YC W26 Demo Day In-Depth Review: The Truth Behind 200 Startups, Copilot Is Dead, AI Agents Fully Take Over

Original Title: What I Learned From 199 Pitches at the YC W26 Demo Day
Original Author: Rathin Shah, Ex-founder of Spenny
Original Translation: Deep Tide TechFlow

Deep Tide Guide: This is not just a simple Demo Day observation report. After attending 199 pitches live, the author uses data and case studies to reveal the underlying logic of current AI startups: why 60% of companies are all-in on AI, why the copilot concept has almost disappeared, and why the fastest revenue-generating companies are founders “selling back to their old employers.”

More importantly, he points out the deadly risks behind seemingly popular tracks, as well as those overlooked areas that could produce the next legend.

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

Core Lessons for Founders

About Market/Problem Statements

1. AI is not a category; it is infrastructure. 60% of the batch are AI-native. Another 26% are AI-enabled. Only 14% have no AI. The question isn’t “Are you using AI?” but “What foundational models does your AI do things that out-of-the-box models can’t?”

2. Replacement, not assistance. The core theme is “AI employees,” not copilot or helper. The pitch is always “we replace [expensive human roles] end-to-end,” with pricing a small fraction of that person’s salary. Copilot is assistance. Agent is action. The industry has moved forward.

3. Find your “Claude Code” in your field. Every profession now has structured outputs AI can generate: contracts, CAD files, financial models, surgical plans, specifications. Find a profession with an hourly rate of $100-500+, tools with 10-30 years of history, and clear validation steps. Broad fields include tax planning, civil engineering, management consulting, clinical trials, patent drafting, music production.

4. Consider service models. About 20% of the batch are building AI-native service companies (law, recruiting, accounting, insurance), charging results-based fees but enjoying software profit margins. They show the fastest revenue growth in the batch. Pattern: start with services → generate revenue and data → automate → upgrade to platform.

5. B2B dominance. AI agents replace B2B knowledge workers. 87% are B2B. Only 14 companies (about 7%) are consumer-facing. Current AI capabilities perfectly match business workflows. This is a good business, but the legendary companies in this batch are likely the outliers: uranium exploration, lunar hotels, robot cowboys, parasitic drug companies.

6. Build data flywheels. Every customer interaction should improve your product. LegalOS trained on 12,000 visa applications → 100% approval rate. Improves perfectly with each hire. Without a data flywheel, you’re just a wrapper.

7. Don’t build general AI wrappers. “AI for everything” loses to “AI replacing a specific $80K/year job.” Deepen into unsexy industries. The best opportunities are in industries you’d never pitch at a cocktail party.

8. Consumer absence signals opportunity. Zero education companies. Zero social apps. Zero mental health/fitness. Zero government tech. Historically, the least funded categories produce the highest abnormal returns. Founders cracking AI-native entertainment, social, or education will dominate the entire category.

9. Hardware is back. 18% of the batch have hardware components (robots, drones, wearables, space tech). This is a significant jump from recent batches. Companies founded by SpaceX/Tesla alumni building physical products are the most differentiated in the batch.

About Distribution Channels

10. Distribution channels are a prerequisite, not an afterthought. Among the top 15 growth companies, 60% gained customers through founder networks or YC networks. If your first 20 customers require “figuring out distribution,” you chose the wrong market.

11. Your previous employer is your first market. Dominant GTM actions (about 35% of B2B): founders who worked in the industry for years, left, then sell back their network. Their business cards are their distribution channels.

12. PE M&A channels are severely underestimated. Ressl AI and Robby independently found that PE-backed M&A buyers urgently need profit improvement tools. One PE deal = 50-200 outlets.

13. Choose markets where you already have distribution networks. Companies struggling with GTM are almost always those that build product 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 speed. Founders who have done the work they now want to automate can close deals in days. Others take months. Proximitty (10k+ ARR in under 3 weeks): CEO was a McKinsey banking risk advisor. Corvera (33K MRR in 4 weeks): CEO runs a CPG brand.

15. Co-founder relationships are your moat. 46% of the batch are 2-person teams. The strongest teams have worked together for years: former colleagues, classmates, siblings, repeat co-founders. If you haven’t launched something with your co-founder, you haven’t validated the most important part of entrepreneurship.

16. Domain expertise beats degrees. The most convincing founders have firsthand experience with the problem: a dentist building surgical AI, an aircraft maintenance manager building mechanical tools, a lobbyist building policy AI. “Big tech” backgrounds are the baseline, not a differentiator.

About Pitching

17. Crazy closing lines matter. When 199 companies pitch in one day, you need to be the one they talk about over drinks. “The first AI Oscar will be born on Martini.” “You can book a lunar hotel in 2032.” Make your vision concrete, falsifiable, and quotable.

