Y Combinator（YC）has recently released its Summer 2026 Requests for Startups（RFS）, outlining the directions this season especially wants founders to pursue. From the directions YC is asking about this season, the startup opportunity is no longer just “adding a chatbot into an existing product,” and it’s not about building another AI mini-tool that helps people write emails, organize meetings, or generate slide decks.
YC is instead focused on how AI is transforming more complex systems inside companies: internal knowledge, professional services delivery, semiconductor supply chains, hardware manufacturing, agriculture, healthcare, space electronics, and counter-drone defense.
In other words, AI startups are shifting from “improving individual productivity” to “rebuilding organizational and industry workflows.” If a company is AI-native from day one, what it sells might no longer be a software package, but rather a service that’s been restructured by AI, an enterprise operating system, or even a new capability in the supply chain.
From leading indicators to lagging indicators—does YC agree?
Five Crowns Capital partner Meng Xing recently published a Silicon Valley inspection report, pointing out a key turning point forming in the current AI startup ecosystem: after the speed of AI iteration has accelerated sharply, Y Combinator (YC), which used to be seen as a startup barometer, may be gradually shifting from a “leading indicator” to a “lagging indicator.”
(Use AI to increase output or reduce costs? A hundredfold efficiency didn’t translate into a hundredfold revenue boost—but no one in Silicon Valley dares to call it off)
When Meng Xing sat in the audience at the YC W26 batch Demo Day in March this year, he put his pen down after the fifth company’s pitch. The reason wasn’t that these companies weren’t trying hard enough—it was that the topics were too similar. Among the more than 100 companies in this batch, about 80% are building vertical agents, such as helping lawyers organize documents, helping customer service route tickets, and helping HR screen resumes.
If you put these same ideas in October last year, they might still have made investors think “That’s an interesting concept.” But by 2026, with Claude Code moving from a developer tool to an interface that nearly everyone can use, and Opus 4.6 further lowering the bar for vibe coding to the floor, many vertical agents that haven’t yet built business moats are no longer as scarce as they were before. A typical engineer could even replicate a similar product with one weekend’s effort.
That also puts pressure on YC’s batch system, which has helped it succeed. From applications, screening, admission to the batch, refinement, and finally the Demo Day pitching, YC’s rhythm is designed for a world where product and market changes are relatively stable and slow-moving. But at today’s pace of AI iteration, five months is enough to trigger multiple shifts in the paradigm. When model capabilities, development tools, and user habits are all rewriting themselves quickly, once a startup idea enters the batch process, by the time it reaches the public demo, it may have already gone from the frontier to consensus—or even to a red ocean.
Interestingly, YC’s latest season of Requests for Startups (RFS) is also trying to respond to this change. In the Summer 2026 RFS, YC clearly states that AI has stopped being just a single feature, and has become the foundation of software, services, hardware, and even the physical world.
AI is no longer just a feature—it’s the underlying assumption for companies and industries
In the Summer 2026 RFS, YC says that AI has stopped being merely a feature within products, and has started becoming the new foundation for software, services, silicon chips, and the physical world. This sentence really captures the theme of this list: AI startups can’t just stop at “plugging a model into a product.” They must rethink the original job, and whether the company and industry should be rewritten.
So, the topics in this RFS are rarely about a simple consumer-app-style vision of entrepreneurship. More precisely, YC isn’t searching for the next AI application that blows up based on UI and traffic. Instead, it’s focusing on areas that have historically been difficult for software to fully transform.
For example, professional services used to depend heavily on human labor and process experience; corporate knowledge is scattered across Slack, email, tickets, and meetings; semiconductor supply chains still rely heavily on manual coordination; hardware manufacturing in the U.S. iterates far more slowly than in Shenzhen; agriculture is still tied to spraying pesticides over large areas; and counter-drone defense faces a structural problem where the attacker’s costs are far lower than the defender’s.
These aren’t problems that can be solved by simply adding a chatbot. They require AI to be embedded into workflows, data, hardware, supply chains, and decision systems—becoming part of the entire way operations run.
From selling software to selling outcomes: AI-native service companies
In the RFS, YC partner Gustaf Alströmer proposes the direction of “AI-Native Service Companies,” which may be the key to understanding the turn of the next wave of AI entrepreneurship.
In the past few years, most AI startups have been building copilots—helping people complete work faster. They sell software, and users still need to operate tools, judge results, and deliver the work. But what YC is interested in now is the next step: companies no longer just sell tools—they sell services directly.
