Farewell to the Era of Simply Writing Code: YC Partner Discusses How Experienced Entrepreneurs and Independent Founders Use AI to Break the Deadlock
Garry Tan
President & CEO, Ycombinator
Original Statement
1. Core Data and Macro Changes in Venture Capital
• Surge in Hard Tech Proportion: Among the projects selected by YC in the past 12-18 months, the proportion of hard tech (projects involving the physical world/atoms) has jumped from 8% to 20%.
• Distribution of Sub-sectors: Robotics has risen from 1% to 6%-7%; domestic industrial manufacturing has climbed from 4% to 10%; defense technology has increased from 1.5% to about 5%; semiconductors and photonics are close to 4%; energy and computing power infrastructure has risen to nearly 3%.
• Change in Founder Educational Background: In the current summer batch, 1 in every 6 founders holds a PhD, indicating a significant increase in the demand for a strong research and academic background in deep tech fields.
• Business Launch and Revenue Growth Break Historical Records:
• Previously, YC teams often had zero revenue upon selection, with a median monthly recurring revenue (MRR) of about $8,000 at graduation; now this median has surged to $20,000.
• There have even been extreme cases where revenue broke one million dollars (Seven Figures) within just three months of the batch, whereas traditionally, a similar scale would typically require over 18 months.
• Proportion of Solo Founders Soars: The proportion of solo founders in selected teams has skyrocketed from 5% to 18%-19%, setting a new historical high.
2. Underlying Drivers of the Hard Tech (Atoms) Explosion
• Generative Code Eliminates Engineering Bottlenecks: Previously, hard tech startups faced stringent limitations such as supply chain issues and complex hardware-software integration, with top software engineers being a key scarce resource. AI-assisted coding and code generation have significantly lowered this barrier, allowing startups to advance full-stack hardware development without needing to hire large engineering teams.
• Cutting-edge Models Feed Back into Scientific Breakthroughs: AI not only changes software but also significantly accelerates fundamental scientific research, enabling deep tech startups to achieve technology validation at earlier stages.
• Three Major Macro Driving Factors:
• Successful Examples of Space Commercialization: Breakthroughs by benchmark companies like SpaceX have driven investment in space infrastructure, leading to alternatives for sovereign satellite networks (such as Exosat) and orbital energy solutions (like Beyond Reach Labs developing satellite solar power for space data centers).
• Geopolitical and National Defense Security Demands: The new generation of founders is directly entering the defense sector, breaking through the traditional cost-plus consulting model of the military-industrial complex. For example, Icarus is developing solar-powered high-altitude reconnaissance/communication aircraft, and Nine Mothers is developing computer vision-driven short-range defense systems against drones, both quickly securing lucrative military contracts.
• Restructuring of Domestic Supply Chains and Dual-Use: Companies like Nox Metals are revitalizing old factories in Detroit to rebuild the domestic metal supply chain, with revenues growing at an ultra-high compound growth rate similar to software SaaS, primarily driven by the immediate procurement needs of emerging defense tech companies.
3. Evolution of Computing Power and Infrastructure: Chips, Optical Interconnects, and Power
• Supply-Demand Reversal in Computing Hardware: Even the rental cost of older model GPUs (like A100) has seen a rare upward trend, driving the entire industry chain to expand, from data center site construction and power battery hybrid solutions to underlying processors.
• New Architectures and Low-Precision Computing: The requirements for floating-point precision in large models continue to decrease (from FP32, FP16 down to FP8 and even lower), prompting teams exploring alternative solutions with ternary architectures or custom architectures (like Bot, Lamb Labs).
• Optical Interconnects Break Network Bottlenecks: The rapid surge in GPU computing power has made traditional electronic switches within data centers a communication throughput bottleneck. Companies like Dipole Labs are beginning to develop all-optical switching systems to achieve end-to-end all-optical communication.
