Author: Rathin Shah

Translated by: Deep Tide TechFlow

Deep Tide's Introduction: This is not a simple Demo Day observation report. After attending 199 pitches live, the author reveals the underlying logic of AI entrepreneurship today with data and examples: why 60% of companies are all in on AI, why the copilot concept has almost disappeared, and why the founders who receive revenue the fastest are those who sold back to their old employers. More importantly, he points out the fatal risks behind those seemingly hot tracks and the overlooked blank areas that could give birth to the next legend.

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

Core Lessons for Founders

About Market/Problem Statement

1. AI is not a category; it is infrastructure. 60% of the batch is AI-native. Another 26% is AI-enabled. Only 14% do not involve AI. The question is not "Are you using AI?" but "What does your AI do that basic models out of the box cannot?"

2. Replacement, not assistance. The core theme is "AI employees," not copilot, not assistants. The pitch is always "we end-to-end replace [expensive human roles]," priced at a small fraction of that person's salary. Copilot is assistance. Agents are action. The industry has moved forward.

3. Find your "Claude Code" for your field. Every profession has structured outputs AI can now generate: contracts, CAD files, financial models, surgical plans, specifications. Look for occupations with hourly rates above $100-$500, tools that have been around for 10-30 years, and clear validation steps. Broad fields: tax planning, civil engineering, management consulting, clinical trials, patent drafting, music production.

4. Consider service models. About 20% of the batch is establishing AI-native service companies (law, recruiting, accounting, insurance), charging based on results while enjoying software profit margins. They showed the fastest revenue growth in the batch. The pattern is: start with services → gain revenue and data → release automation → upgrade to platform.

5. B2B dominates. AI agents replace B2B knowledge workers. 87% are B2B. Only 14 consumer-facing companies (about 7%). Current AI capabilities unlock perfect matches for business workflows. It’s a good business, but the legendary companies in this batch are likely to be outliers: uranium mining companies, lunar hotels, robotic cowboys, 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 improving with every hire. Without a data flywheel, you are just a wrapper.

7. Don’t build generic AI wrappers. "AI for everything" loses to "AI replacing a specific $80,000 salaried position." Dig deep into an unsexy industry. The best opportunities are in industries you would never pitch at a cocktail party.

8. The absence of consumers is an opportunity signal. Zero education companies. Zero consumer social. Zero mental health/fitness. Zero government tech. Historically, the least funded categories have produced the greatest abnormal returns. Founders cracking AI-native entertainment, social, or education will monopolize entire categories.

9. Hardware is back. 18% of the batch has hardware components (robots, drones, wearables, space technology). This is a significant jump from recent batches. The physical product companies built by SpaceX/Tesla alumni are the most differentiated in the batch.

About Distribution Channels

10. Distribution channels are prerequisites, not afterthoughts. 60% of the top 15 growth companies acquired customers through founder networks or YC networks. If your first 20 customers need to "figure out distribution channels," you have chosen the wrong market.

11. Your former employer is your first market. Dominant GTM moves (about 35% of B2B): founders have spent years in the industry, left, and then sold back to their networks. Their business cards are distribution channels.

12. PE acquisition channels are severely underestimated. Ressl AI and Robby independently found that PE-backed acquirers urgently need profit improvement tools. One PE deal = 50-200 points of sale.

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

About Teams

14. Founder-market fit is the strongest predictor of revenue velocity. Founders who have actually done the work they are now automating can close deals in days. Others take months. Proximitty (less than 3 weeks to $700k ARR): CEO was a risk consultant at McKinsey. Corvera ($33k MRR in 4 weeks): CEO ran a CPG brand.

15. Your co-founding relationship is your moat. 46% of the batch are two-person teams. The strongest teams have worked together for years: former colleagues, classmates, siblings, repeat co-founders. If you haven’t released anything with your co-founder, you haven’t validated the most important part of entrepreneurship.

16. Domain expertise beats degrees. The most compelling founders have personally experienced the problem: a dentist building surgical AI, an aircraft maintenance manager building mechanic tools, a lobbyist building policy AI. "Former big company" is the baseline, not the differentiator.

About Pitches

17. A crazy ending is important. When 199 companies pitch in one day, you need to be the one they talk about while drinking. "The first AI Oscar will be born at Martini." "You can book a lunar hotel for 2032." Make your vision specific, falsifiable, and quotable.

About What to Avoid

18. Avoid undifferentiated agent infrastructure. 8-10 companies are building agent monitoring/testing/compression. Foundational model providers will natively build these. If "[existing DevOps tools] but for AI agents" describes you, you are in dangerous territory.

