Author: Rathin Shah

Compiled by: Deep Tide TechFlow

Introduction: This is not just a simple Demo Day observation report. After attending 199 pitches, the author uses data and case studies to reveal the underlying logic of current AI startups: why 60% of companies are all in AI, why the copilot concept has almost disappeared, and why the founders who "sold back to their old companies" are the ones who generate revenue the fastest. More importantly, he points out the fatal risks behind seemingly hot sectors and those overlooked yet potentially legendary untapped areas.

I attended YC Winter Demo Day 2026. 199 companies. Here are my observations: data, patterns, and everything you need to know if you're a future founder.

Core lessons for founders

Statement regarding the market/issue

1. AI isn't a category; it's infrastructure. 60% of batches are AI-native. Another 26% are AI-enabled. Only 14% have no AI. The question isn't "Are you using AI?" but rather "What things has your AI done that basic models couldn't do out of the box?"

2. Replacement, not assistance. The core theme is "AI employees," not copilots, not assistants. The pitch will always be "We provide end-to-end replacement for [expensive human roles]," priced at a fraction of that person's salary. Copilots are assistants. Agents are the agents. The industry has moved forward.

3. Find the "Claude Code" for your field. Every profession has structured outputs that AI can currently generate: contracts, CAD files, financial models, surgical plans, specifications. Find a profession where practitioners earn $100-$500 per hour or more, the tools are 10-30 years old, and there are clearly defined validation procedures. Broad fields include: tax planning, civil engineering, management consulting, clinical trials, patent drafting, and music production.

4. Consider the service model. Approximately 20% of batches are building AI-native service companies (legal, recruitment, accounting, insurance), charging based on results but enjoying software profit margins. They demonstrate the fastest revenue growth among batches. The pattern is: start with services → generate revenue and data → release automation → upgrade to a platform.

5. B2B Dominance. AI agents are replacing B2B knowledge workers. 87% are B2B. Only 14 companies are consumer-facing (approximately 7%). Current AI capabilities are unlocking a perfect match for business workflows. This is a good business, but the legendary companies in this batch are likely to be the outliers: uranium exploration companies, lunar hotels, robot cowboys, and parasite drug companies.

6. Build a data flywheel. Every customer interaction should make your product better. LegalOS, trained on 12,000 visa applications → 100% approval rate. Perfectly improves with every hire. Without a data flywheel, you're just a wrapper.

7. Don't build a generic AI wrapper. "AI for everything" loses to "AI replacing a specific $80,000-a-year job." Delve into an unsexy industry. The best opportunities are in industries you'd never pitch at a cocktail party.

8. The absence of consumers is an opportunity signal. Zero education companies. Zero consumer social media. Zero mental health/fitness. Zero government technology. Historically, the least funded category has produced the largest outlier returns. Founders who crack AI-native entertainment, social, or education will dominate the entire category.

9. Hardware is making a comeback. 18% of the batches included hardware components (robotics, drones, wearables, space technology). This is a significant jump from recent batches. The physical product companies founded by SpaceX/Tesla alumni are the most differentiated in these batches.

Regarding distribution channels

10. Distribution channels are a prerequisite, not an afterthought. 60% of the top 15 fastest-growing companies acquire customers through founder networks or the Y Combinator network. If your top 20 customers need to "figure out distribution channels," you've chosen the wrong market.

11. Your former employer is your first market. The dominant GTM (Gross-Total Marketing) strategy (approximately 35% of B2B): Founders spend years in the industry, leave, and then sell back their network. Their business cards become the distribution channel.

12. The PE M&A channel is severely underestimated. Ressl AI and Robby independently discovered that PE-backed M&A firms urgently need profit improvement tools. One PE deal equals 50-200 sites.

13. Choose markets where you already have a distribution network. Companies struggling with GTM are almost always those that built the product first and then asked, "How do we sell it?" Winners ask, "Who can I already reach, and what do they urgently need?"

About the team

14. Founder-market fit is the strongest predictor of revenue velocity. Founders who have actually done the work they're now automating can close deals within days. Others take months. Proximitty ($700,000 ARR in less than 3 weeks): CEO is a risk advisor at McKinsey & Company. Corvera ($33,000 MRR in 4 weeks): CEO runs the CPG brand.

15. Your co-founder relationships are your moat. 46% of batches are 2-person teams. The strongest team collaborations have lasted for years: former colleagues, classmates, siblings, and repeat co-founders. If you haven't released anything with your co-founders, you haven't validated the most important part of starting a business.