4. Reconstruction of Software and SaaS: From Tools to End-to-End Agents
• Software is not disappearing but undergoing a qualitative transformation: The proportion of companies providing "end-to-end task completion" in the batch has increased from 10% to over 25%.
• From Point Recording to "Direct Work":
• Traditional SaaS is essentially point solutions or record systems that require human manual operation;
• The new generation of products directly replaces complete workflows (such as full-process insurance brokerage, clinical appointment pre-registration, medical billing settlement, etc.), with customers more willing to pay a premium for "final results".
• Example (Juicebox): Initially screening talent profiles through LLM, it has upgraded to having Agents automatically complete bulk outreach, schedule coordination, and preliminary interview arrangements, leading to a doubling or even several-fold increase in average revenue per user (ARPU).
• Data Silos and AI Harness Competition:
• Existing system record giants (like Salesforce, Slack) are evolving into AI harness frameworks. Companies with core workflow data and collaborative relationships can solidify their moat if they can make their software the main battlefield frequently called upon by Agents.
5. Invisible Money-Making Tracks: Laboratory Data and Reinforcement Learning (RL) Environments
• Quietly Profitable Niche Markets: In the past two years, YC has invested in dozens of startups selling high-quality training data or reinforcement learning (RL) simulation environments to leading AI laboratories, with several teams established only a few years ago achieving annual revenues exceeding tens of millions or even hundreds of millions of dollars.
• Capital Expenditure Flow: Leading large model laboratories are estimated to spend billions of dollars on RL environments and custom data construction, focusing on areas like financial deep simulation and long-horizon tasks.
• Extending to the Physical World: Laboratories need to train embodied intelligent models, driving rapid revenue generation for companies collecting industrial operational network data, first-person perspective (Egocentric), and tactile/operational data (like Practis Robotics, Deep Reach, Human Archive, etc.).
6. Robotics Technology Paradigm: Fine-tuning and Specialized Data are Key
• Complexity of Real-World Dimensions: Language models (LLMs) abstract the real world into language symbols, while robots need to handle continuous degrees of freedom in 3D physical space, requiring extremely high real-time responsiveness and physical interaction fault tolerance.
• General Base + Vertical Scene Fine-tuning: Startups accessing general robotic base models (like Physical Intelligence's PI model) generally cannot be used out of the box and must combine proprietary video and action data from vertical scenes for deep fine-tuning (for example, Boost Robotics focuses on cable plugging and unplugging in data centers, while Ultra combines thousands of hours of packing data to delve into packaging flows).
7. Evolution of Founder Portraits: The Dividend Period for Experienced Practitioners and Solo Teams
• Experienced Founders Welcome Spring: Entrepreneurs in their late 30s, 40s, or even over 50 have significantly expanded advantages.
• Business Acumen Surpasses Pure Coding: After Agentic Coding has greatly reduced software construction costs, knowing "what to build" and having a deep understanding of vertical business pain points and industry insights have become the most scarce core barriers.
• Agent Management Equals Team Management: Seasoned practitioners with rich engineering team or corporate management experience show stronger control when breaking down complex tasks and scheduling multiple concurrent coding Agents.
• The Threshold for Solo Entrepreneurship Has Been Greatly Lowered:
• In the past, running a company solo required extremely rare all-round talent (with top-notch technical, business, and sales abilities); now, with the help of AI, the threshold for a single person to prototype and validate commercialization has been significantly lowered.
• Dynamic Partner Structure: The modern model tends to have a solo founder start and gain early market traction, and then introduce full-time partners and teams based on business needs at a more mature stage.
ABAB AI Insight
YC 2026 Startup Panorama: AI is rewriting the underlying rules of entrepreneurship, capital, and industry.
If we only interpret the recent data released by Y Combinator as "AI startups are still hot," "robots are popular again," and "there are more solo founders," we actually underestimate this round of changes.
What truly deserves attention is not which track has suddenly become hot, but rather a larger structural change:
AI is simultaneously lowering the "cost of creating software" and the "cost of controlling the physical world."
