Picture this: You walk into Y Combinator’s Demo Day in fall 2025, and virtually every pitch starts with “We’re using AI to...” One founder claims their system will run your entire marketing operation. Another is building 40-gram drones that hunt mosquitoes. A third just fired up a fusion reactor in orbit (well, not yet…).
If you squint, the whole batch looks like a sci-fi anthology written by an optimistic AI and a mad scientist who got into the same coffee supply.
Here’s what’s actually happening. Over half the companies are building AI agents to do specific jobs: filing freight claims, reconciling accounts, troubleshooting server crashes, generating ad campaigns. The rest split between generative media tools (turn your kid’s crayon sketch into a cartoon), AI infrastructure (the picks and shovels for this gold rush), and a handful of genuinely wild bets on drones, space reactors, and robotic factories.
This is a guided tour through that landscape. We’ll map the consensus plays everyone’s making, spot the outliers that make you do a double-take, and figure out what it all means for the near future. By the end, you’ll see how this batch sketches a world where AI is the default coworker and even dusty old industries get a futuristic makeover.
The New Normal: Your Coworker Doesn’t Eat Lunch
Imagine an office where one “employee” never eats, sleeps, or complains about Mondays. That’s because it lives in the cloud. Walking through YC’s Fall 2025 batch feels like touring a factory stamping out AI assistants, each tailored to a different job. Marketing, logistics, finance, customer support: pick a department, someone’s building an AI for it.
Why now? Generative AI crossed a capability threshold. What used to require armies of humans and careful programming can now be handled by models you can spin up in an afternoon. Founders smell efficiency gains, and there’s a whiff of FOMO too. Nobody wants to be the sucker building non-AI software in 2025.
Let’s look at what they’re building:
Digital coworkers for business operations
Several teams are automating the grunt work nobody wants to do. Prox built a “digital logistics teammate” that lives inside Microsoft Teams and handles tedious tasks for freight companies. Instead of a human spending hours filing claims or chasing invoices, Prox’s AI does it in seconds.
Zalos positions itself as a virtual finance analyst that reconciles accounts and spots anomalies so your finance team doesn’t become a bottleneck. Questom focuses on custom merchandise companies (think Custom Ink), automating sales inquiries and support tickets.
The vision: an AI embedded in each department, handling the repetitive stuff so humans can focus on decisions.
AI agents for developers and IT
Even engineers are hiring AI interns now. Deeptrace builds agents that troubleshoot production outages, digging through logs to find root causes while the human on-call gets extra sleep. Redapto creates self-improving support systems that learn from every customer interaction and automatically update their knowledge bases.
These aren’t just productivity tools. They’re a meta-layer: AI helping teams build and maintain other software, including more AI.
Copilots for niche professionals
Many of these assistants target specific jobs. Uplane is an “AI autopilot” for ad campaigns that generates ads, builds landing pages, and optimizes spending across platforms. Apriora automates recruiter tasks. One startup built an AI that detects patent infringement and generates claim charts in minutes. Telemetron offers an AI support platform specifically for hardware companies like medical device makers.
If your job involves repetitive data work or form-filling, someone in this batch is trying to hand it to an AI.
What does this mean for you?
These AI coworkers will probably show up in your daily tools soon. They’ll triage emails, prep spreadsheets, handle level-one support tickets. Best case: you’re freed from drudgery to focus on creative and strategic work. Worst case: they become competition for certain entry-level jobs. Either way, expect your relationship with work to shift. You might manage an AI agent, collaborate with one, or discover the service you hire quietly uses one behind the scenes.
The consensus from YC founders? If a task is dull, repetitive, and rules-based, an AI is coming for it.
The Creative Copilots: Don Draper Meets the Algorithm
Don Draper wouldn’t recognize the new creative team. Instead of cigar smoke and sketchpads, imagine a room of AI avatars churning out campaign ideas in minutes. A marketing director can think up an ad concept over breakfast and have polished assets by lunch. No film crew, no agency retainer.
This isn’t fantasy. It’s literally the pitch of multiple startups in this batch.
AI ad and video studios
Absurd brands itself as an absurdly fast AI ad studio. They claim they can deliver production-quality launch videos in 72 hours using a “multi-agent orchestration layer” guided by human creative directors. Their AI agents collaborate over Slack and a canvas tool, generating shots and scenes the client watches materialize in real time. The results? Viral-style videos that have already racked up hundreds of thousands of views.