16. Domain expertise trumps academic qualifications. The most compelling founders have firsthand experience with the challenges: dentists building surgical AI, aircraft maintenance supervisors building machine tools, and lobbyists building policy AI. Having worked at a major tech company is the foundation, not a differentiating factor.

About the roadshow

17. A crazy closing is important. When 199 companies are pitching in one day, you need to be the one they're talking about while drinking. "The first AI Oscars will be born at Martini." "You can book a hotel on the moon in 2032." Make your vision specific, falsifiable, and quotable.

What to avoid

18. Avoid undifferentiated agent infrastructure. 8-10 companies are building agent monitoring/testing/compression. The underlying model provider will build these natively. If "[existing DevOps tools] but for AI agents" describes you, that's a danger zone.

19. Avoid AI-native services without a data moat. Fastest revenue but lowest defensiveness. Core technologies can be replicated in weeks. Traditional companies adopt AI in 12-18 months. Without proprietary data or embedded distribution, the moat is thin.

20. Avoid commoditized workflow wrappers. AI performs a well-defined task, while GPT-5 can natively do the same thing within 6 months.

on site

199 pitches. Fresh startups coming out of the YC oven have a unique smell: exciting, high-energy, and never dull.

Some unforgettable moments:

A startup is pitching its first hotel on the moon, with an invitation from the White House and a $500 million letter of intent.

Robot cowboys use autonomous drones to herd cattle.

An AI demo company generates its own pitch deck in real time during a demo.

A company randomly zoomed in on Tehran, Iran, during a satellite image demonstration (the entire room went silent).

Martini's founder ended the interview with the line, "Martini will win the first Oscar for an AI-made film!"—a line that would either make investors roll their eyes or reach for their checkbooks.

The hardware demo area was bustling with activity: robots, drones, microscopes with life science proteins, vehicle radar. Real, tangible physical objects. This is more than just a batch SaaS dashboard.

After listening to 199 pitches, you stop hearing about individual companies and start seeing patterns. Here are my findings.

Macro figures

Total number of companies: 199

Business Model:

B2B: 174 (87%)

B2C: 14 (7%)

B2B2C: 11 (6%)

Product type:

Pure software: 163 (82%)

Hardware + Software: 24 (12%)

Pure hardware: 12 (6%)

AI Classification:

AI-native (AI is the product): 120 (60%)

AI-enabled (existing workflows + AI): 52 (26%)

Non-AI: 27 (14%)

Traction:

Estimated median ARR: Approximately $50,000 to $100,000

Estimated median growth: Approximately 30-50% MoM

Companies with an ARR greater than $1 million: Approximately 5%

No income: Approximately 50%

Major industries: B2B software (59%), industrial (15%), healthcare (10%), fintech (8%), and consumer (4%).

Of the 14 companies that target consumers, Y Combinator officially categorizes only 7 as "consumer." The rest are consumer products disguised as businesses, falling under the categories of B2B, healthcare, or fintech.

Top Ten Themes

1. AI Agents Replace Entire Job Functions

The core theme: Not copilot, but a complete replacement.

Beacon Health replaces pre-authorized administrative personnel

Perfectly end-to-end replacement for recruiters

Lance replaces the front desk staff at 50+ Marriott/Hilton/Hyatt hotels.

Mendral (co-founder of Docker) replaces DevOps engineers

Canary as a replacement for QA

The "copilot" framework saw a decrease from approximately 4% of roadshows in early 2025 to 1% in W26.

2. "Claude Code in the X Domain"

Claude Code and Cursor have proven that agent-based AI works effectively with code. The founders of W26 are applying the same paradigm to every profession with structured output:

REV1 for Mechanical Engineers (3D→2D Drawings)

Avoice for architects (specifications, documents)

Synthetic Sciences for Scientific Research

Maywood for investment bankers

Alt-X for real estate underwriting (work directly in Excel)

Cardboard for video editing

Mango Medical generates surgical plans in minutes instead of days.

3. AI-native professional services ("service business, software economics")

Instead of building tools for existing companies, we'll build AI companies that compete with them:

Four AI law firms (Arcline, General Legal, Vector Legal, LegalOS)

AI recruitment agency (Perfectly)

AI Accounting (Balance)

AI Insurance Broker (Panta)

AI Policy Consulting (Fed10, founded by three former lobbyists)

Panta explicitly states: "A service business with software economics." It charges based on results and operates on software profit margins because AI does 80% of the work that humans do (20%). Arcline has 50+ startup clients. LegalOS boasts a 100% visa approval rate.