19. Avoid AI-native services without data moats. Fastest revenue but least defensible. Core technology can be replicated in weeks. Traditional companies will adopt AI within 12-18 months. Without proprietary data or embedded distribution, the moat is thin.

20. Avoid commoditized workflow wrappers. AI does a clearly defined task, while GPT-5 might natively do the same thing in 6 months.

On-site

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

Some memorable moments:

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

A robotic cowboy herds cattle with autonomous drones.

An AI demo company generates its pitch deck live during the demo.

A company casually zooms in on Tehran, Iran, while demonstrating satellite images (the whole room went quiet).

The founder of Martini ends with, "The first Oscar for an AI-made movie will be won by Martini!" This line either made investors roll their eyes or pull out their checkbooks.

The hardware demo zone was bustling: robots, drones, microscopes with life science proteins, onboard radar. Real, tangible physical things. This was not just a batch SaaS dashboard.

After attending 199 pitches, you no longer hear individual companies but start to see patterns. Here are my findings.

Macro Numbers

Total Number of Companies: 199

Business Models:

B2B: 174 (87%)

B2C: 14 (7%)

B2B2C: 11 (6%)

Product Types:

Pure Software: 163 (82%)

Hardware + Software: 24 (12%)

Pure Hardware: 12 (6%)

AI Classification:

AI Native (AI is the product): 120 (60%)

AI Empowered (existing workflows + AI): 52 (26%)

Non-AI: 27 (14%)

Traction:

Estimated Median ARR: Around $50,000 - $100,000

Estimated Median Growth: About 30-50% MoM

Companies with ARR > $1 million: About 5%

No revenue: About 50%

Major Industries: B2B Software (59%), Industrial (15%), Healthcare (10%), Fintech (8%), Consumer (4%).

Only 14 companies are consumer-facing; YC officially categorized only 7 as "consumer." The rest are consumer products disguised as enterprise, classified under B2B, healthcare, or fintech.

Top Ten Themes

1. AI Agents Replace Entire Job Functions

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

Beacon Health replaces administrative personnel who approve tasks in advance.

Perfectly end-to-end replaces recruiters.

Lance replaces front desks at over 50 Marriott/Hilton/Hyatt hotels.

Mendral (Docker co-founder) replaces DevOps engineers.

Canary replaces QA.

The "copilot" framework has dropped from about 4% in early 2025 to 1% in W26 pitches.

2. "Claude Code in X Field"

Claude Code and Cursor have proven that agentified AI is effective for code. W26 founders are applying the same paradigm to every profession with structured outputs:

REV1 for mechanical engineers (3D to 2D drawings).

Avoice for architects (specifications, documentation).

Synthetic Sciences for scientific research.

Maywood for investment bankers.

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

Cardboard for video editing.

Mango Medical generates surgical plans in minutes instead of days.

3. AI Native Professional Services ("Service Businesses, Software Economics")

Not building tools for existing companies but creating AI firms that compete with them:

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

AI recruiting agency (Perfectly).

AI accounting (Balance).

AI insurance brokers (Panta).

AI policy consulting (Fed10, founded by three former lobbyists).

Panta states clearly: "A service business with software economics." Charging based on results, operating with software profit margins, as AI does 80% and humans do 20%. Arcline has 50+ startup clients. LegalOS has a 100% visa approval rate.

Bear case: having people in the loop limits margins to 60-80%. Liability is real. Moat concerns: if the core technology is "LLM + domain prompts + human review," what stops replication? The emerging answer: start with services → release automation → upgrade to platform. Services are the wedge; software is the moat.

4. Infrastructure of the Agent Era

Every layer of the tech stack is being rebuilt for agents:

Agentic Fabriq = "Okta for agents."

Sponge (three former encryption leads from Stripe) = financial infrastructure for agents.

Moda/Sentrial = Datadog for agent reliability.

Salus = runtime guardrails.

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

Zatanna converts pre-LLM SaaS into agent-queryable databases.

Risks: foundational model providers natively build these. About 30% of competitive overlap confirms it is crowded.

5. Vertical AI in "Unsexy" Industries

The highest ROI is in industries overlooked by technology:

Zymbly automates airplane maintenance paperwork (5-minute repairs require 45 minutes of documentation).

GrazeMate builds robotic cowboys, autonomous drones for herding. When they pitched, you couldn’t help but laugh. It sounds absurd until you learn the founder grew up on a ranch with 6,000 head of cattle.

OctaPulse does computer vision for fish farming.

Squid solves grid planning (inefficiencies of $760 billion annually, still using spreadsheets).