About What to Avoid

18. Avoid undifferentiated agent infrastructure. 8-10 companies building agent monitoring/testing/compression. Foundational model providers will build these natively. If your description is “[existing DevOps tools] but for AI agents,” you’re in the danger zone.

19. Avoid AI-native services without data moats. Fastest revenue but lowest defensibility. Core tech can be copied in weeks. Traditional companies will adopt AI in 12-18 months. Without proprietary data or embedded distribution, moats are thin.

20. Avoid commoditized workflow wrappers. AI that performs a well-defined task, as GPT-5 might do natively in 6 months.

Live Event

199 pitches. Fresh startups out of the YC incubator have a unique vibe. Excitement, high energy, never dull.

Some unforgettable moments:

A startup pitching the first hotel on the Moon, with White House invitations and a $500 million term sheet

Robot cowboys using autonomous drones to herd cattle

An AI demo company generating its own pitch deck in real-time during the demo

A company zooming into satellite images to Tehran (the whole room went silent)

Martini founder ending with “The first AI-made movie Oscar will be won by Martini!”—such lines make investors either roll their eyes or reach for their checkbooks

Hardware demo area bustling: robots, drones, microscopes with biological proteins, vehicle radars. Real, touchable physical objects. This is not just a batch SaaS dashboard.

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

Macro Numbers

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

· Non-AI: 27 (14%)

Traction:

· Estimated median ARR: about $500k–$100K

· Estimated median growth: about 30-50% MoM

· Companies with ARR > $1M: about 5%

· No revenue: about 50%

Main Industries: B2B software (59%), industrial (15%), healthcare (10%), fintech (8%), consumer (4%).

Only 14 companies target consumers; YC officially classifies only 7 as “consumer.” The rest are consumer products under corporate labels, categorized as B2B, healthcare, or fintech.

Top 10 Themes

1. AI Agents Replacing Entire Job Functions

Core theme.

Not copilot, but complete replacement.

· Beacon Health replacing authorized administrative personnel

· Perfectly replacing recruiters end-to-end

· Lance replacing front desks at 50+ Marriott/Hilton/Hyatt hotels

· Mendral (co-founder of Docker) replacing DevOps engineers

· Canary replacing QA

The “copilot” framework dropped from about 4% of pitches early 2025 to 1% in W26.

2. “Claude Code” in X Domain

Claude Code and Cursor prove that agentized AI is effective for code. Founders in W26 are applying the same paradigm to professions 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 Business, Software Economics”)

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

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

· AI recruiting agencies (Perfectly)

· AI accounting (Balance)

· AI insurance brokers (Panta)

· AI policy consulting (Fed10, founded by three ex-lobbyists)

Panta explicitly states: “A service business with software economics.” Charging results, operating with software profit margins, because AI handles 80% of human work. Arcline has over 50 startup clients. LegalOS has 100% visa approval rate.

Bearish reasons: human-in-the-loop limits profit margins to 60-80%. Responsibility is real. Moat issue: if core tech is “LLM + domain prompts + human review,” what prevents copying? Emerging answer: start with services → automate → upgrade to platform. Service is the wedge; software is the moat.

4. Infrastructure for the Agent Era

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

· Agentic Fabriq = “Agent’s Okta”

· Sponge (ex-Stripe crypto leads) = agent’s financial infrastructure

· Moda/Sentrial = agent reliability’s Datadog

· Salus = runtime guardrails

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

Zatanna turns SaaS before LLM into queryable databases for agents.

Risk: foundational model providers will build these natively. This layer has about 30% overlap, confirming it’s crowded.

5. Vertical AI in “Unsexy” Industries

Highest ROI in industries overlooked by tech:

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

· GrazeMate builds robot cowboys, autonomous drone herders. When they pitch, you can’t help but laugh. Sounds absurd until you learn the founders grew up on a ranch with 6,000 cattle.

· OctaPulse does computer vision for fish farming

· Squid solves grid planning (annual $760B inefficiency, still using spreadsheets)

These founders go deep. Scout Out founder is fourth-generation in construction. LegalOS co-founder grew up in a family immigration law firm (over 10k hours from age 12). Zymbly co-founder was a Virgin Atlantic aircraft maintenance chief. The best opportunities are in industries you’d never pitch at a cocktail party.

6. Revival of Physical AI/Robots

18% of the batch have hardware components:

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

· Origami Robotics makes robotic hands

· RoboDock’s 60-day MVP deployment went viral, securing a $100K Waymo contract

· Fort (ex-Tesla engineers) tracks strength training; Whoop/Oura still can’t do this

· Pocket shipped over 30k units, with annual revenue of $27 million

Hardware demo area is the most lively part of the day.

