Author: Luo Xia Leo, Luo Xia in AI
From 2026 to now, YC has invested in three batches: Winter 199 companies, Spring 196 companies, Summer 235 companies, a total of 630 companies. At the same time, it updated the Request for Startups twice.
The list is being shared by everyone, but it serves as a recruitment advertisement, detailing what YC wants to receive. The real answers are in the 630 companies it has already funded.
I reviewed the public descriptions, industry classifications, and founder profiles from these three batches to answer four specific questions: What directions is YC investing in, how crowded are these directions, who are these companies selling to, and what do the selected individuals look like.
Here's the conclusion: Over 90% of the money YC has invested this year does not target To C, and more than half is focused on repairing pipes for AI.
Let’s first outline the directions, then look at the data, and finally present a few points that I believe are genuinely useful for entrepreneurs.
01 The 10 Directions It Invests Real Money In
Arranged from most to least companies, after merging similar items these are the ten directions. Following each is an approximate quantity from these three batches and a specific example.
01 Agent Infrastructure: The layer that provides ID cards for AI employees
There are 55 in the Spring batch, and a total of 89 in the S26 batch with agent and AI infrastructure combined; this is the most densely populated segment this year. They are all working on very specific dirty work: Inkbox provides agents with email, phone numbers, and 2FA, enabling them to be recognized by the system like employees; Agentcard handles agent payments; Glen manages memory; Hyperprobe performs runtime debugging; Archal verifies what the agent has done and automatically opens repair PRs.
02 AI Employees: Not making tools, directly taking over a position
There are 56 in Spring and 52 in S26. Rational describes itself as an "AI employee" of an accounting firm, with Last Accounting Company and Billow AI in the same segment; Osmaura and Erinys are in legal; Locke handles government affairs; Truffle oversees the operations of restaurant kitchens. The commonality is that the product pages list job titles, not function names.
03 Inference Costs: Turning billing itself into a business
At least four companies in S26 are handling different parts of this chain. Understudy Labs captures production flow, fine-tunes cheaper open-source models from it, and then uses automated quality gates to decide which requests can be routed; Conifer handles inference routing; Stoa and OpenRelay are on the GPU side.
04 Robots and Embodied Intelligence: Prices have been lowered
About 23 companies in S26. Nori offers a 1,688 USD dual-arm home robot, where the selling point is the price rather than performance, and it allows users to train skills at home and share them; Salem Robotics does inspections of nuclear facilities; Libra Robotics handles photovoltaic installations; Manifold and Proprio do warehouse handling. Companies like Robocurve, Instance, and Markov have also emerged, focusing on evaluations for robots.
05 Operating Systems for Factories, Logistics, and Supply Chains: The largest group in S26
About 41 companies, the largest single cluster in this batch. Pango creates an agentic OS for e-commerce logistics; Control Seat adds a layer of intelligence to industrial SCADA systems; Alloovium interprets construction documentation; Torus serves as an engineering assistant for physical infrastructure projects.
06 Defense and Space: Buyers have taken the initiative to state their needs
Among the 26 hardware companies in Spring, 13 focus on defense; in S26, there are groups dealing with precision munitions, drone swarms, anti-drone technologies, and battlefield situation analysis (ISENGARD, Earendil Robotics, Guild, Hop Aero). The "Future of American Defense" line in RFS is signed by U.S. Army Secretary Daniel P. Driscoll, not a YC partner.
07 Data Centers and Energy: Computing power is starting to look for a place
Atomarine builds floating data centers at sea, using seawater for cooling; Pacific creates modular data centers that can be quickly deployed to unfinished infrastructure sites. YC itself wrote a line in RFS about "offshore computing power," citing the blockage of land-based electricity and approvals.
08 Healthcare: From selling software to holding licenses themselves
Approximately 20 to 24 companies in S26. Taiga handles AI medical billing for independent clinics, Care GP operates a suite for Australian general practice clinics, and RonanRX has become a full-stack pharmaceutical manufacturer—delivering personalized peptides and GLP-1 directly to individuals. Notably, many companies are no longer selling software to licensed institutions but instead obtaining the licenses themselves.
09 Finance and Insurance: Underwriting, Settlement, Reconciliation
Between 17 to 26 companies in S26 (with slight differences in criteria). Klaimee provides agentic insurance, Vestris handles mortgage settlement processes, and Spectre Intelligence focuses on intelligence and risk management. The proportion in this segment is declining, which will be discussed later.
10 The Internet for Agents: Reworking the web
Today's web pages, APIs, and documentation are for human viewing. Context.dev and Rindler are converting them into a format usable by agents; Harsha Gaddipati's "self-maintaining API" in RFS represents another aspect of this same topic—automatically updating client codebases upon interface changes.