These founders dig deep. Scout Out’s founder is a fourth-generation construction worker. LegalOS co-founder grew up in a family immigration law firm (over 10,000 hours logged per person by age 12). Zymbly co-founder was the aircraft maintenance manager at Virgin Atlantic. The best opportunities are in the industries you would never pitch at a cocktail party.

6. Physical AI/Robotics Renaissance

18% of the batch includes hardware components:

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

Origami Robotics builds robotic hands.

RoboDock gained popularity with MVP deployment in 60 days, securing a $100,000 contract with Waymo.

Fort (three former Tesla engineers) tracks strength training, something Whoop/Oura still can’t do.

Pocket shipped over 30,000 units, achieving $27 million in annual revenue.

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

7. Defense and National Security

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

Seeing Systems builds AI strike drones for the UK Royal Marines.

DAIVIN! builds tankless diving gear for US Special Operations.

Defense budgets are large, contracts long, and reputations transferable to commercial.

8. Data as a Moat

When everyone has the same foundational models, proprietary data is the primary defense:

Shofo: the world’s largest indexed video library.

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

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

Pattern: every customer interaction makes the product better. Without a data flywheel, you are merely a wrapper.

9. Hard Tech and Space

The boldest pitches. GRU Space is building the first hotel on the moon by 2032. When they pitched, the room recalibrated: half thought they were crazy, half thought they might actually do it. $500 million in letters of intent, a White House invitation, and over 1 billion views. Beyond Reach Labs builds solar arrays the size of a football field in orbit (power demand will increase 500 times by 2030). Terranox uses AI to discover uranium deposits (a single find = $200-700 million).

Ditto Biosciences may have the most creative argument: parasites have evolved proteins that control the human immune system over millions of years. Ditto uses AI to identify them and design immunotherapies. Evolution has already solved the problem; they are merely reading the answer.

10. AI Native Research and Science

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

Aemon (twin brothers, published papers at ICLR/EMNLP before age 20) computes under $10 to create a world record on NP-hard math problems, beating Google DeepMind.

Ndea, co-founded by Mike Knoop from Zapier and François Chollet, creator of Keras, focuses on building AGI that can innovate.

Founders: Patterns from 429 People

Demographics:

About 60% immigrants/international.

86% male, 14% female.

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

55% studied Computer Science; 45% did not.

Background:

About 30% from large companies.

About 25% have previous entrepreneurial experience.

About 12% from finance/trading (Citadel, Jane Street, Jump).

About 12 founders from SpaceX, the vast majority building hardware and aerospace.

Team:

46% are two-person teams, 15% are solo.

Most common prototype: two technical co-founders with different expertise (about 35%), not the classic "hacker + salesperson."

19% of companies have at least one co-founder with a PhD.

How they met: about 35% college classmates, about 25% former colleagues, about 15% repeat co-founders, about 10% family/siblings.

Founders who are domain experts have the most compelling stories: Adrian Kilian (dentist → Mango Medical surgical AI), Robbie Bourke (25 years in aviation → Zymbly), Pamir Ehsas (external legal advisor for OpenAI → Arcline), Conor Jones (many years inside the National Grid → Squid).

Some observations:

Deep domain expertise + buildable technical co-founders = the strongest companies in the batch.

The most successful teams have either built and sold companies together before or have worked side by side in the same company solving the same problem they are now trying to tackle.

31% of companies have at least one founder with a PhD or researcher background, primarily concentrated in healthcare/biotech, hard tech, and AI infrastructure.

How They Found Their Markets

B2B (88% of the batch)

"I’ve experienced this pain point" (about 40%): the strongest pattern. The End Close founder spent 6 years at Modern Treasury handling over $1 trillion in payments. The Squid founder spent many years inside the National Grid. They do not need customer discovery; they are the customers.

"I built this platform to replace" (about 20%): The Docker co-founders built Mendral. ML scientists from TikTok built Perfectly. They know the architecture intimately and see where AI creates step changes.

"50 conversation sprint" (about 15%): systematic discovery. Ritivel had over 50 pharmaceutical conversations before writing code. Ressl AI started consulting and discovered that deals had the most glue work.

"Infrastructure prophecy" (about 15%): argument-driven. "If agents exist, they need authentication" → Agentic Fabriq. Risk: building for a future 2-3 years away.

"Research → commercialization" (about 10%): CellType (Yale professor + DeepMind). Valgo co-founder actually wrote the textbook on safety-critical systems.