This means that the entrepreneurial paradigm dominated by the internet, mobile internet, and SaaS over the past twenty years is undergoing a very deep migration:
Entrepreneurship is re-entering Atoms from Bits, moving from selling tools to selling results, from piling people to scheduling machines, from software companies to industrial companies.
In September 2026, YC disclosed in "The State of Startups in 2026" that over the past 12-18 months, the proportion of Hard Tech in YC projects has risen from about 8% to 20%; there has been significant growth in robotics, manufacturing, defense, semiconductors, photonics, and power infrastructure. At the same time, nearly one-fifth of new companies are now solo founders, and the speed at which companies generate revenue has also noticeably accelerated.
Behind these data is actually a new logic of capital.
────────────────
1. The entrepreneurial logic of the past twenty years is being truly rewritten for the first time.
To understand today, we must first understand the past.
From 2005 to 2020, the classic route for tech entrepreneurship in the United States was:
Software is eating the world.
The reason is very simple.
Software has extremely strong economic attributes:
The marginal cost of replication is nearly zero;
The cost of global distribution is very low;
Capital expenditure is low;
Gross margins can reach 70%, 80%, or even 90%;
Companies can achieve exponential growth without needing a lot of fixed assets.
Therefore, venture capital naturally favored SaaS, mobile apps, marketplaces, FinTech, and social networks.
This is also why the most important question in Silicon Valley over the past decade has often been:
"Is there any traditional industry that can be softwareized?"
Today, the question has changed.
The new question is becoming:
"Is there any real-world work that can be completely executed by AI + software + machines?"
These are two completely different questions.
The former adds a software interface on top of the real world.
The latter allows software to truly operate in the real world.
────────────────
2. The growth of Hard Tech from 8% to 20% is not as simple as "hardware becoming popular again."
The data disclosed by YC is very noteworthy:
Hard Tech has risen from about 8% to 20%;
Robotics has risen from about 1% to 6%-7%;
Domestic industrial manufacturing in the U.S. has risen from about 4% to 10%;
Defense has risen from about 1.5% to 5%;
Semiconductors and photonics are close to 4%;
Energy and power infrastructure are close to 3%.
Why?
Many people attribute this to "the return of American manufacturing."
This only explains part of it.
What is really happening is:
The costs that previously limited Hard Tech entrepreneurship are simultaneously decreasing.
────────────────
The first layer: decreasing software costs.
In the past, making robots required not just mechanical knowledge.
You also had to do:
Control software;
Machine vision;
Firmware;
Sensor fusion;
Motion planning;
Backend systems;
Data processing;
Simulation;
Testing tools.
Therefore, a so-called "robotics company" might actually need to employ dozens of software engineers first.
Today, coding agents are changing this structure.
A team of ten or even five people can accomplish the software workload that previously required dozens.
So the biggest short-term contribution of AI to robotics may not necessarily be the "robot brain."
Instead, it might be:
Cutting down the software engineering costs required to manufacture robots.
This is a point that is very easy to overlook.
────────────────
3. One of the greatest industrial values of AI may not be replacing programmers, but rather freeing engineers to create in the real world.
Many people discuss AI coding and only ask one question:
"Will programmers lose their jobs?"
This question is too narrow.
What capital truly cares about is:
When the production of software suddenly becomes ten times cheaper, what will humans do with this released engineering capability?
The answer may be:
Build robots;
Build drones;
Build satellites;
Build factories;
Build power grids;
Build new chips;
Build energy systems;
Build industrial equipment.
In other words:
AI may not necessarily make the world more virtual; rather, it may bring the tech industry back to the physical world.
This is a very counterintuitive aspect of this round of the AI revolution.
The internet era shifted a lot of capital from Atoms to Bits.
The AI era may see a reverse process:
Bits begin to help humans regain control over Atoms.