7. Defense and National Security

Milliray (three PhDs from Oxford/St. Andrews) built drone detection radars for NATO (batch sales of $470K)

Seeing Systems built AI attack drones for the UK Royal Marines

DAIVIN! built underwater diving gear for US special forces

Big defense budgets, long contracts, transferable credibility to commercial markets.

8. Data as Moat

When everyone has the same foundational models, proprietary data becomes the main defensive moat:

· Shofo: the world’s largest indexed video library

· Human Archive: collected data from thousands of families in Asia, after dropping out from Stanford/Berkeley, for humanoid robots

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

Pattern: each customer interaction makes the product better. Without a data flywheel, you’re just a wrapper.

9. Deep Tech and Space

Most daring pitches. GRU Space is building the first hotel on the Moon before 2032. During their pitch, the room recalibrated: half thought they were crazy, half thought they might do it. $500M term sheet, White House invitations, over 1 billion views. Beyond Reach Labs is building orbital solar arrays the size of a football field (power needs increase 500x by 2030). Terranox uses AI to discover uranium deposits (single discovery worth $200-700 million).

Ditto Biosciences might have the most creative thesis: parasites evolved proteins that control the human immune system over millions of years. Ditto uses AI to identify and design immune therapies. Evolution has already solved the problem; they just read the answers.

10. AI-native Research and Science

Talking Computers deploy an AI scientist fleet (ARR over $1 million)

Aemon (twin brothers, published papers at ICLR/EMNLP before age 20) set world records on NP-hard math problems with less than $10 in compute, beating Google DeepMind

Ndea, co-founded by Mike Knoop of Zapier and François Chollet of Keras, aims to build truly innovative AGI.

Founders: Patterns from 429 People

Demographics:

· About 60% are immigrants/international

· 86% male, 14% female

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

· 55% studied CS; 45% did not

Backgrounds:

· About 30% from big tech

· About 25% have previous startups

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

· Only SpaceX has about 12 founders, most building hardware and aerospace

Team:

46% are 2-person teams, 15% solo

Most common prototype: two technical co-founders with different expertise (about 35%), not the classic “hacker + sales”

19% of companies have at least one PhD founder

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

Becoming a domain expert founder is the most convincing story: Adrian Kilian (dentist → Mango Medical surgical AI), Robbie Bourke (20+ years in aerospace → Zymbly), Pamir Ehsas (external legal counsel for OpenAI → Arcline), Conor Jones (years inside the power grid → Squid).

Some observations:

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

The most successful teams either built and sold companies together before or worked side-by-side in the same company solving the same problems they now tackle.

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

How They Find Markets

B2B (88% of the batch)

“Experienced this pain point myself” (about 40%): the strongest pattern. End Close founder spent 6 years handling over $1 trillion in payments at Modern Treasury. Squid worked inside the national grid for years. They don’t need customer discovery—they are the customer.

“Built a platform to replace” (about 20%): Docker co-founded Mendral. TikTok’s ML scientist built Perfectly. They deeply understand architecture and see where AI creates leapfrogs.

“50+ dialogue sprints” (about 15%): systematic discovery. Ritivel had 50+ pharma conversations before coding. Ressl AI started with consulting, finding the most glue work in deals.

“Infrastructure prophecy” (about 15%): hypothesis-driven. “If agents exist, they need certification” → Agentic Fabriq. Risk: building for a future 2-3 years out.

“Research to commercialization” (about 10%): CellType (Yale professor + Google DeepMind). Valgo co-founders really wrote the textbook on safety-critical systems.

B2C (7% of the batch)

“I am the user” (about 50%): Fort founder is a weightlifter disappointed with wearables. Doomersion founder combines short videos and language learning.

“Format transformation” (about 25%): existing behaviors + new media. Pax Historia: love for strategy games + AI-generated alternate history.

“Hardware wedge” (about 25%): physical products create data loops that software can’t copy.

Meta-lesson: No successful W26 company was born from hackathons or “what if we do with AI” brainstorms. Every one stems from deep personal experience or obsessive customer discovery.

How They Find Distribution Channels

Data is clear: founder networks are the fastest-growing mechanism for B2B companies. Among the top 15, 60% gained their first customers through founder or YC networks.

B2B Patterns:

“Sell to former employer peers” (about 35%): Fed10’s three ex-lobbyists, their business cards are their distribution.

“YC as launchpad” (about 25%): Cardinal does outbound for 40+ YC companies, Palus Finance signed 33 in a few weeks.

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

“PE M&A channels” (about 8%): one deal equals 50-200 outlets.

“Systematic outbound” (about 15%): limited buyer list with quantifiable pain points.