Five of the ten directions will ultimately need to extend beyond the screen.
02 Over 90% of the 630 Companies Do Not Intend to Do To C
First, let’s look at the industry distribution for the Summer 2026 batch based on YC’s own categorization:
Figure 1 Industry Distribution of YC Summer 2026 Batch (n=235)
B2B ██████████████████████52.3%
Industrial ██████████23.0%
Healthcare ████8.5%
Fintech ███6.8%
Consumer ██5.5%
Real Estate Construction █2.6%
Breaking down the industrial, healthcare, finance, and construction sectors, their clients are similarly businesses and organizations. Therefore, a more accurate interpretation is: over 90% of this batch targets businesses and institutions, with only 5.5% truly aimed at individual consumers, and education has only one company.
This has two direct implications for entrepreneurs.
One implication is the customer acquisition path. Over 90% of the companies will follow the same route—finding a specific position, calculating its annual expenditure, and then comparing it against this cost. This path is crowded, but it has defined payers.
Over 90% of companies are competing for the same set of payers; in the remaining segment of under 10%, almost no one is interested.
The other implication is its converse. There are only a dozen companies in the consumer space, just one in education, while YC simultaneously included "AI consumer products for 1 billion people" and child AI tutoring in RFS. There is a disconnect between what is desired and what has been invested; this gap will be addressed separately.
03 In Half a Year, Money Has Shifted from "Creating Applications" to "Repairing Pipes"
If we only look at the four percentages, this year's changes are summed up. This illustrates the shift between the Spring and Summer batches:
Figure 2 Four Shifts in Six Months (Spring 2026 → Summer 2026)
AI Infrastructure 8% ↑ 20%█████████
AI Application Products 55% ↓ 39%██████████████████
Industrial 12.8% ↑ 23.0%██████████
Fintech 10.2% ↓ 6.8%███
The number of companies making AI applications dropped by 30%, while those handling AI foundational infrastructure increased by one and a half times. This isn't a sudden shift in YC's preferences but reflects the normal order after a category has scaled: first there are many applications, and then someone extracts the dirty work of redoing all the applications for sale.
The same thing happened when SaaS gained traction. The first group to profit weren't SaaS providers but Heroku, Twilio, Stripe.
When a category's billing alone warrants the creation of a company to optimize it, it indicates that it has already reached a scale.
The doubling of the industrial segment is the other half of the story. More intuitively: for every four companies YC invests in, one will ultimately have its value tied to a piece of equipment, a production line, or a warehouse, while a year ago this ratio was closer to one out of eight.
04 The Selected Individuals: Two People, 5.8 Years of Experience, Half Are Not Americans
This section of data is primarily compiled from the profiles of founders from the W26 and Spring batches, based on third-party statistics rather than official YC releases.
Team Size. Two co-founders are the absolute mainstream—among the 199 in W26, 129 are two-person teams, accounting for 64%; solo founders number 22, accounting for 11%; teams of three or more make up about a quarter. The number of solo founders receiving funding has noticeably increased compared to a few years ago, indicating that the old rule "there must be co-founders" is loosening.
Experience. The average years of work experience of founders before joining YC is 5.8 years; those in the agent direction tend to be younger, with a median of 4.8 years. They are not student entrepreneurs, nor are they veterans; they are in the age group that has "worked a round in a big company and realized how bad certain specific processes are."
Schools and Previous Employers. Among the W26 founders, there are 30 from Berkeley, 22 from Stanford, and 18 from Harvard, collectively accounting for about 16%. Amazon appears in 14 of the companies, followed by Apple in 12; Meta and McKinsey have significant representation in the agent sector.
Nationality. In the Spring batch, founders with international backgrounds make up half, with the top six sources in the public statistics being:
Figure 3 YC Spring 2026 International Founder Source Top 6 (by number)
United Kingdom ██████████████████████33 People
India ███████████████████29 People
France ██████████████21 People
Canada ███████████17 People
Germany █████████14 People
Switzerland ███████10 People
Two points are worth noting for Chinese readers. First, China is absent from this top-six list. Second, another set of data shows that among U.S.-based YC founders, 29% are of Indian descent, up from 7% in 2008 and only about 15% in 2019.
I did not find authoritative statistics on the proportion of Chinese founders, so I won't reach a conclusion. However, it can be confirmed that 91% of international founders ultimately established their company headquarters in the U.S., with 76% of Spring batch companies headquartered in San Francisco.
YC is not a remote project. It is a ticket that requires you to move.
