Y Combinator's Requests for Startups is a list of "we wish someone would solve this." The Summer 2026 edition's central premise: AI is becoming infrastructure, not a feature, which forces software, services, hardware, and physical-world systems to be rebuilt from the foundation. Recurring themes: chemical-free agriculture, AI replacing services, personalized medicine, structuring company knowledge, counter-swarm defense, and the agent-era software/chip/supply-chain stack.

1. What RFS Is, and the Big Premise for Summer 2026

The RFS is YC's tradition of publicly sharing the problems they hope founders will tackle. They emphasize: this is part of what YC funds, not all of it. A specific RFS idea is "extra signal," not a requirement.

The unifying theme of Summer 2026 is that AI now sits underneath software, services, silicon, robotics, and the physical world — not on top of it. The next wave of startups is the one that rebuilds these layers from scratch.

"AI is starting to become infrastructure, not a feature."

2. Chemistry-Free Agriculture 🌱

Modern farming runs on chemicals — pesticides, herbicides, fertilizers. They worked for a while, but residue accumulates in food, water, and soil; concerns about glyphosate and similar compounds keep growing; and pests and weeds adapt and become resistant. Farmers are stuck in a worsening loop: more chemicals, less effect, more cost, more risk.

What's changed:

• AI vision can identify individual weeds and pests in real time
• Cheap sensors and cameras make dense field-level monitoring possible
• Precision robotics can act at the level of single plants, not whole fields
• Biological alternatives (microbes, peptides, RNA-based solutions) have moved from sci-fi to practical
• Engineering plant defenses — making crops less dependent on external inputs
• AGI-driven scientific breakthroughs to amplify all of the above

"A company that helps farmers reduce chemical use by 90% while producing more food won't just be a good business — it'll be a generational one."

3. AI-Native Companies That Do the Work Instead of Selling Tools

The arc:

1. Past — services were turned into SaaS.
2. 2023–2025 — services got AI copilots.
3. Now (2026) — AI-native companies don't sell software, they deliver the service itself.

Why it matters: services dwarf software in total spending and are already heavily outsourced — easy to replace with AI products. Targets called out specifically:

• Insurance brokerage
• Accounting, tax, audit
• Compliance
• Healthcare administration (claims, paperwork)

"Instead of giving you tools, just do the job."

4. Agent-Driven Personalized Medicine 🧬

Two revolutions are converging:

• Personalized diagnostics get cheap — genome sequencing is dropping faster than Moore's Law; new early-detection tools proliferate.
• n=1 gene therapy delivery and manufacturing get cheap — mRNA vectors enable patient-specific drugs; the FDA is more open to single-patient trials.

Combined with intelligent agents (think Claude Code-style harnesses) that can analyze personal health data — diagnostics, genomic scans, EHR, wearables — patients get sharper personal risk assessment and access to therapies that were previously gated to the wealthy or healthy.

5. The Bottleneck Isn't the Model — It's the Company's Knowledge

For company automation, the real obstacle isn't model quality; it's domain knowledge scattered across heads, old emails, Slack threads, tickets, and databases. Humans navigate this fuzzy mess by memory and intuition. Agents can't.

The new primitive: a company brain — not search, not a doc chatbot, but a living map of how the company actually operates, exposed as AI-executable skill files. Refunds, exception approvals, incident response — all structured so an agent can execute them safely and consistently.

"I think every company in the world is going to need one of these."

6. Counter-Swarm Defense 🚨

The framing is alarming: cheap Iranian drones reportedly disabled an AWS data center; nothing stopped them. Worse, a coordinated swarm of a thousand drones is unstoppable today.

The asymmetry:

"A Patriot missile costs $3M. An FPV drone costs $500."

Existing counter-drone defense is a tangled mix of radar, cameras, jammers, interceptors, and human spotters. Fine for hobby drones; useless for swarms. YC wants:

• High-capacity interceptors that neutralize 50 drones at once, not one
• Sensor + defender fusion software with a unified real-time picture
• Non-kinetic defenses: aerosols that disable rotors, streamers that entangle swarms
• Direct attack on the autonomy stack itself, since RF jamming will fade as drones go more autonomous

"Drone defense is becoming less about operating weapons and more about running a real-time distributed system. The winner will look more like Cloudflare than Raytheon."

7. Coding Agents Make Users Their Own Forward-Deployed Engineers 🛠️

Pre-AI, every user got the same UI with minor personalization. Real customization happened only in enterprise via Forward Deployed Engineers. Now, coding agents are good enough that users can be their own FDE.

Software companies need to ship not finished UI but shared primitives, expecting users to dramatically modify the final interface. Open questions:

• Should source code be exposed for the agent to operate on?
• Is mutation limited to the front end, or can middleware be modified at runtime?

8. Inference Chips for Space

Reusable rockets (SpaceX, Stoke Space) are blowing open orbital throughput, and demand for in-space inference compute will follow. YC explicitly calls out space inference chips with optimization for:

• Mass
• Thermal characteristics
• Radiation tolerance

"The market for inference chips in space is going to be enormous."