Velvet (built by a team from Meta’s FAIR lab and Adobe) wants to replace your video agency with a web app where you type a concept and get a polished video out. Claybird offers something similar: studio-quality ads with minimal effort.
Generative content for e-commerce and social media
It’s not just big brands. SellRaze built a video-driven shopping platform where sellers can easily create engaging short videos for their products (imagine TikTok meets eBay). Their AI auto-generates video listings. They claim 200,000 sellers and $1M in annual revenue already.
Bluma calls itself “Canva for short-form video ads,” simplifying the process so small businesses can crank out marketing content at speed.
AI for entertainment and education
On the playful end, consider Pixley. This startup gives kids a magic wand to create cartoons. As the founders put it: “With Pixley, any drawing becomes an animated character, and every episode is shaped by the child’s imagination.” It’s like an AI Pixar that turns crayon sketches into personalized Saturday morning cartoons.
The tech leverages generative AI for images and storytelling, trying to make screen time more active and creative. Instead of kids passively consuming content, they become directors of their own shows.
The theme across all these: user gives high-level idea, AI produces polished result. Lower cost, faster turnaround, and the ability to A/B test a hundred variations since an AI can spit out endless tweaks.
There’s a witty irony here. For years, creatives feared “AI will replace us.” Now startups pitch “AI will empower you to replace your need for others.” The marketing manager becomes a one-person content studio. The kid becomes a cartoonist without learning animation.
It democratizes creation, but it also floods the world with more content. If every brand can crank out dozens of “viral” ads a day with AI, do any of them truly go viral? The noise might just increase. You might end up drowning in an ocean of AI-made videos, each fighting for your scroll-through attention.
Building the Picks and Shovels of the AI Gold Rush
During the original Gold Rush, the surest way to succeed was selling picks and shovels to the miners. In 2025’s AI boom, a bunch of YC startups are doing exactly that: building the infrastructure and developer tools all these AI applications need under the hood.
Think of them as the backstage crew making sure the AI stars can perform without falling through a trapdoor.
Multi-agent and data control infrastructure
Agentic Fabriq has a clear thesis: as companies deploy AI agents internally, they need a control center to manage who can access what data. Fabriq positions itself as “the single control plane for data permissioning” across an organization’s AI and human users. It’s like an identity manager, but for a world where some of the “users” are bots.
The fact this exists tells you enterprises are seriously considering deploying multiple AI agents in production. Enough to worry about messy questions like governance and compliance.
Unsiloed AI tackles a different bottleneck: feeding unstructured documents into AI reliably. They built vision models to parse complex documents (text, tables, images) into structured, queryable data. Think of it as supercharged OCR, cleaning up data so AI apps can work on top without choking on garbage input.
Tools for developers working with AI
Several startups realized building with AI (especially agents) is hard, so they’re making tools to help developers prototype, test, and maintain AI systems.
Sanctum lets developers launch new features into a parallel universe of AI-simulated users before exposing them to real customers. They create AI models that behave like your users (based on session recordings), so you can catch bugs and UX issues early. It’s a novel twist on QA: instead of test scripts, you have AIs wandering through your app clicking things to see what breaks.
Hyperspell tackles a known weakness of AI agents: short-term memory. Agents without memory are impressive but forgetful. If you’ve used ChatGPT, you’ve seen how it can forget instructions from earlier in a conversation. Hyperspell gives agents a memory upgrade so they stay consistent through multi-step tasks.
AI-enhanced coding and software design
Specific has a wild dream: “a platform for engineers to build backend services where engineers don’t write code at all, just natural-language specs and tests.” Describe your software in English, and their AI builds it behind the scenes. This is the logical extreme of GitHub Copilot: not just auto-completing lines, but no lines at all.
Sourcebot is more grounded: an open-source platform to help developers understand huge codebases with AI assistance. Thousands of engineers at big companies already use it. Navigating legacy code is hard, and an AI that acts as smart code search is valuable. It’s essentially Google for code, with some intelligence baked in.
As our tools (AI agents) become more powerful, the meta-tools to manage them become essential. This batch assumes AI won’t just be a novelty; it’ll be ubiquitous. That means we need guardrails, debuggers, and memory keepers for the AI itself.