“Wedge products” (about 7%): narrow entry points, expanding everywhere.

B2C: product itself is the distribution channel. Doomersion got 15K downloads in 2 weeks with zero paid marketing. Pax Historia built tens of thousands of DAU through organic growth. Hardware founders bet on physical presence creating word-of-mouth.

Biggest takeaway: Companies struggling with GTM almost always build product first and then ask “how do we sell?” Winners ask “who can I already reach, and what do they desperately need?” and build that.

Excellent Pitch Analysis

Seven components distinguish memorable pitches from vague ones:

1. Hook

Three effective prototypes:

Shocking data: “Bringing a drug to market takes 500k days. We want to do it in 5 days” (Rhizome AI)

Reframing: “Every file you upload uses the 1974 protocol” (Byteport)

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

2. Specific, concrete problem

“Half of technical work is paperwork” (Zymbly) beats “automating backend workflows.”

3. (One-sentence) Credibility bomb

“Andrea wrote Docker’s first line of code” (Mendral). “Our team invented the MPIC standard protecting every HTTPS connection on the internet” (Crosslayer Labs).

4. Market (inevitable, not just big)

“Satellite power demand: increase 500x before 2030” (Beyond Reach Labs). Explains why now and why it’s inevitable, not just TAM size.

5. Traction (speed > absolute numbers)

“$33K MRR in 4 weeks” (Corvera) beats “100K ARR” without a timeframe.

6. Unique insight

“Parasites evolved proteins that control the human immune system over millions of years. We read their answers” (Ditto Bio). “Insurers can’t price autonomous systems because of no historical claims data” (Valgo).

7. Crazy closing lines

“The first AI Oscar will be won on Martini.” “Book a lunar hotel in 2032” (GRU Space).

Vague pitches: generic “AI for [industry],” team credentials unrelated to the problem, and (key) no crazy closing lines.

Overlap: YC’s Multiple Bets

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

High overlap: LLM context compression (Token Company vs. Compressr), medical legal docs (Wayco vs. Docura Health), robot data (Human Archive vs. Asimov)

Medium: legal startups (Arcline vs. General Legal vs. Vector Legal), AI SRE (IncidentFox vs. Sonarly), agent monitoring (Sentrial vs. Moda), pre-authorization (Ruma Care vs. ClaimGlide vs. Beacon Health)

What it tells you: YC bets on markets, not companies. Three legal startups = a large, real market with room for multiple winners. Two companies that look similar on Demo Day will be completely different by Series A. The most differentiated companies have zero overlap: Terranox, Zymbly, GrazeMate, Ditto Bio. In each case, founders’ domain expertise is their moat.

Obvious Absences

· Zero education companies

· Zero government tech

· Zero social apps

· Zero mental health/fitness

· Almost no market presence

· Almost no pure crypto (blockchain used as pipeline, never as product point)

· Consumer companies at historic lows (only 14 total, 7 officially categorized)

Industry share jumped from 3.6% in W24 to 14.1% in W26, a 4x increase.

“The shift from atoms to bits” is real within YC.

Reverse interpretation: W26’s composition is a snapshot of what’s fundable now, not what will be valuable in 10 years. The missing legends in this batch are consumer and social founders, who will arrive in 2-3 batches once AI capabilities match their ambitions.

What Could Fail

Undifferentiated agent infrastructure. 8-10 companies building agent monitoring/testing/compression. Foundational model providers will build these natively. Enterprise buyers will default to existing vendors.

AI-native services without data moats. Fastest revenue but lowest defensibility. Core tech can be copied in weeks. Traditional companies will adopt AI in 12-18 months. Without proprietary data or embedded distribution, moats are thin.

Single-person technical founders in the sales market. Construction, insurance, freight: if no one can speak the language on-site, growth stalls.

“AI for [industry]” without domain depth. Flag: starting descriptions with “We use advanced LLM agents…” instead of specific customer pain points.

Long-cycle deep tech with no revenue. Conceptually sound but prone to burn cash.

Commodity workflow wrappers. Single-task AI, GPT-5 might do the same natively in 6 months.

Five Traits of the Fastest Companies

1. Sell results, not tools

2. Founders have customer relationships before product exists

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

4. Customers are desperate, not curious (Proximitty: banks with $2B+ bad loans; Ruma Care: clinics rejected $150K reimbursement)

5. MVP is embarrassingly simple: they describe results, not architecture

The gap between “launch and learn” and “build and hope” is where most failures happen in this batch.

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

Original Link

Click to learn about BlockBeats’ job openings

Join the official BlockBeats community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Group Chat: https://t.me/BlockBeats_App

Twitter Official Account: https://twitter.com/BlockBeatsAsia