Reasons for a bearish outlook: Human intervention limits profit margins to 60-80%. The responsibility is real. The moat issue: If the core technology is "LLM + domain hints + human review," what prevents replication? Emerging answer: Start with a service → release automation → upgrade to a platform. The service is the wedge; the software is the moat.

4. Infrastructure in the Agent Era

Each technology stack layer is rebuilding the agent:

Agentic Fabriq = "Agent's Okta"

Sponge (three former Stripe crypto heads) = agent's financial infrastructure

Moda/Sentrial = Datadog for agent reliability

Salus = Runtime guardrail

21st (1.4 million developers) = AI-first UI React components

Zatanna transforms the previous SaaS model of LLM into an agent-queryable database.

Risk: These are built natively by the underlying model providers. The approximately 30% competitive overlap in this layer confirms its congestion.

5. Vertical AI in the "unsexy" industry

The highest ROI is found in industries neglected by technology:

Zymbly automates aircraft maintenance paperwork (5 minutes of repair requires 45 minutes of documentation).

GrazeMate is building robotic cowboys and autonomous drones for herding cattle. When they pitch, you can't help but laugh. It sounds absurd until you learn the founders grew up on a ranch with 6,000 cows.

OctaPulse uses computer vision for fish farming.

Squid solves grid planning (a $760 billion annual problem of inefficiency still using spreadsheets).

These founders have deep expertise. The founder of Scout Out is a fourth-generation construction worker. The co-founder of LegalOS grew up in a family immigration law firm (each with over 10,000 hours of experience since age 12). The co-founder of Zymbly was the head of aircraft maintenance at Virgin Atlantic. The best opportunities are in industries you'd never be pitching at a cocktail party.

6. The resurgence of physical AI/robotics

18% of the batches contain hardware components:

Remy AI and Servo7 built warehouse robots that learn from human demonstrations (80% of warehouses are zero-automated).

Origami Robotics builds robotic arms

RoboDock's 60-day deployment MVP goes viral, securing a $100,000 Waymo contract.

Fort (three former Tesla engineers) tracked strength training, something Whoop/Oura still can't replicate.

Pocket has shipped over 30,000 units, generating annualized revenue of $27 million.

The hardware demonstration area was the most vibrant part of the day.

7. National Defense and National Security

Milliray (three Oxford/St. Andrews PhDs) built a drone detection radar for NATO (batch sales of $470,000).

Seeing Systems builds AI-powered strike drones for the British Royal Marines.

DAIVIN! builds tankless diving equipment for US Special Operations Forces

Defense budgets are large, contracts are long, and credit can be transferred to commerce.

8. Data is the moat.

When everyone has the same underlying model, proprietary data is the primary defense:

Shofo: The world's largest indexed video library

Human Archive: Dropped out of Stanford/Berkeley, moved to Asia, and collected data from thousands of families for humanoid robots.

LegalOS: 12,000 successful visa applications → 100% approval rate

Model: Every customer interaction makes the product better. Without a data flywheel, you're just a packaging machine.

9. Hard Technology and Space

The most audacious roadshow ever. GRU Space is building the first hotel on the moon by 2032. When they pitched, the rooms were recalibrated: half thought they were crazy, and half thought they could do it. $500 million in letters of intent, a White House invitation, and over 1 billion views. Beyond Reach Labs is building a solar array the size of an orbital football field (power requirements will increase 500 times by 2030). Terranox uses AI to discover a uranium deposit (single discovery = $200-700 million).

Ditto Biosciences' most innovative argument is this: parasites evolved proteins that control the human immune system over millions of years. Ditto uses AI to identify them and design its own immune therapies. Evolution has already solved the problem; they're just reading the answers.

10. AI-native research and science

Talking Computers deploys a fleet of AI scientists (ARR over $1 million)

Aemon (twin brothers, who published papers at ICLR/EMNLP before the age of 20) set a world record for solving an NP-hard math problem with less than $10, beating Google DeepMind.

Ndea, co-founded by Mike Knoop of Zapier and François Chollet, creator of Keras, is clearly designed to build an innovative AGI (Automatic Generative Intelligence).

Founders: A model from 429 people

Demographics:

Approximately 60% of immigrants/internationals

86% male, 14% female

Top schools: Berkeley (approximately 45), Stanford (approximately 35), MIT (approximately 20), Waterloo (approximately 15).