9. The Real US Hardware Gap: Iteration Speed, Not Supply Chain

YC has been investing more in hardware — medical devices, home robots, aerospace — but the US is too slow vs. China (Shenzhen specifically). In Shenzhen, design-to-new-component cycles can run in a day; in the US the same loop takes weeks. The compounded gap is enormous.

The diagnosis: it's not just supply chain, it's iteration speed. Shenzhen has dense supplier networks, fast turnarounds, and tight design-to-manufacturing coordination. The US doesn't. Some startups (Hlabs W26, Prototyping.io P26) are filling small parts of this stack, but most of it is empty.

YC wants startups that:

• Produce parts dramatically faster
• Enable rapid hardware iteration
• Tightly integrate design, manufacturing, and logistics

10. Lunar and Space Industrial Capability

A bigger play: in-space industrial capacity, especially using lunar regolith. Electrolyze regolith to extract silicon, aluminum, iron, and titanium. 3D-print structures from molten regolith. Lower gravity and absent atmosphere actually make some processes more efficient than on Earth.

11. Inference Chips Designed for Agent Loops

Most AI silicon is built for the "prompt → response" world. Agents don't work that way. They run loops: tool calls, branching, backtracking, dozens of context-preserving steps. GPUs hit only 30–40% of theoretical utilization on this workload because work is bursty across:

• Memory-bound model calls
• I/O-bound tool use
• CPU-bound orchestration

That utilization gap is the opening for purpose-built silicon. Specifically for agent loops:

• Fast model-to-model context switching
• Native speculative decoding
• Memory architecture matched to a persistent KV cache across the execution graph

"Groq's real insight wasn't the chip — it was the compiler that made the chip work."

12. Unbundling Legacy SaaS

Investors fear AI is destroying trillions in software market cap. YC flips that into a startup opportunity:

"Bad news for incumbents — good news for startups."

SaaS won because writing custom software was prohibitively expensive. AI dropped that cost 10–100×, and "tens of millions of lines of legacy code" stops being a moat.

Attack vectors, escalating:

1. Clone the existing product, sell at 1/10 the price.
2. Redesign the workflow AI-native — not "chatbot bolted on 2010 UI."
3. Bundle ten-point solutions into one suite.
4. Replace a $50K/seat product with open source plus services and hosting.

Don't stop at easy targets like project management. Go after the supposedly impregnable: chip design, ERP, industrial control, supply chain.

13. Software for Agents, Not Humans

The "next billion users" of the internet won't be humans — they'll be AI agents. They already browse, research, buy, and use legacy CRMs. Today's software is built for human clicks: slow, inconsistent, brittle.

"Agents need a completely different foundation."

What matters in the agent era is machine-readable interfaces: APIs, MCP, CLIs. Documentation has to be strong enough that agents can discover, sign up for, and use a tool without human help. Most major software categories will be rebuilt agent-first.

"Everyone's building agents. The bigger opportunity may be building the software the agents depend on."

14. AI Opens Mega-Enterprise Sales to Tiny Teams

The classic advice "startups should sell to startups" still holds, but AI has opened F100-class customers to early teams. Three reasons:

1. Big-company decision makers are actively searching for teams that can solve core problems with AI.
2. With AI, a 2–3 person team can ship something a Fortune 100 will buy in months, not years.
3. Big-company leaders increasingly know what to build internally vs. buy externally — and what the cost of inaction is.

"It is no longer unusual for a startup's first customer to be one of the world's largest companies."

15. Semiconductor Supply Chain 2.0

A frontier AI chip touches ~1,400 process steps, dozens of countries, 5 months of lead time — and is still managed via spreadsheets, SAP, and phone calls. In 2021, a $300 chip shortage prevented $210B in vehicle production, with near-zero visibility down to second- and third-tier suppliers.

It's gotten harder:

• TSMC advanced packaging is the AI compute bottleneck; NVIDIA has 60%+ of capacity
• HBM memory is sold out through 2026
• Export controls shift quarterly
• US fabs from CHIPS Act require essentially building new supply chains

Real-time allocation tracking, multi-tier risk monitoring, export compliance — none of this is well-built. Adding features to SAP won't solve it. The opportunity is for teams who actually understand wafer allocation and packaging constraints.

16. The Queryable Company

The shared trait of AI-native companies: the entire company is queryable. Every meeting recorded, every ticket tracked, every customer interaction captured. With that input, an intelligence layer can read the company and learn how it works.

The shift: open-loop → closed-loop. Don't make a decision and check the result weeks later — let a system monitor progress, compare to expected state, and adjust.

The current barrier is integration drudgery — Slack, Linear, GitHub, Notion, call recordings, etc. The product YC wants isn't a dashboard; it's a connective layer that turns company outputs into a self-improving loop and warns when something is wrong.

Closing

The Summer 2026 RFS message is consistent. With AI now strong enough, the bottleneck is the real-world stack — data, knowledge, supply chains, hardware, defense systems, interfaces. Teams rebuilding those layers stand to capture huge ground. YC specifically calls out chemistry-free agriculture, services delivered by AI, personalized medicine, company brains / AI OS, counter-swarm defense, agent-first software, and chips/semiconductors for the agent era.