B2C (7% of the batch)

"I am the user" (about 50%): The Fort founder was a disappointed weightlifter with wearables. The Doomersion founder swiped through short videos and learned languages, combining them.

"Format conversion" (about 25%): existing behaviors + new media. Pax Historia: love for strategy games + AI replaced history.

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

Meta lesson: No successful W26 company was born from a hackathon or "What if we use AI to..." brainstorming session. Each arose from profound personal experience or obsessive customer discovery.

How They Found Their Distribution Channels

The data is clear: founder networks are the #1 mechanism for the fastest-growing B2B companies. 60% of the top 15 growth companies acquired their initial customers through founder networks or YC networks.

B2B Models:

"Sell to former employer peers" (about 35%): Three former lobbyists of Fed10 whose business cards serve as distribution channels.

"YC as a launchpad" (about 25%): Cardinal made outbound calls for 40+ YC companies, Palus Finance signed 33 in just a few weeks.

"Open source" (about 10%): 21st has 1.4 million developers, only effective for infrastructure.

"PE acquisition channel" (about 8%): one deal = 50-200 points of sale.

"Systematic outbound" (about 15%): limited buyer lists have quantifiable pain points.

"Wedge product" (about 7%): narrow entry, expanding everywhere.

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 DAUs, experiencing organic growth. Hardware founders bet that physical presence generates word-of-mouth.

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

Analysis of Exceptional Pitches

Seven components distinguish memorable pitches from blurry ones:

1. Hook

Three effective prototypes:

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

Reframing: "Every file you’ve uploaded uses a protocol from 1974" (Byteport).

"I am the problem": "I spent 6 years building reconciliation at Modern Treasury, handling $1 trillion" (End Close).

2. Problem (specific, not generic)

"Technicians spend half their time on paperwork" (Zymbly) beats "We automate backend workflows."

3. Team (credibility bomb in one sentence)

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

4. Market (inevitable, not just large)

"Satellite power demands will increase 500 times before 2030" (Beyond Reach Labs). The strongest market pitches explain why now and why it’s inevitable—not just how big the TAM is.

5. Traction (speed > absolute numbers)

"$33k MRR within 4 weeks" (Corvera) beats "10k ARR" without a time frame.

6. Unique Insight

"Parasites evolved proteins that control the human immune system. We read their answers" (Ditto Bio). "Insurance companies can’t price autonomous systems because historical claims data doesn’t exist" (Valgo).

7. Crazy Ending

"The first AI Oscars will be born at Martini." "Book your lunar hotel for 2032" (GRU Space).

Blurred pitches: generic "AI for [industry]," team qualifications unrelated to the problem, and (crucially) no crazy ending.

Competitive Overlap: YC's Multiple Bets

About 30% of companies in the batch have direct competitors. Only about 5% face significant overlap.

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

Medium: startup 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: YC bets on markets, not companies. Three startup law firms = the market is real and big enough to accommodate multiple winners. Two companies that look the same 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.

Notable Absences

Zero education companies.

Zero government tech.

Zero consumer social.

Zero mental health/fitness.

Almost no market.

Almost zero pure crypto (blockchain as a pipeline, never as a product argument).

Consumer is at a historical low (only 14 companies, only 7 officially categorized).

Industrial jumped from 3.6% in W24 to 14.1% in W26, a fourfold increase. The transition from "atoms vs bits" is real within YC.

Reverse interpretation: The composition of W26 is a snapshot of what is fundable now, not what will be valuable in 10 years. 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 are doing agent monitoring/testing/compression. Foundational model providers will natively build these. Enterprise buyers default to existing vendors.

AI native services without data moats. Fastest revenue, lowest defensibility. Core technology can be replicated in weeks. Traditional companies adopt AI within 12-18 months.

Solo technical founders in relationship-sold markets. Construction, insurance, freight: it stalls if no one can walk onto a job site and speak the lingo.

"AI for [industry]" without domain depth. Red flag: descriptions starting with "We use advanced LLM agents..." instead of specific customer pain points.

Long-cycle deep tech without revenue. Not conceptually wrong, but the failure mode is burning through cash.

Commoditized workflow wrappers. Single-task AI that GPT-5 may natively do the same in 6 months.

Fastest Companies Share Five Traits

1. Sell results, not tools.

2. Founders have customer relationships before the product exists.

3. Charge from day one: no free tiers, no pilot hell.

4. Customers are desperate, not curious (Proximitty: banks with over $2 billion in bad loans; Ruma Care: clinics denied $150,000 reimbursement).

5. MVP embarrassingly simple: they describe outcomes, not architecture.

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

Exciting times ahead! There has never been a better time to build.

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