────────────────
4. Why is America suddenly willing to invest in "factories" again?
There is actually a second main line behind this:
National security.
The globalization logic of the past forty years is essentially:
Whoever has the lowest production cost produces there.
So:
Semiconductors went to Asia;
Electronics manufacturing went to China;
Textiles went to Southeast Asia;
Raw material processing became globalized;
Supply chains pursued Just-in-Time.
This is a typical maximization of economic efficiency.
However, entering the 2020s, the U.S. began to recalculate another cost:
What happens if war, pandemics, sanctions, or geopolitical conflicts occur and the supply chain is disrupted?
Thus, the optimization goals of businesses and governments shifted from:
Efficiency
to gradually focusing on:
Resilience.
Efficiency → Resilience.
This changes the value of many assets.
In the past, domestic factories in the U.S. might have meant:
High costs.
Today, they may mean:
Supply chain security.
This is why defense, industrial manufacturing, chips, energy, drones, and even mineral resources have become industries that venture capital can invest in again.
────────────────
5. SpaceX has changed not just aerospace, but the entire investment philosophy of Hard Tech.
To understand this round of Hard Tech, one must understand SpaceX.
One of the greatest historical significances of SpaceX is not just that it built rockets.
But it proved one thing:
Industries traditionally thought to require government, large contractors, and decades of R&D cycles can also be redesigned by startups.
In the past, the U.S. military-industrial system relied heavily on Cost-Plus Contracts:
You spend how much;
The government gives a certain profit on top of the costs.
This model naturally leads to a problem:
Insufficient incentives to reduce costs.
Silicon Valley's model is the opposite:
First, bear the R&D risks yourself;
Produce the product;
Scale it up;
Reduce unit costs;
Then sell in large quantities.
SpaceX brought this business logic into aerospace.
Anduril brought a similar logic into defense.
As a result, many young entrepreneurs are starting to ask:
Why can't radar be a startup?
Why can't drones be a startup?
Why can't missiles be a startup?
Why can't satellites be a startup?
Why can't nuclear energy be a startup?
Why can't shipyards be a startup?
This is the underlying background of the cultural shift in American Hard Tech entrepreneurship.
────────────────
6. The real bottleneck of AI is continuing to spread from "chips" outward.
In the past three years, when capital markets talk about AI, the easiest thing to think of is NVIDIA.
But the industry chain has begun to gradually spread outward.
The true system of AI computing power can be roughly understood as:
Chips → Memory → Network → Data Centers → Power → Land → Cooling → Power Grid.
Any link that is insufficient can become a bottleneck.
So today, AI is no longer just about the semiconductor industry.
It is actually consuming:
Energy;
Utilities;
Industrial real estate;
Optical communication;
Power electronics;
Nuclear energy;
Natural gas power generation;
Energy storage;
Transformers;
Transmission infrastructure.
YC also specifically mentioned that computing power is increasingly becoming a physical infrastructure issue.
This is very important for investors.
────────────────
7. Truly smart investments do not necessarily always invest in "the sexiest things."
Historically, there is a very classic investment rule during gold rushes:
When everyone goes to mine for gold,
The people selling shovels often make money more steadily.
AI is the same.
The first layer is models:
OpenAI, Anthropic, Google DeepMind.
The second layer is chips:
NVIDIA, AMD, ASIC.
The third layer is infrastructure:
Networks;
Optical communication;
Data centers;
Power;
Cooling;
Storage.
The fourth layer is applications.
So a very important investment mindset is:
Don't just ask who has the smartest model.
Also ask:
If AI usage expands 100 times in the future, what will be in serious shortage?
This question is often more valuable than "Who will be the next ChatGPT?"
────────────────
8. Why might optical communication become increasingly important?
Today, many people understand data centers as:
Many GPUs placed together.
This is actually inaccurate.
The essence of a truly large AI system is a "supercomputer" made up of thousands of computing chips.
GPUs must exchange data frantically with each other.