If you are in China, with a team in China and customers in China, the cost-effectiveness of the YC path needs to be recalculated—not because you can't get in, but because a significant portion of the value of the resource network after you enter cannot be realized.
05 YC Gives You 500,000 USD, Takes 7%, Then What
The terms haven't changed much over the years, but many people only remember half of them.
Figure 4 The Two Amounts in YC's Standard Agreement
█████125,000 USD for 7% equity, post-money SAFE, immediately priced
███████████████375,000 USD no valuation cap MFN SAFE, converting according to your next round's terms
Total 500,000 USD. What really affects dilution is the first amount: 7% is fixed, regardless of your previous valuation. The second amount has no cap; the more expensive your next round is, the smaller the percentage it gets converted into at the time of conversion—thus it is essentially betting that you can raise money at a good price after graduation.
So, what price is that after graduation? Following investors from these two batches, the benchmark given is 4 million USD funding, pre-money valuation of 40 million USD, and having 1 million USD ARR before pitching is now not uncommon. This is an estimate from the investor side, not publicly reported data from YC, but the direction is clear:
Choosing the right direction is no longer an advantage; having chosen the right direction and already generated income constitutes an advantage.
Putting these two numbers together, YC's business model is quite straightforward: using 500,000 USD to buy 7% plus a subsequent conversion right, investing in over six hundred companies, and just a few companies reaching billions of dollars can cover everything. This machine allows for a high error rate—but you cannot afford to make mistakes.
06 Five Useful Points from These Numbers
First, don’t be the 90th agent. Instead, work on the layer that those 89 companies need to redo. The criteria are simple: if every agent company has to build their own from scratch and they are all reluctant to do so, this is worth extracting and selling. Identity, payments, memory, evaluation, inference costs, all emerged this way.
Second, write job names on your product pages, not function names. These two batches of companies collectively completed a shift in wording: no longer saying "helping teams become more efficient," but rather "replacing this position." The pricing method has also changed—charging by seats assumes your employees are using my software; once the narrative shifts to replacement, pricing can only be based on workload, results, or directly benchmarked to the annual salary of that position.
Third, those directions that YC repeatedly mentions but where the records are still empty are the true areas they are recruiting for. The consumer sector has only 5.5%, education has only one company, and cryptocurrency is nearly absent in the batch structure, while those three segments are specifically highlighted in RFS. RFS was originally intended to fill gaps in recruitment ads—it indicates "what is missing from the applications we received."
First, clarify how much the position you want to replace costs the customer annually. Everything else is just packaging.
Fourth, if you are dealing with the physical world, first confirm that your strength is not in the model. The threshold for the industrial segment lies in whether you can access that batch of data, enter that site, and pass through that procurement process. If your answer is "our model is well-tuned," then you are likely not in a competitive position within this track.
Fifth, the defense sector requires a separate assessment for Chinese entrepreneurs. It has seen the fastest growth this year, but the benefits are heavily tied to the U.S. procurement system. The direction can be referenced, but the path cannot be simply copied.
07 Three Cold Waters: Lists Lagging, Tracks Crowded, YC Can Be Wrong
Lists can lag. The Summer RFS once mentioned "Software for Agents," and by the time this was publicly written, there were already 89 companies in the S26 batch. There is typically a gap of about one to two batches between the appearance of RFS and the large-scale deployment of corresponding companies. The moment you read the list, the first wave is already present.
The track is already crowded. In the ten directions above, the first two combined account for over one-third of this year's new companies. You are not the only one seeing the opportunity.
YC can also be wrong. It invests in over six hundred companies a year; it is a probability machine, not a prophet. The batch structure reflects what applications it received and its current preferences, which do not equate to market conclusions. Offshore data centers, 1,688 USD home humanoid robots, companies issuing ID cards to agents—all may be gone in three years, which is completely acceptable within YC’s model.
YC can be wrong 620 times; you can only afford to be wrong once.
There is also a matter that this batch of companies collectively has not addressed: when an agent truly replaces a position, who is responsible for its mistakes. Compliance, auditing, and insurance layers appeared in RFS but remain very thin in the batch. This is both a risk and a gap mentioned earlier.
08 One Sentence Summarizing These 630 Companies
What they truly indicate is not which direction will win, but that the boundaries of the word "software" are expanding: part has shifted to the invisible pipes beneath agents, while another part has moved to devices, production lines, and warehouses outside the screen.
The entries on RFS will change with the next issue, but this movement will not.
If there is one thing to take away from this article: first see what YC has stated, then see what it has invested in, and only choose the areas where the two are inconsistent.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
