I always wanted to analyze VC startup batches. Study what they build. Write about the trends. But I never had time. Now we have LLMs, so here we go:
Y Combinator's Summer 2025 batch marks a shift. 88% of the ~160 companies are AI-native[1]. Not AI-powered as a feature. AI-native as the core.
This isn't hype. It's industrialization.
The winning startups do one of two things:
Three forces define this batch:
For founders: "AI-powered" means nothing now.
You need deep domain expertise + proprietary data + complete workflow solution. Or build the infrastructure everyone depends on.
The shift happened fast. Each YC batch in 2025 moved the industry forward:
Winter 2025: Companies introduced "AI agents" as a new category. The idea: software that can act autonomously, not just respond to prompts.
Spring 2025: Everyone built "Cursor for X"—AI assistants modeled after Cursor (the AI code editor). These tools help humans work faster. A copywriter gets an AI that suggests better headlines. A designer gets an AI that generates variations.
Summer 2025: The next step. Instead of assisting humans, these systems own entire workflows[2]. The human doesn't get faster—the human disappears.
The shift: copilot → autopilot
Solva: $245K ARR in 10 weeks[3]
Keystone: Fixes bugs autonomously[3]
Same pattern everywhere: Mortgage applications. Warehouse logistics. Medical records. Complete workflows, automated.
"AI-powered" is no longer a moat. Everyone says it.
Three things you must have:
YC partners push every founder: "Why this AI agent?" Domain depth is the answer[2].
30% of the batch builds developer tools, cloud infrastructure, or AI tooling[4]. That's high.
This is the picks-and-shovels play.
When everyone's building AI agents, everyone needs the same infrastructure. Five categories emerge:
Deploying AI agents is harder than deploying a website. Agents make API calls, maintain state, run background tasks, and need to scale when usage spikes.
Dedalus Labs ("Vercel for AI agents")[2] solves this:
The infrastructure complexity disappears. You focus on building your agent, they handle keeping it running.
Pricing AI products is chaos. You want to charge a subscription, but also meter usage (tokens, API calls). Add credits for prepaid customers. Add enterprise add-ons. Stripe wasn't built for this.
Autumn ("Stripe for AI")[2] handles the complexity:
Already processing payments for hundreds of AI apps and 40 YC startups. When everyone has the same pain point, solving it once for everyone becomes valuable.
You ship an AI feature. A week later, customers complain it's giving wrong answers. How do you know your AI doesn't hallucinate? How do you measure if the new model version is actually better?
Design Arena solves this by crowdsourcing rankings[2]:
Large AI labs pay for this data to improve their models. The feedback loop closes: better evaluation → better training data → better models.
You build an AI tool for healthcare. An enterprise wants to buy it. First question: "How do you handle HIPAA compliance? Do you log patient data? Where are the servers? What about the EU AI Act?"
You're a 3-person startup. You don't have a compliance team.
BitPatrol, Casco, Mindfort[2] build tools for this:
Enterprise customers demand this before signing contracts. It's not optional anymore. One data breach kills your company.
20 software development companies—the largest single category[5]
We're past "Can AI do this?" and into "How do we operationalize AI at scale?"
This is the same pattern as the 2000s web boom:
Phase 1 (early 2000s): People built web apps. Social networks, e-commerce, SaaS tools. Everyone focused on the application layer.
Phase 2 (mid 2000s): Builders realized they were all solving the same problems. Hosting. Payments. Version control. Authentication.
Phase 3 (late 2000s): AWS (2006), Stripe (2010), GitHub (2008), Twilio (2008) emerged. Infrastructure companies. They enabled everyone else. Eventually, infrastructure became half the market cap.
We're in Phase 2 for AI right now. Everyone building agents needs the same things. The infrastructure companies emerging today will be worth billions.
Winter 2025: 10% weekly revenue growth across the entire batch[6]
YC CEO Garry Tan: "That's never happened before in early-stage venture"
Summer 2025 companies hitting significant ARR within weeks:
Multiple companies: $1M–$10M ARR with fewer than 10 people[2]
1. AI coding tools compress timelines
2. Enterprise buyers fast-track trials
3. YC's network provides immediate traction
4. Usage-based pricing generates immediate revenue
Five years ago, having revenue at Demo Day was impressive. Three years ago, it was expected. Now it's baseline.
Having a live product with paying customers by Demo Day[9] is now baseline, not exceptional.
What this means for founders: The competition moved faster. You can't spend 3 months just "building in stealth." You need to ship fast, get customers fast, and prove the business model works—all during the YC batch.
The startups that can't do this don't get follow-on funding. Simple as that.
80–85% of Summer 2025 is B2B or enterprise-focused[4]
Only 15–20% build consumer products
This represents a fundamental shift from YC's early consumer app days.
Think about YC's famous companies from the 2000s: Airbnb (consumer), Dropbox (consumer), Reddit (consumer), Instacart (consumer). Those days are gone.
Now it's all enterprise software, B2B SaaS, and infrastructure tools. The incentives changed.
1. Clearer willingness to pay
2. Faster revenue
3. Funding environment
4. Risk profile
The few consumer companies that made it aren't about AI at all. They're about genuine human needs:
RealRoots: Women struggle to make friends after college. Existing apps don't work—they're awkward, forced, full of flakes. RealRoots uses AI to match compatible women and coordinate group experiences. $782K monthly revenue[4] from people paying to solve loneliness.