55% studied Computer Science; 45% did not.

background:

Approximately 30% of former major companies

Approximately 25% had prior entrepreneurial experience.

Approximately 12% of former finance/trading firms (Citadel, Jane Street, Jump)

SpaceX alone has about 12 founders, the vast majority of whom are involved in hardware and aerospace construction.

team:

46% are teams of two, and 15% are solo.

The most common prototype: two tech co-founders with different expertise (about 35%), not the classic "hacker + sales" combination.

19% of companies have at least one PhD founder.

How they met: approximately 35% were university classmates, approximately 25% were former colleagues, approximately 15% were repeat co-founders, and approximately 10% were family/siblings.

Becoming a domain expert who is also a founder is the most compelling story: Adrian Kilian (dentist → Mango Medical surgical AI), Robbie Bourke (25 years in the aviation industry → Zymbly), Pamir Ehsas (external legal counsel for OpenAI → Arcline), Conor Jones (many years within State Grid → Squid).

Some observations:

Deep domain expertise + collaborative technical capabilities = the strongest company in a batch

The most successful teams either previously built and sold companies together, or worked side-by-side within the same company to solve the same problem they are now facing.

31% of the companies have at least one founder with a PhD or who is a researcher, primarily focusing on medical/biotechnology, hard technology, and AI infrastructure.

How did they find the market?

B2B (88% of batches)

"I've personally experienced this pain point" (approximately 40%): The strongest model. The founder of End Close spent six years at the Modern Treasury processing over $1 trillion in payments. The founder of Squid spent years within the State Grid. They don't need customers to discover them; they are the customers.

"I built this platform to replace" (approximately 20%): Docker co-created Mendral. TikTok's ML scientists built Perfectly. They have a deep understanding of the architecture and see where AI can create a step-by-step change.

"50-Conversation Sprint" (approximately 15%): Systematic Discovery. Ritivel conducted 50+ pharmaceutical conversations before writing code. Ressl AI started with consultations and discovered that transactions involved the most glue work.

"Infrastructure Prophecy" (approximately 15%): Argument-driven. "If agents exist, they need to be certified" → Agentic Fabriq. Risk: Building for the future 2-3 years from now.

"Research → Commercialization" (approximately 10%): CellType (Yale professor + DeepMind). Valgo co-founders actually wrote a textbook on safety-critical systems.

B2C (7% of batch)

"I am a user" (approximately 50%): The founder of Fort is a weightlifter who was disillusioned with wearable devices. The founder of Doomersion combined watching short videos and learning languages.

"Format Conversion" (approximately 25%): Existing behavior + new medium. Pax Historia: A passion for strategy games + AI replacing history.

"Hardware wedge" (approximately 25%): Physical products create data loops that software cannot replicate.

The fundamental lesson: No successful W26 company was born from a hackathon or a "What if we did it with AI..." brainstorming session. Each one stemmed from deep personal experience or a passionate discovery of customer needs.

How did they find distribution channels?

The data is clear: Founders networks are the number one mechanism for the fastest-growing B2B companies. 60% of the top 15 fastest-growing companies acquired their first customers through founder networks or the Y Combinator network.

B2B model:

"Selling to former employers and peers" (approximately 35%): Three former lobbyists for Fed10 used their business card holders as distribution channels.

"YC as a launchpad" (approximately 25%): Cardinal handles outbound calls for 40+ YC companies, and Palus Finance signed 33 deals within weeks.

"Open source" (approximately 10%): 21st Century has 1.4 million developers, but it's only effective for infrastructure.

"PE M&A Channel" (approximately 8%): One transaction = 50-200 branches

"Systematic outbound calling" (approximately 15%): Limited buyer list with quantifiable pain points.

"Wedge-shaped products" (approximately 7%): narrow entry point, expanding in all directions.

B2C: The product itself is the distribution channel. Doomersion gained 15,000 downloads in 2 weeks with zero paid marketing. Pax Historia built tens of thousands of daily active users (DAU) through organic growth. Hardware founders bet on physical existence to generate word-of-mouth.

Biggest takeaway: Companies that struggle with GTM are almost always the ones that build the product first and then ask, "How do we sell it?" Winners ask, "Who can I reach, and what do they urgently need?", and then build that.

Analysis of Excellent Roadshows

Seven components distinguish a memorable roadshow from something vague:

1. Hook

Three prototypes are valid:

Shocking data: "It takes 500,000 days to bring a drug to market. We want to reduce it to 5 days." (Rhizome AI)

Reframed: "Every file you've uploaded uses the 1974 license" (Byteport).