Thus, another issue arises:
Communication speed.
Even if a single GPU is fast, if the data transfer between chips is too slow, the entire cluster still cannot run.
This is why:
High-speed networks;
Optical modules;
Silicon photonics;
Photon switching;
CPO;
Data center interconnect
will become increasingly important.
The future war for AI infrastructure is not just:
Whose GPU FLOPS are higher.
But:
Who can make hundreds of thousands or even millions of accelerators work together.
This has escalated from a "chip war" to a:
System engineering war.
────────────────
9. The biggest change in AI software: SaaS is shifting from "selling tools" to "selling labor."
This is what I believe to be the most important change in the entire YC content.
In the past, SaaS companies sold:
Tools.
Salesforce provided tools for salespeople.
Adobe provided tools for designers.
QuickBooks provided tools for accountants.
ServiceNow provided tools for IT departments.
Software improved human productivity, but:
Ultimately, it was still humans who completed the work.
Agents change this logic.
In the future, more and more software will directly tell customers:
You don't need to operate me.
Just hand the tasks over to me.
For example, traditional recruiting SaaS:
Search for candidates;
Screen;
Send emails;
Schedule meetings;
Record CRM.
In the future, AI Recruiting Agents may directly:
- Find candidates;
- Analyze candidates;
- Send emails;
- Follow up;
- Schedule interviews;
- Update systems.
What customers are buying is no longer:
Software Seat.
But rather:
Completed recruitment.
────────────────
10. This may change the pricing model of the entire software industry
The most classic business model of SaaS in the past was:
Per Seat.
For example:
$50 per employee per month.
100 employees:
$5,000 MRR.
But what if AI leads to a decrease in the number of employees?
The number of seats would actually decrease.
Thus, the next generation of AI companies will increasingly tend to:
Charge by results;
Charge by tasks;
Charge by transactions;
Charge by cost savings;
Charge by revenue generated.
This contains a huge change in the business model:
The TAM of software companies begins to shift from "IT Budget" to "Labor Budget."
This distinction is very significant.
American companies spend a lot of money on software every year.
But one of the biggest costs for companies has always been:
Salaries.
If AI can truly take on work, then the market size that AI companies can theoretically enter may far exceed that of traditional SaaS.
────────────────
11. This is also why Agent companies may grow very quickly
YC stated that many companies entering YC previously had almost no revenue, and by the end of the batch, the median MRR was about $8,000; recently this figure has reached about $20,000, and there have been cases of teams going from zero to seven-figure revenue within a three-month batch.
This does not mean that all AI companies can grow like this.
But it reflects an important phenomenon:
The distance from AI products being a Prototype to a Commercial Product is shrinking.
In the past:
Writing software took half a year;
Hiring engineers;
Testing;
Deploying;
Selling.
Today:
One domain expert + AI Coding Agent,
can produce something sellable in a few weeks.
Thus, the core bottleneck for startups has shifted.
Previously:
Can you build it?
Today it is increasingly becoming:
Do people actually want it?
In other words:
After the technical risk decreases,
market judgment becomes more important.
────────────────
12. "Knowing how to code" is shifting from a barrier to a basic skill
This is something many tech entrepreneurs must realize.
In the past, if a person could:
Write front-end;
Write back-end;
Deploy servers;
Design databases;
Make APIs;
they might be considered very rare entrepreneurial talent.
These skills are still important today.
But AI is rapidly commodifying them.
Thus, value is beginning to shift to several other abilities:
Knowing what is worth doing;
Understanding the real pain points of an industry;
Having customers;
Having data;
Having distribution channels;
Having a brand;
Having regulatory licenses;
Having physical infrastructure.
This is what is referred to as:
Shifting from How to Build to What to Build.
YC itself has also emphasized in this discussion that knowing what to build is becoming more important than simply knowing how to build.