Pingo AI: People want to learn languages by speaking, not by doing Duolingo exercises. Pingo gives you an AI conversation partner that sounds native and corrects you naturally. $250K monthly revenue, 70% monthly growth[7].
They work because they solve real problems that existed before AI. The technology is the means, not the end.
Notice: You could build both of these without AI—they'd just be worse. The AI makes the product work at scale. That's different from "AI demo looking for a problem."
After several software-heavy batches, Summer 2025 brought back hard tech[2]:
Robotics
Advanced Semiconductors
Defense
Biotech
Climate
AI enables the solution, but the application targets physical-world problems.
AI-designed chips. AI-controlled robots. AI-optimized weather models.
Hard tech combined with AI creates real moats that pure software can't replicate:
Regulatory barriers: Defense contractors need government clearances. Medical devices need FDA approval. These take years and millions of dollars. Your competitor can't just copy your code.
Hardware complexity: Building robots isn't like deploying a web app. You need supply chains, manufacturing, physical testing. Capital intensive. Long iteration cycles. Hard to replicate.
Domain expertise: Understanding drone warfare or semiconductor physics takes years. You can't learn it from blog posts. The knowledge itself becomes a moat.
Pure software AI products have none of these advantages. Anyone can clone your UI, use the same API, and undercut your price.
Summer 2025 founders: young with strong technical depth[1]
YC's average founder age: ~29 years → ~26 years[1]
A pattern from tech history is repeating: the "mafia."
The PayPal Mafia gave us Tesla (Elon Musk), LinkedIn (Reid Hoffman), YouTube (Jawed Karim, Steve Chen), Palantir (Peter Thiel). People who worked together at PayPal went on to build the next generation of companies.
Now it's happening with AI research labs. Clusters from DeepMind, Google Brain, MIT AI labs are spawning companies together. They have the technical depth. They know what's possible with AI because they built the foundation models.
Where founders come from predicts what they build. Ex-DeepMind researchers build agent infrastructure. Ex-Google engineers build AI dev tools. The pattern is clear.
The negative space matters[2]
What's missing tells the story:
Crypto/Web3: Virtually absent
Pure consumer social apps: Very few "next Instagram" plays
EdTech and GovTech: YC flagged these as least common
"Grow now, monetize later": This playbook is dead
General-purpose AI platforms: Nobody's building the next foundation model
When 88% of a batch is AI-native[2], overlap is inevitable.
Look at any vertical: AI for code. At least 10 companies in this batch alone. AI for bookkeeping. Another 5-7 companies. AI for customer service. Same story.
These aren't all going to survive as standalone businesses. Some will get acquired. Some will pivot. Some will shut down. Consolidation is coming.
The winners will be those who moved fastest, built the best product, or locked in customers early. Everyone else becomes a footnote.
Many AI startups are essentially UI layers over APIs[2]. They call OpenAI's API or Anthropic's API, add a nice interface, and sell it as a product.
The problem: two existential threats.
First threat: The API provider (OpenAI, Anthropic, Google) decides to add the same feature you built. Suddenly you're competing with the company whose API you depend on. They have the model, the distribution, and the brand.
Second threat: Your competitor can replicate your product easily. If you're just a UI over someone else's API, anyone else can build the same UI. No proprietary technology. No data moat. Easy to copy.
AI startups targeting healthcare, finance, insurance, or defense face complex compliance requirements[2].
In healthcare: HIPAA regulations govern how you handle patient data. One violation and you're done.
In finance: You need licenses, audits, and compliance programs before you can process transactions.
In Europe: The EU AI Act just created new obligations for "high-risk AI systems." Insurance underwriting, credit scoring, hiring—all regulated now.
The problem for startups: You can build a working product in three months. Getting through compliance takes a year or more. Enterprise customers won't buy until you're compliant. Startups that ignore this hit walls when they try to sell.
Most YC startups raise a seed round that lasts 12-18 months[2]. After that, they need a Series A to keep going.
The problem: Seed capital is flowing easily right now for AI startups. Investors are excited, checks are getting written.
But Series A is different. Those investors want to see real traction. Revenue growth. Customer retention. Market fit. The bar is higher than it used to be.
What happens: Startups that can't show strong metrics when their seed money runs out will struggle to raise a Series A. That's the cliff—you either have the metrics to jump to the next level, or you don't. No more "we'll figure out business model later."
1. Don't build another general AI agent
2. Infrastructure plays are high-conviction
3. Domain expertise is the new advantage
4. Revenue from day one
5. AI + physical world is undercrowded
The opportunity shifted. The winners won't build impressive AI demos. They'll solve real problems completely, or build the infrastructure making everyone else's AI reliable.
[1]: Y Combinator S25 Batch Analysis: 160 AI Startups & Founder Insights | Extruct AI
[2]: Y Combinator S25 Batch Profile and AI Trends - Catalaize
[3]: The 9 most sought-after startups from YC Demo Day | TechCrunch
[4]: Y Combinator S25 Batch Profile and AI Trends - Catalaize
[5]: Coding with LLMs in the summer of 2025 – an update | Hacker News
[6]: Y Combinator startups are fastest growing, most profitable in fund history because of AI | CNBC
[7]: YC's Latest Trendsetters: 9 Startups Preferred by Top VCs | 36Kr
[8]: A quarter of startups in YC's current cohort have codebases that are almost entirely AI-generated | TechCrunch
[9]: YC Summer 2025 Demo Day | Y Combinator
[10]: Exclusive: YC-backed Perseus Defense Closes $6M Seed Round | Tectonic Defense
This article generated by AI in February 2026