"I am the problem": "I spent six years building reconciliation at the Modern Treasury, handling $1 trillion." (End Close)

2. The problem (specific, not general)

"Technicians spend half their time on paperwork" (Zymbly) is better than "our automated back-end workflows".

3. Team (A one-sentence reputation bomb)

"Andrea wrote the first line of code for Docker" (Mendral). "Our team invented the MPIC standard to protect every HTTPS connection on the Internet" (Crosslayer Labs).

4. Market (inevitable, not just large)

"Satellite power requirements: 500-fold increase by 2030" (Beyond Reach Labs). The strongest market roadshow explains why now and why this is inevitable, not just how large TAM is.

5. Traction force (speed > absolute value)

"$33,000 MRR in 0 to 4 weeks" (Corvera) outperforms "$100,000 ARR" without a time frame.

6. Unique Insights

"Parasites have evolved proteins that control the human immune system. We read their answers." (Ditto Bio) "Insurance companies cannot price autonomous systems because historical claims data is unavailable." (Valgo)

7. A Crazy Closing Remark

"The first AI Oscar will be born on Martini." "Book the lunar hotel for 2032" (GRU Space).

The pitch was vague: a general "AI for [industry]", no connection between the team's qualifications and the questions, and (crucially) no catchy closing remarks.

Overlapping Competition: YC's Multiple Bets

Approximately 30% of companies have direct competitors within a batch. Only about 5% face truly high overlap.

High overlap: LLM context compression (Token Company vs. Compresr), medical legal documentation (Wayco vs. Docura Health), robotic data (Human Archive vs. Asimov)

Intermediate Level: Entrepreneurial Law (Arcline vs. General Legal vs. Vector Legal), AI SRE (IncidentFox vs. Sonarly), Agent Monitoring (Sentrial vs. Moda), Prior Authorization (Ruma Care vs. ClaimGlide vs. Beacon Health)

What it tells you: Y Combinator bets on the market, not the companies. Three startup law firms = a market that's real and big enough to accommodate multiple winners. Two companies that look identical at Demo Day will be completely different by Series A. The most differentiated companies have zero overlap: Terranox, Zymbly, GrazeMate, Ditto Bio. In each case, the founder's domain expertise is the moat.

Obvious absence

Zero Education Company

Zero Government Technology

Zero-cost social networking

Zero mental health/fitness

Almost zero market

Almost zero pure encryption (blockchain is used as a conduit, never as a product argument).

Consumer demand is at a historic low (out of 14 companies, only 7 are officially categorized).

Industry jumped from 3.6% in W24 to 14.1% in W26, a fourfold increase. The "atom vs. bit" shift is real within YC.

A reverse interpretation: W26 is a snapshot of what's currently fundable, not what will be valuable 10 years from now. The legendary companies missing from this batch are those consumer and social founders who will arrive in 2-3 batches, once AI capabilities catch up with their ambitions.

What might fail?

Undifferentiated agent infrastructure. 8-10 companies handle agent monitoring/testing/compression. The base model provider will natively build these. Enterprise buyers default to their existing vendors.

AI-native services lacking a data moat. Fastest revenue generation, lowest defensiveness. Core technologies can be replicated in weeks. Traditional companies adopt AI in 12-18 months.

A lone tech founder in the relationship-based sales market. Construction, insurance, freight: stagnation occurs if no one can understand the jargon of the construction industry.

"AI for [industry]" lacks domain depth. A hallmark: descriptions that begin with "We use advanced LLM agents..." instead of addressing specific customer pain points.

Long-term, deep tech projects with no revenue. The concept is correct, but the failure model is burning through all the money.

Commercial workflow wrappers. Single-task AI, GPT-5 could potentially do the same thing natively within 6 months.

The fastest companies share five characteristics

1. Sell results, not tools.

2. The founders had customer relationships before the product existed.

3. Charged from day 1: No free tier, no pilot purgatory.

4. Customers are desperate, not curious (Proximitty: a bank with over $2 billion in bad loans; Ruma Care: a clinic that was denied $150,000 in reimbursement).

5. MVPs are awkwardly simple: they describe the outcome, not the architecture.

The gap between "launch and learn" and "build and hope" is where most of the deaths in this batch will occur.

The future is exciting! There has never been a better time to build.

Written on March 25, 2026, a few days after YC W26 Demo Day.