That means customers don’t need to buy an AI software package to train employees to use. Instead, they hand tasks like insurance brokerage, accounting, taxes, audits, compliance, and medical administration to an AI-native company to complete. In this setup, software becomes an internal productivity tool, not the main product sold to customers.
This shift matters a lot. The services market is much larger than the software market, and many professional services are already outsourced by enterprises. If AI-native companies can complete work with lower costs, faster speed, and more consistent quality, they won’t only be attacking the SaaS layer—they’ll be attacking the entire cost structure of the services industry.
Company brain: what corporate AI automation lacks isn’t models—it’s internal knowledge
Another key focus is “Company Brain.” YC partner Tom Blomfield believes the biggest bottleneck for enterprise AI automation is no longer model capability, but domain knowledge. Every company has a large amount of critical knowledge scattered in employees’ heads, old email threads, Slack discussions, customer service tickets, and databases. The reason companies can operate is that humans roughly know who to ask about something, which process has exceptions, and which decisions aren’t actually written in formal documents.
But AI agents can’t rely on this kind of fuzzy memory. To let AI truly enter enterprise automation, companies need a new foundational layer that organizes scattered knowledge into an updatable, executable map of how the company operates.
YC emphasizes that this isn’t a company search tool, and it isn’t a document chatbot. It’s a system that lets AI understand “how the company works.” For example: how refunds are handled, how pricing exceptions are decided, how engineering incidents are escalated, and how customer service responds differently depending on circumstances. Once these workflows are structured, AI can execute tasks safely and consistently.
This is also where AI entrepreneurship is starting to become harder—and more valuable. The future moat might not just be models or interfaces, but who can convert an organization’s or an industry’s implicit knowledge into workflows that AI can execute.
Enterprise AI operating system: turning companies from open loop to closed loop
Similar to Company Brain, YC partner Diana Hu proposes “The AI Operating System for Companies.” She observes that the best AI-native companies have already made the entire organization queryable: every meeting, every ticket, every customer interaction, and every product decision can be understood and used by an intelligence layer.
This changes a company from open loop to closed loop. Traditional enterprises often make decisions first, then check the results weeks later. But AI-native companies can continuously monitor what’s actually happening, compare it with the target state, and then adjust execution accordingly.
The problem is that doing this today requires a lot of integration work. Slack, Linear, GitHub, Notion, meeting recordings, customer service systems, and all kinds of internal tools must be connected. YC therefore believes the opportunity for startups lies in building a connection layer that automatically turns every work trace created inside a company into data that AI can understand, reason over, and execute. This isn’t another dashboard—it’s a foundational system that allows the company itself to form a self-improvement feedback loop.
The next-generation software might not be built for humans—it might be built for agents
YC also proposes “Software for Agents.” The core assumption behind this direction is: in the future, there will be a large number of AI agents on the internet that replace humans to do research, buy things, manage CRM, operate tools, and complete tasks. But today’s software is still designed for humans to click. Humans need forms, buttons, and dashboards; agents need API, MCP, CLI, machine-readable documents, and programmable registration, authorization, and usage flows.
So once everyone is building agents, the bigger opportunity might be building software for agents. These software products no longer treat humans as the only users—they treat agents as first-class citizens.
This also means that the product design logic of traditional SaaS may need to be rewritten. In the past, software companies worked hard to design better user interfaces. In the future, software companies may have to design both human interfaces and agent interfaces, and in some scenarios, machine-readable interfaces may be more important than human UI.
AI coding reduces software costs, and legacy SaaS becomes the target
In “SaaS Challengers,” YC partner Jared Friedman points out that the market is debating whether AI coding will bring an end to SaaS. For existing software companies, that could be bad news. But for startups, it might actually be a once-in-a-decade opportunity.
SaaS has been able to build moats in the past because software development costs are high—five-person startups could never quickly rewrite Salesforce or large enterprise systems. But when AI coding cuts software production costs by 10 to 100 times, the barriers built from millions of lines of code and years of accumulated features may become less unbreakable.
YC encourages founders not to start only with simple project management tools, but to challenge more complex, more expensive systems that were previously considered difficult to dislodge—such as chip design software, ERP, industrial control systems, and supply chain management. The focus here isn’t “copy an old SaaS and lower the price.” It’s using an AI-native product architecture to rethink the workflow itself. If software from day one assumes AI agents will participate in execution, the product doesn’t necessarily need to look like the SaaS of the past.
Dynamic software interfaces: users become the deployment engineers themselves
In “Dynamic Software Interfaces,” YC proposes another software design change. In the past, everyone used the same interface—at most adjusting the theme, layout, or recommended content. But after coding agents mature, users may be able to significantly modify the software they use.