────────────────
13. Why might entrepreneurs in their 40s and 50s enter a golden age?
In the past, Silicon Valley had a very strong myth of youth:
In their 20s;
Dropping out of school;
Knowing how to program;
Living in a garage;
Creating the next Facebook.
Today, the situation may be changing.
Why?
Because AI is reducing the scarcity of "implementation capability."
And entrepreneurs in their 40s may have another asset:
20 years of industry knowledge.
For example, someone who has worked in insurance for 20 years knows:
Which steps waste the most time;
Which claims are most prone to errors;
What customers are truly willing to pay for;
Where regulations are most troublesome;
Which data is most critical.
In the past, they might not have known how to program, so they couldn't start a business.
Today, AI can help them quickly complete:
Prototypes;
Back-end;
Databases;
Automation;
CRM;
Data analysis.
Thus, their 20 years of experience is suddenly "software-ized."
This represents a very important new entrepreneurial paradigm:
Domain Expert + AI.
────────────────
14. The most dangerous people in the future may not be young programmers, but "industry veterans who understand both the industry and AI"
Take a simple example.
A 22-year-old engineer starts a business in the insurance industry.
He needs two to three years to understand the industry.
Another person has already worked in an insurance company for 15 years.
In the past, the second person wouldn't know how to write software, so the first person had the advantage.
Now, the second person uses an AI Coding Agent.
Then:
His industry experience + customer relationships + business understanding
may directly outweigh pure technical advantages.
So one of the most noteworthy people in the future entrepreneurial market will be:
AI-native industry insiders.
That is:
Understanding the industry;
Understanding customers;
Understanding business;
And then using AI to productize their experience.
────────────────
15. What does the rise of Solo Founders really mean?
YC's latest discussion states that the proportion of single-founder companies has approached 18%-19%, while historical levels are much lower. Paul Graham previously mentioned that the percentage of single-founder companies in YC's summer batch rose from about 9% the previous year to about 18%. YC's official stance still reminds applicants: Solo Founders can apply, but generally having co-founders still makes it easier to succeed.
So do not misunderstand:
Solo Founder ≠ Always only one employee.
What has really changed is:
When a company must hire a second person.
In the past:
Idea → Cofounder → Engineer → Designer → Product → Revenue.
In the future, it may be:
Idea → Founder + AI → Product → Revenue → Team.
In other words:
Team building occurs after product validation.
This has a huge impact on entrepreneurial economics.
────────────────
16. The equity structure of entrepreneurs may change as a result
Traditional startups:
Two founders.
50% + 50%.
Or:
60% + 40%.
Without a product, a large amount of equity has already been distributed.
In the future, a Founder may:
Achieve MVP on their own;
Acquire customers;
Reach $20K MRR;
And then find partners.
At this point, their negotiation position is completely different.
In the past:
"We are betting on an idea together."
In the future:
"This company is already running, you come join."
This will actually improve the capital efficiency of the Founder.
────────────────
17. But do not misinterpret it as "the era of one-person companies has arrived"
We must remain calm here.
AI can replace many execution tasks.
But companies still have a lot of things that are very difficult for AI to replace:
Sales;
Financing;
Recruiting;
Organizational management;
Negotiation;
Customer relationships;
Strategic partnerships;
Brand;
Leadership;
High-risk decision-making.
Broader data from Carta also shows that while the proportion of Solo Founders has indeed been rising over the long term, the proportion of Solo Founders receiving VC funding is significantly lower than their company creation share. By 2025, the proportion of single founders in Carta companies has reached about 36%, but the proportion receiving VC funding remains significantly lower.
So the correct understanding is not:
"Teams are useless."
But rather:
The minimum viable team is shrinking.
This is a completely different conclusion.
────────────────
18. AI data may be repeating the gold rush of 1849
Another extremely important industry that ordinary entrepreneurs rarely pay attention to is:
Training data and RL Environment.
