We ran every YC Request for Startups through our validator. Here's what came back.
A few weeks ago YC published its Summer 2026 Request for Startups — sixteen ideas the partners are actively trying to fund. It's about as authoritative a list of "good startup ideas" as exists. So we did the obvious experiment: we pasted each RFS into Unycorn, our startup idea validator, as-is. No reframing, no founder pitch, no cherry-picking — just the YC blurb in, full report out. The point wasn't to argue with YC. The point was to see where our model agrees with the smartest investors in the world, and where it disagrees, why.
AI for Low-Pesticide Agriculture
Garry Tan's pitch is that AI vision plus precision robotics can cut pesticide use by 90% while raising yield — an unsolvable problem that suddenly isn't. The TAM is genuinely massive, but the validator docks the idea on go-to-market: farmers are notoriously slow buyers, and the capex story for ag-robotics has burned a lot of investors. The signal worth watching is whether retrofit beats greenfield — software-only solutions riding on existing John Deere hardware score very differently from end-to-end robotics plays.
AI-Native Discovery Engines
Jon Xu argues that science is leaving the copilot era and entering the closed-loop era — models proposing hypotheses, robots running experiments, results feeding back in. Demand pressure is sky-high in drug discovery and materials science; the validator's concern is wedge. "Discovery engine" as a category is broad enough to mean almost anything, and the companies that win will pick one painfully specific corner of the workflow and own it before scaling.
AI-Native Service Companies
Gustaf Alströmer's pitch is the most obvious-in-hindsight RFS on the list: the services market is many times larger than the software market, and AI lets you sell the work itself instead of a tool. The model flags the category as crowded — every YC batch since W24 has had bookkeeping, accounting, and compliance plays. Differentiation is no longer "can AI do this work" but "which vertical has the worst incumbents and the most reluctant buyers" — that's where pricing power lives.
AI Personalized Medicine
Ankit Gupta's pitch is that genome sequencing, n-of-1 therapies, and agent-driven analysis are all collapsing in cost at once, and personalized medicine is finally feasible. The validator agrees on tailwinds and is brutal on regulatory drag — every dollar of value created here has to clear an FDA path that punishes startup pace. The companies that win will look like infrastructure for clinics and labs already doing this work, not direct-to-patient.
Company Brain
Tom Blomfield wants someone to build the "company brain" — a living map of how a business actually works, sitting between fragmented internal knowledge and the AI agents trying to act on it. Demand pressure is real; defensibility is the worry. "Company brain" is essentially RAG-over-enterprise-systems, a category Glean, Notion AI, and every hyperscaler is already attacking — winning requires a non-obvious wedge most teams haven't found yet.
Counter-Swarm Defense
Tyler Bosmeny wants founders to flip the economics of drone defense, where a $3M Patriot interceptor currently stops a $500 FPV drone. The strategic importance is undeniable and the timing is sharp post-Ukraine. The validator's concern is buyer concentration — the only customers are governments, deal cycles run in years, and primes (Lockheed, RTX) hold the integration relationships. The winning archetype here is Palantir-shaped, not YC-batch-shaped, and very few founders have the patience for that.
Dynamic Software Interfaces
Ankit Gupta's second RFS argues that coding agents will let every user become their own forward-deployed engineer, with personalized interfaces over shared software primitives. Beautiful vision; tough validation. The model flags weak demand pressure — most users don't actually want to customize software, they want defaults that work. The opening is likely in prosumer power-user tools (Notion, Linear, Raycast) where the customization muscle already exists, not mass-market apps.
Electronics in Space
Philip Johnston (Starcloud) wants inference chips designed for orbit — mass-, thermal-, and radiation-optimized silicon to serve the compute demand reusable rockets will unlock. The thesis is provocative; the validator is skeptical of near-term unit economics. Inference still happens on Earth more cheaply than orbit for essentially every realistic workload, and the moment that changes is downstream of solar costs, downlink bandwidth, and orbital cooling — three problems that aren't a startup's to solve.
Hardware Supply Chain
Nicolas Dessaigne wants someone to close the iteration-speed gap between Shenzhen and the US, where the same hardware loop takes days versus weeks. Demand pressure scores high — every hardware founder has the same complaint. Defensibility is the question: parts-on-demand businesses tend to compress to commodity margins fast unless the startup owns the design tooling layer too. The winners here will be software companies that ship parts, not the other way around.
Industrial Capabilities in Space
Adi Oltean's RFS is to mine and manufacture on the moon — electrolysis of regolith, 3D printing in vacuum, raw materials extracted in situ. The validator agrees this is generational technology and prices in a 15-year timeline. For founders, that means picking a piece of the stack that has a paying customer before lunar industry exists — NASA contracts, defense interest, or terrestrial spinoffs from the same physics.
Inference Chips for Agent Workflows
Diana Hu's pitch is sharp: today's GPUs run at 30-40% utilization on agent workloads because agents loop, branch, and backtrack in ways prompt-response chips weren't designed for. Demand is real and rising; the validator is cautious on capital intensity. Silicon startups need ~$100M to first product, and the architectural insight has to be defensible against NVIDIA's compiler team plus three well-funded incumbents. The founder bar here is "ex-Groq, ex-NVIDIA, with a working prototype" — not zero-to-one.
SaaS Challengers
Jared Friedman is asking founders to attack legacy SaaS now that AI has collapsed the cost of producing software 10-100x. The model loves the timing and flags the obvious failure mode: cloning a SaaS for one-tenth the price gets you one-tenth the revenue, not the same revenue. The interesting wedge is products incumbents can't follow — open-source replacements for $50K-per-seat tools, or category-collapsing bundles that re-architect the workflow rather than copying it.
Software for Agents
Aaron Epstein argues that the next trillion users are agents, and they need machine-first software — APIs, MCPs, CLIs, ruthless documentation — not buttons designed for humans. Demand pressure is climbing fast; the validator's concern is that "rebuild every category for agents" is too broad to be a company. The winners pick one category (commerce, CRMs, dev tools, payments) and become the default agent-facing layer there.
Supply Chain 2.0 for Semiconductors
Diana Hu's second RFS targets the $50K-cars-held-up-by-$300-chips problem: a $500B+ supply chain still run on spreadsheets and SAP. The validator scores demand high, timing perfect (CHIPS Act, TSMC bottlenecks, export controls), and defensibility surprisingly strong — domain expertise in wafer allocation is rare enough to be a moat. The founder profile that works is ex-TSMC, ex-NVIDIA, ex-Applied Materials — not generalist enterprise SaaS.
The AI Operating System for Companies
Diana Hu's third RFS imagines the closed-loop company — every meeting, ticket, and customer interaction queryable by an intelligence layer that learns and adjusts in real time. The model flags strong demand and very weak defensibility: this is the same problem Glean, Notion, and every hyperscaler is racing toward, and the integration moat dissolves the moment LLMs get better at ingesting messy enterprise data (i.e., now). The wedge is probably a vertical — agencies, law firms, dev teams — not a horizontal OS.
A few patterns jumped out across the run.
First, AI-native ideas score higher on demand pressure than they do on defensibility. The market is real; the moat is what kills you.
Second, the deep-tech ideas (space, semiconductors, swarm defense) score lower on timing than they should on importance — the validator keeps flagging long capex cycles and narrow buyer pools. Fair criticism, even when the prize is huge.
Validation isn't a substitute for conviction. YC's job is to back founders who see something the model can't. Our job is to give you a sober read before you spend a year of your life on an idea. Validate yours →