the accelerated YC path."
VCs are mostly staying quiet. Don't bite the hand that feeds you, right?
But everyone else sees it.
Here's what we know about YC's historical success rates:
Across all batches, about 5-5.5% of companies reach $100M+ valuations. About 6.5% become unicorns. About 10% achieve an exit.
Summer 2013 had roughly 50 companies. It produced DoorDash (now $43B+), plus several other significant exits. That batch likely hit 10-15% with $100M+ valuations.
Fall 2025 just started, so we can't measure success yet.
But here's the thing, when you fund 140 companies instead of 50, and 54% of them are AI wrappers that can be obsoleted by OpenAI's next API update, what do you think the success rate will be?
Even if the absolute number of successful companies stays the same (5-7 per batch), the percentage drops dramatically. Let's just say 1-2% instead.
YC's bet is that funding more companies increases the odds of finding the next Stripe. But it also means most of the batch is noise.
Is that worth it? For YC's fund returns, probably. For the brand and reputation? We're seeing the answer in real-time.
SendGrid is the most famous YC rejection. Robert Morris reviewed the application alone and scored it so low that no other partner looked at it. He called it a "spam company" for obvious reasons.
SendGrid went to TechStars instead and was acquired by Twilio for $2 billion.
After that, PG changed YC's process. Any application Robert Morris scored negatively would also get reviewed by Paul. They recognized the blind spot and fixed it.
Other notable rejections: Udemy, Unacademy, Veed.io, Calm, Buffer, Algolia (in 2013).
This shows PG's filter was genuinely rigorous. Even experienced partners made judgment calls that seemed reasonable (an email delivery service could look like spam) but turned out wrong.
The fact that YC systematically improved after missing SendGrid proves they took quality seriously.
Now?
"Brainrot IDE" got funded.
An IDE that adds an
<iframe /> to Stake.com and TikTok in VS Code so you can gamble and scroll while waiting for AI responses.What judgment call led to that? Who knows.
It's happening across tech, not just YC:
- Twitter killed third-party clients
- Reddit broke API access with absurd pricing
- YouTube constantly breaks downloaders
- Instagram kills scraping tools
- Spotify sends cease and desists to open-source developers
Platforms want total control. They want to be the only way to access their data. Innovation they can't monetize gets crushed.
YC is doing the same thing, just in reverse.
Instead of restricting access, they're flooding the market. Instead of quality control, they're playing quantity games. Instead of taste, they're optimizing for fund returns.
The result is the same: the ecosystem suffers.
Developers see YC batches full of AI wrappers and think "I guess that's what investors want", real innovation gets crowded out by derivative plays.
Capital flows to "[Product] for [Niche]" instead of genuinely hard problems.
If you're building something real, YC's decline is actually good news.
It means the competition is weaker. Half the batch is building ChatGPT wrappers that will be obsolete in 12 months. The other half is trying to be "Lovable for X."
If you're solving actual hard problems, you stand out more.
But it also means the YC brand is worth less. Getting into YC used to mean "smart people vetted your idea is real." Now it means "you had a decent application during the AI bubble."
The acceptance rate is 0.6%, but that's not because the bar is high, it's just that 33,000 people applied and they can only fund 200.
YC used to represent something.
It was proof that you didn't need an MBA or connections. You needed an idea and the ability to build. The Collison brothers were teenagers from rural Ireland. Brian Chesky sold cereal boxes to survive. Drew Houston made a demo video because he couldn't make Demo Day, he just had a working product and wanted to show it off.
These weren't insiders, but hackers.
And now? 50%+ of founders are from Stanford/MIT. A huge percentage are ex-FAANG. Many are second-time founders. Already funded by YC alumni angels.
And the companies being funded aren't trying to change the world, they're trying to raise a Series A.
That's what happens when you optimize for fund returns instead of impact.
YC became the thing it was supposed to disrupt: an insider's club where pedigree matters more than product.
Someone will build the next Stripe. The next Airbnb. The next company that actually matters.
It probably won't be a YC company.
Not because YC can't find them, but because when you're funding 500+ companies a year and 54% of them are AI wrappers, you're not optimizing for world-changing ideas.
You're optimizing for portfolio returns.
And that works financially, YC's portfolio is worth hundreds of billions and their fund returns are probably great by now.
But the mission itself is dead.
Paul Graham used to write about building the future, relentlessly resourceful founders and ideas that seem crazy until they work.
Now YC funds "Brainrot IDEs."
The future is being built somewhere else, but probably not YC.
All data in this post comes from YC's public batch listings and published reports. The Fall 2025 analysis is based on 112 companies with public descriptions. Batch size data compiled from TechCrunch reports, YC's blog, and Hacker News discussions.
If you're building something real and YC rejected you, don't worry, they rejected SendGrid too.