YC mentioned in the program that it has invested in more than a dozen companies over the past few years that provide data or reinforcement learning environments for large AI Labs, some of which have annual revenues exceeding ten million dollars.
Why?
Because what AI Labs are truly lacking today is not just GPUs.
They also lack:
High-quality training tasks;
Expert data;
Real workflows;
Long-horizon Tasks;
Verifiable results;
Simulated environments;
Robot interaction data.
────────────────
19. The competition for the next generation of AI may shift from "internet text" to "real-world experience"
What is the core data for the first generation of LLMs?
The internet.
Wikipedia;
Reddit;
Web pages;
Books;
GitHub;
Papers.
The problem is:
This data is quickly being consumed.
What should future models do to continue improving?
They must acquire new data.
Thus, data is beginning to develop in three directions:
The first type: Expert data
Doctors;
Lawyers;
Investment bankers;
Accountants;
Engineers;
Scientists.
The second type: Interaction data
AI performing tasks;
Making mistakes;
Retrying;
Learning.
The third type: Physical world data
Robots grasping things;
Walking;
Plugging in;
Packing;
Welding;
Operating machines.
This means that an important asset in the next decade may be:
The ability to collect data from the real world.
────────────────
20. What robots truly lack is not "a bigger ChatGPT"
This is also a mistake that many robot entrepreneurs easily make.
The input for LLMs is basically:
Tokens.
Robots face:
3D space;
Forces;
Speed;
Friction;
Position;
Vision;
Touch;
Time.
The real world does not have Ctrl+Z.
If a language model says something wrong, it can answer again.
If a robot grabs the wrong thing:
It may damage equipment;
Injure people;
Cause production line shutdowns.
So robot models will heavily rely on:
Specialized scene data.
This is also why YC emphasizes in discussions that general robot models may still require a lot of fine-tuning for specific tasks.
────────────────
21. This means that the robotics industry may not ultimately have just one "OpenAI"
The structure of the robotics industry is likely to be more decentralized than that of LLMs.
Because:
Warehouse robots need a set of data;
Surgical robots need a set of data;
Factory robots need a set of data;
Agricultural robots need a set of data;
Household robots need another set of data.
Therefore, it is very likely that in the future we will see:
General Foundation Models
Industry Data
Dedicated Hardware
Vertical Software
Forming a large number of vertical robotics companies.
This is very similar to the SaaS era.
AWS is the infrastructure.
But on top of that, we still saw the emergence of:
Shopify;
Salesforce;
Uber;
Airbnb;
Stripe.
Robots may be the same.
────────────────
22. The real big opportunity is not "AI + X", but AI changing the cost curve of X.
In the future, when judging a startup project, it’s best not to ask:
"Is there AI involved?"
This is a very low-level question.
You should ask:
Has AI changed the unit economic model of this industry?
For example:
Previously, an insurance claims adjuster handled 20 cases a day.
After AI:
One person handles 200 cases.
This is called:
10× Productivity.
Or:
Previously, industrial inspection required engineers.
Now cameras + AI complete it automatically.
This is called:
Labor Replacement.
Or:
In the past, software development required $50M.
Now it costs $5M.
This is called:
Cost Curve Compression.
Only when these changes occur does AI truly create commercial value.
────────────────
23. One of the most important capital logics in the next decade: upgrading from "software leverage" to "intelligence leverage."
In the past, the most important leverage for internet companies was:
Software Leverage.
A programmer writes code once;
One hundred million people use it.
In the future:
AI adds a second layer of leverage:
Intelligence Leverage.
One person's judgment;
Through AI;
Can simultaneously manage:
10 Agents;
100 Agents;
In the future, even thousands of Agents.
So the basic organizational unit of enterprises may change.
In the past:
One Manager managed 10 people.
In the future:
One Founder may manage:
Dozens of AI Agents + a small number of core employees.
The scale of the enterprise and the value of the enterprise will begin to decouple.