It’s a bit meta, yes: AI to manage AI. But it tracks with computing history. We built operating systems to manage programs, version control to manage code, and now agent orchestration systems to manage AIs.
The smartest YC founders know that in a gold rush, selling shovels can be as lucrative as finding gold. Here, the “shovels” are developer platforms, APIs, and enterprise middleware. Quieter than consumer apps, but potentially stickier businesses.
Beyond Software: When Startups Go Sci-Fi
Amid all the AI software pitches, a handful of founders strolled in with ideas straight out of a Neal Stephenson novel.
One team is building 40-gram drones that hunt and kill mosquitoes to curb malaria. Another is launching the world’s first in-orbit fusion reactor to power industries in space. In a garage (or lab), engineers are assembling fully autonomous robotic factories, dreaming of a future where scaling hardware is as easy as deploying code. One startup is literally tackling hypersonic missiles with space-based interceptors.
These are the iconoclasts. The startups that zigged when everyone else zagged into AI SaaS. They’re few, but they make up for it in audacity.
Public health meets robotics
Tornyol is perhaps the most visually striking idea: micro-drones that kill mosquitoes. The founders combined smartphone microphones, car parking sensors, and control algorithms to turn cheap toy-sized drones into automated mosquito hunters.
Why? Mosquitoes still cause over a million deaths a year via malaria and dengue. If you can make mosquito control 100 times cheaper (their claim), you could save lives and make governments very happy.
It’s an edge case in the batch, neither pure software nor enterprise. But it’s tackling a huge global problem with a fresh approach: real-world search-and-destroy AI. Timing matters too. Drone components and AI vision have gotten so affordable and lightweight that a mosquito-seeking drone sounds plausible in 2025. Ten years ago, it would’ve sounded bonkers.
New space race stuff
Zephyr Fusion wants to put a fusion reactor in orbit. Fusion power has been “20 years away” forever. Zephyr’s twist is trying it in space, where a reactor could run more freely (and any unfortunate events stay far above our heads). If you buy their premise, a working fusion source in orbit could enable all kinds of space industry.
Wardstone is building the next generation of missile defense: satellites armed with kinetic interceptors to knock out hypersonic missiles from space. This sounds like something only nations would do, yet here’s a startup taking it on. The macro backdrop is obvious: geopolitics and defense tech are more salient now, and technologies like sensors and miniaturized interceptors have advanced.
YC getting into defense tech is itself notable. These companies assume that even defense and space, historically government-dominated, are now startup territory thanks to venture funding and tech crossover. SpaceX and Anduril taught us it’s not impossible.
Automation of hardware manufacturing
Tensr is “building fully autonomous robotic factories,” making scaling hardware as effortless as scaling on AWS. They’re Berkeley robotics grad students who won an autonomous IndyCar racing competition (self-driving racecar at 160mph). Now they’re applying similar tech to manufacturing.
The assumption: recent leaps in robotics (computer vision, control, machine learning for manipulation) plus demand for onshore manufacturing make it the right time for “lights-out” factories. With global supply chain challenges fresh in memory, an autonomous micro-factory could be a game-changer for hardware startups. Imagine iterating on physical products as quickly as software.
Tensr treats manufacturing as code: you deploy a factory like deploying a server.
AR and next-gen interfaces
AirCaps created lightweight AR glasses that caption and translate conversations in real-time. Think of it as live subtitles for life. They’re initially targeting people with hearing loss (a meaningful assistive tech angle), but the device also offers language translation and on-the-fly summaries.
AR has been “on the cusp” for years, with many failures (Google Glass). But AirCaps is going after a use case that’s both high value and technically feasible now, thanks to advances in real-time speech recognition and translation. Their existence suggests new optimism for AR hardware: components are lighter, and AI is powerful enough that smart glasses can genuinely enhance conversations.
What ties these frontier startups together? Willingness to embrace risk and hardware that most of their batchmates avoid. They can’t iterate weekly or A/B test their way to success. You can’t MVP a fusion reactor like you can a SaaS app.
These founders share boldness (or craziness, depending who you ask) to tackle problems that might take years of R&D. It’s refreshing amid the quick-turn AI apps. If even one of these hits, it could redefine an industry.