For example, with the same email client, some people might want it to function like a task list, others like a calendar, and others might want it to directly become a customer tracking system. In the past, such customization required enterprise software’s forward-deployed engineers; in the future, coding agents may let users do it themselves.
This will force software companies to rethink how they deliver products. In the future, companies may not only deliver a fixed product, but a set of primitives that can be reassembled by agents. The final interface of the software might no longer be determined entirely by the developer—it may be jointly generated by the user and the agent.
From agriculture and healthcare to defense: AI is starting to enter the physical world
In this list from YC, multiple topics point to the physical world.
In agriculture, YC proposes “AI for Low-Pesticide Agriculture.” This direction isn’t building an agriculture information platform. Instead, it combines AI vision, low-cost sensors, cameras, robots, and biotechnology to break agriculture’s cycle of “spray more pesticides, but the results get worse.”
In healthcare, YC focuses on “AI Personalized Medicine.” As the costs of gene sequencing, personalized diagnostics, wearable devices, and mRNA-based therapies fall, AI agents have the opportunity to help analyze individuals’ health data—making medical recommendations and treatments more personalized.
In defense, YC proposes “Counter-Swarm Defense.” Low-cost drones give attackers a major cost advantage, and the traditional model of intercepting cheap drones with expensive missiles doesn’t make sense. YC is therefore looking for a new type of defense system that can handle drone swarms, including sensor fusion, high-capacity interception, non-kinetic defense, and even attack methods targeting the autonomous drone systems themselves.
Together, these topics show that AI entrepreneurship is moving beyond simple screen-based applications, and beginning to address sensing, control, cost, and safety problems in the physical world.
Space, hardware, and semiconductor supply chains are also included in the AI startup map
YC’s Summer 2026 RFS also places space, hardware, and semiconductor supply chains on the priority list. In the space direction, YC focuses on space electronics, especially inference chips in space. As reusable rockets increase humanity’s ability to send objects into space, computation demand in space will also increase, and chips need to be redesigned to meet constraints like weight, thermal management, and radiation.
In the hardware supply chain area, YC believes the iteration speed of U.S. hardware companies is still far behind Shenzhen. The issue isn’t just the supply chain itself, but the iteration speed across design, production, logistics, and parts sourcing. Companies that enable hardware teams to complete design and prototype production faster may become the foundational infrastructure for the next generation of hardware startups.
In the semiconductor supply chain area, YC points out that an advanced AI chip goes through roughly 1,400 manufacturing processes, spanning more than a dozen countries, taking months to complete—but supply chain management still relies heavily on spreadsheets, SAP, and phone calls. Advanced packaging, HBM, export controls, and multi-tier supplier risks all require new real-time management tools.
These directions aren’t “AI applications” in the traditional sense, but they’re closer to the real bottlenecks in the AI industry going forward: compute, hardware, supply chains, manufacturing speed, and physical deployment capability.
The real signals revealed by YC Summer 2026 RFS
The real signal in YC’s Summer 2026 RFS is that the main battlefield of AI entrepreneurship is shifting from “application-layer little tools” to “reconstructing industry fundamentals.”
This doesn’t mean consumer AI has no opportunity, nor that small tools can’t become good products. But at least judging from YC’s list, the most highly watched startup directions are no longer simply building an AI app that’s easy to demo, easy to go viral, and easy to replicate. Instead, they’re moving into areas where workflows are complex, data is fragmented, delivery responsibility is heavy, industry knowledge runs deep, and even hardware and the physical world are involved.
This also gives “AI-native” a clearer definition. An AI-native company isn’t just adding a chatbot to the bottom-right corner of a website, nor is it merely plugging LLMs into old workflows. It should redesign from day one how work gets done, how knowledge is stored, how software is used by agents, how services are delivered, and how the company forms its own self-improvement loop.
If the previous wave of AI entrepreneurship was about who could wrap models into a product the fastest, then the next wave may be about who can embed AI into the parts of industry that truly run. In other words, the AI companies that will be truly valuable in the future might not sell a tool. They might sell a new kind of company capability: understanding workflows, executing work, integrating data, taking responsibility for outcomes—and turning complex systems that used to be maintained only by human experience into infrastructure that AI can help operate.
This article: the 15 startup directions YC plans to invest in for Summer 2026—AI entrepreneurship isn’t about stuffing a Chatbot into a product. First appeared on Chain News ABMedia.