────────────────
24. The first truly "micro giant enterprises" may emerge in the future.
In the past:
A $10B company usually needed:
Thousands or even tens of thousands of employees.
For the first time in the internet era:
Hundreds of people created tens of billions of dollars in value.
In the AI era, it may further emerge:
Dozens of people creating tens of billions of dollars in value.
Even:
A very small number of people controlling huge cash flows.
So in the future, an important indicator in the entrepreneurial field may no longer be:
The number of employees.
But rather:
Revenue per Employee.
That is:
How much revenue each employee generates.
This is one of the most worthwhile long-term observables in the AI era.
────────────────
25. However, capital ultimately recognizes only one thing: the moat.
AI makes entrepreneurship easier, but it also brings a side effect:
Copying becomes easier.
You can do it in three days.
Others might also be able to do it in three days.
So in the future, the moat of entrepreneurial companies cannot just be:
"I used GPT."
The real Moat will come from:
Data;
Distribution;
Network Effects;
Brand;
Regulation;
Customer Relationships;
Hardware;
Supply Chain;
Economies of Scale;
Capital;
Patents;
Real-world infrastructure.
A very simple judgment:
If OpenAI launches the same features tomorrow, what will be left of your company?
If the answer is:
"Nothing at all."
Then this company is in danger.
────────────────
26. What truly deserves entrepreneurs' attention is the rules behind this round of capital migration.
Summarizing the changes in YC 2026, I believe it can be condensed into six points.
First, technology is being commoditized, and judgment is appreciating.
The importance of knowing how to write code will not disappear.
But knowing what to write will become increasingly valuable.
────────────────
Second, software is starting to move from auxiliary labor to execution labor.
In the past, AI helped you work.
In the future, AI will complete the work for you.
This is the biggest commercial change of Agents.
────────────────
Third, entrepreneurial teams are getting smaller.
But the markets that companies can attack are getting larger.
Small teams + large markets will become a very typical combination in the AI era.
────────────────
Fourth, capital is re-entering Atoms from pure Bits.
Because AI not only changes software.
It is beginning to change:
Manufacturing;
Energy;
Robotics;
Defense;
Data Centers;
Aerospace;
Fundamental Science.
────────────────
Fifth, the endgame of AI is not an App, but a new industrial infrastructure.
Looking back at today in the future, AI will likely be similar to electricity and the internet.
It will not just be an industry.
It will become:
A factor of production for all industries.
────────────────
Sixth, the biggest entrepreneurial opportunities often arise when old industrial structures are recalculated.
Why is it worth paying attention to insurance now?
Why is it worth paying attention to healthcare?
Why is it worth paying attention to manufacturing?
Why is it worth paying attention to defense?
Because:
AI has recalculated labor costs;
Geopolitics has recalculated the value of supply chains;
Computing power demand has recalculated the value of electricity;
Robots have recalculated physical labor.
Whenever a key cost undergoes an order of magnitude change,
New giants may emerge.
────────────────
The most memorable sentence:
The core of entrepreneurial opportunities in the past twenty years has been:
Digitizing the real world.
The entrepreneurial opportunities in the next decade may turn into:
Letting digital intelligence regain control of the real world.
The internet moved the world's information online.
What AI is about to do next may be even greater:
Understanding this information;
Making decisions;
Executing work;
Controlling machines;
Ultimately changing the real world.
Therefore, the Hard Tech resurgence, the explosion of Agents, the growth of Solo Founders, the return of seasoned entrepreneurs, and the warming of robotics and energy infrastructure that YC sees today may seem like many independent trends.
In fact, they are all driven by the same historical change:
The cost of intelligence is rapidly decreasing.
When computing costs fell, the internet was born.
When communication costs fell, mobile internet was born.
And when "intelligence costs" begin to fall,
The next batch of giant enterprises will likely not just create another software.
They will redefine:
How people work, how companies are organized, how machines operate, and how capital is allocated.
This is the real insight worth understanding from this set of YC data.
G