The Founders: Ex-FAANG, PhDs, and the Man in the Warehouse
Meet the characters driving these startups.
There’s the ex-Googler who left a cushy job to solve a problem they saw on the inside. The PhD researcher who realized her cutting-edge work on foundation models could become a product. The scrappy operator who lived the pain point for years, like the guy who literally grew up in a family logistics business and now automates warehouse workflows.
Reading founder bios feels like scanning LinkedIn’s greatest hits: ex-DeepMind, ex-Meta, Stanford CS, Berkeley robotics, second-time founder, Forbes 30 under 30. It’s a concentration of talent that would make a recruiter drool.
Big Tech and AI research alumni
Many founders cut their teeth at Google, Meta, DeepMind, OpenAI, or top research labs and are now striking out on their own. Tell If AI’s team includes ex-DeepMind and Yandex researchers tackling deepfake detection. Velvet’s team worked on video generation at Meta’s FAIR and infrastructure at Adobe.
There’s a brain drain from Big Tech into startups right now, likely fueled by the allure of moving faster and owning equity. Plus, frankly, some layoffs and pivoting in Big Tech’s AI strategy opened doors. For YC, these founders are gold: cutting-edge knowledge and credibility to tackle tough problems.
Domain insiders solving their own pain
Another archetype is the founder with deep domain experience who faced frustration firsthand. The Prox team reportedly “literally grew up in warehouses” and has serious logistics chops. They saw inefficient freight operations and decided to automate them.
Consider Telemetron: one founder’s stint at SpaceX Starlink likely exposed how standard support tools failed for complex hardware, hence an AI platform tuned for hardware support. Play Health’s founding team includes a prior healthtech exit, giving them unique insight in tackling women’s midlife health.
These folks aren’t opportunistically jumping on trends. They’re scratching personal itches or seizing opportunities spotted in previous careers. That magical “founder-market fit” investors love.
Repeat founders and seasoned operators
YC F25 isn’t just fresh grads. There are second-time (or third-time) founders and people who’ve scaled companies before. One of Zalos’ co-founders helped scale a fintech startup to near-unicorn status. That implies he’s seen fast growth and where operations break.
Several teams boast credentials like “previous exit” or “previously built X to Y users.” In general, YC has been attracting more experienced founders recently, and it shows.
Global and diverse talent (but SF-centric)
Many startups have a presence in San Francisco, which is no surprise since YC encourages moving there during the batch. But founders come from everywhere: India’s IITs, Europe’s research labs, North African tech scenes, all converging in Silicon Valley.
One interesting pattern: small, lean teams. Most companies have just two or three employees at this stage (often just the co-founders). It’s striking to see a startup tackling satellite defense or multi-agent infrastructure with literally two people and a laptop (okay, maybe a few servers).
This speaks to a broader tech assumption now: thanks to cloud computing, open-source, and existing AI models, tiny teams can attempt things that used to require an army. The cost of daring greatly has dropped. A couple of brilliant minds can reshape an industry from a co-working space.
If we anthropomorphized the typical F25 founder, we’d get someone in a hoodie emblazoned with both a university lab logo and a FAANG logo, brain buzzing with ideas, slight bags under the eyes from too many late nights debugging. In one pocket, a crumpled note of a pain point they swore to fix. In the other, a sketch of an architecture that might just do it.
They are, in short, builders.
The Big Picture: What It All Means
Stepping back, what do these trends tell us about the world of startups and tech in late 2025?
AI isn’t a vertical, it’s horizontal infrastructure
Just as electricity and the internet seeped into every industry, AI is doing the same. The YC F25 batch treats AI less like a special category and more like a given: a means to an end for nearly any problem. Whether it’s enterprise SaaS, creative tools, healthcare, or defense, founders apply AI to differentiate their products.
The cohort validates that AI has transitioned from research hype to mainstream utility. Nearly every B2B idea here has AI at its core. This is a consensus bet that feels very much “of the moment.”
The risk? If everyone’s doing it, competition will be fierce. Only those with genuine moats (data, distribution, proprietary tech) will survive the thinning.
Human-in-the-loop is still alive
Despite the AI-everywhere vibe, many startups acknowledge AI works best with humans guiding it. Absurd has human creative directors guiding their ad-making agents. Fabriq’s whole point is giving people control over AI access. The Hog arms human marketers with a command center of AI helpers.
This reflects a nuanced view. These founders aren’t naive about AI replacing everyone. It’s about leveraging AI to augment skilled people, letting one person do the work of five.
The witty take: in the future, your intern might be an AI, but you’re still the boss who knows what needs doing.
Clusters reflect “why now” factors
Each cluster we identified has strong timing rationale:
Multi-agent and copilot startups emerged because the tech (GPT-4, 5, etc.) finally works well enough, and businesses actively seek productivity boosts due to labor shortages and remote work.
Generative media startups are here because generative models for images and video crossed a quality threshold recently, and the content boom on TikTok and Instagram means video is the internet’s language, with every brand needing it.
AI infra and dev tools popped up because so many companies tried gluing AI into their apps last year and felt the pain (security issues, debugging, cost optimization). The ecosystem needs better support tools, a classic picks-and-shovels response as the AI gold rush matures.
Frontier tech bets (drones, fusion, space) align with external trends: climate change plus cheaper drones enable mosquito control innovation; geopolitics plus better rocketry enable private missile defense; SpaceX’s success plus advances in plasma physics create appetite for space fusion; rising labor costs plus better robots enable autonomous factories.
These aren’t random. They’re a few years or months into a curve where the startup approach became feasible.
Iconoclast swings versus consensus plays
We see both. The consensus plays are piles of AI B2B SaaS. It’s almost crowded how many are attacking workflow automation or sales enablement with AI. A shakeout will happen. Not all the “AI for X” will survive once the market picks winners.
The iconoclasts (deeptech stuff) face different risks: technology risk and adoption risk. But if they hit, they really hit. Big moats, few competitors. YC’s approach seems to be: fund a lot of the former (some will stick and become nice businesses) and a few of the latter (each a lottery ticket for a potential SpaceX-scale outcome).
From an investor’s perspective, it’s a barbell strategy. From a narrative perspective, it makes the batch more exciting. You get bread-and-butter SaaS and a dash of “crazy dream.”
What’s the vibe of this cohort?
It feels like we’re at a hinge point where the definition of software is evolving. Software isn’t just code written by humans anymore. It’s behaviors learned by models, decisions made by agents, hardware becoming software-defined (factories, drones).
The YC F25 startups collectively signal a future where intelligence is woven into everything. The mundane gets automated. The creative gets turbocharged. The physical gets an AI layer or two.
In a more playful tone: If we anthropomorphize technology eras, the 2025 startup scene is like a teenager who just discovered a new superpower (AI) and is trying it on everything. Often to impressive effect, occasionally to absurdity (did we need AI in a to-do list app? Probably someone in the batch is doing that too).
It’s an enthusiastic, somewhat chaotic burst of experimentation. Some experiments will fail spectacularly. Others will redefine how we live and work.
The Sweeping Takeaway
The YC Fall 2025 batch shows us a landscape dominated by AI-driven ambition, tempered by a few daring detours into hardware and frontier science. The dominant clusters (AI copilots, generative media, AI infrastructure) tell us the present and near-future of startups is all about intelligent automation and augmentation.
Meanwhile, the oddballs (mosquito drones, fusion in space) keep the spirit of wild innovation alive, saying “hey, not everything interesting is an app.”
The founders are highly credentialed, deeply technical, and often solving problems they know intimately. A recipe that bodes well for executing these lofty ideas.
If there’s one sweeping takeaway, it’s this: the lines between science fiction and startup reality are blurring. Need a personal AI assistant? Already here. Want to eradicate disease-carrying bugs with tiny robots? They’re on it. Think we can power a space colony? Someone’s working on the reactor.
The F25 cohort is both a mirror of what’s hot now (AI, AI, AI) and a window into what could be next. It’s as exhilarating as it is overwhelming, and it leaves us with plenty to watch in the years to come.
One final thought:
Ten years ago, if someone pitched “AI that writes your ads” or “drones that hunt mosquitoes,” you’d have laughed them out of the room. Today, they’re getting venture funding and building real products. Whatever the next ten years bring, it’s probably weirder and more ambitious than we imagine. And if this batch is any indication, it’s being built right now, in garages and co-working spaces, by small teams with big dreams and really good WiFi.