Y Combinator is out with its 2026 request for startups
Daniel Levi
Posted On February 4, 2026
Y Combinator has released its 2026 Request for Startups, a tradition that offers founders a glimpse into the ideas and problem areas the world’s most influential startup accelerator wants founders to tackle next.
The 2026 request doesn’t just reflect what YC is interested in funding. It reveals how the firm believes the nature of startups is changing — and where the next generation of category-defining companies will come from.
The throughline is clear. AI has crossed a threshold where it no longer just speeds up existing workflows. It reshapes entire systems. As a result, YC’s focus has expanded well beyond traditional software into finance, government, industry, energy, and physical labor — areas once considered too slow, regulated, or complex for venture-scale innovation.
This year’s request isn’t about chasing trends. It’s about rebuilding foundations. From how products are conceived, to how money moves, to how governments operate, to how physical work gets done, YC is signaling that the biggest opportunities now live where intelligence meets real-world constraints.
While YC stresses that founders don’t need to work on these ideas to apply, the document is more than a suggestion list. It functions as a directional map of where the firm believes the next wave of breakout companies will emerge — shaped by AI-native workflows, shifts in financial infrastructure, government modernization, and a renewed push to rebuild the physical economy.
Taken together, the list reflects a deeper shift: startups are no longer just software companies. They’re becoming systems companies, blending AI, regulation, hardware, energy, and human labor into entirely new operating models.
What follows isn’t a checklist for founders. It’s a map of where the frontier is opening — and what kinds of builders YC believes are best positioned to cross it.
Here’s what stands out.
Y Combinator’s 2026 Request for Startups
Cursor for Product Managers
YC’s starting point is blunt: AI has made it dramatically easier to write code, but code was never the hardest part of building a product people want.
Tools like Cursor and Claude Code shine once the team already knows what to build. But getting to that point is where most products win or die. Product management, whether it’s done by founders, engineers, or dedicated PMs, is about talking to users, understanding the market, synthesizing messy feedback, and deciding which problems are actually worth solving.
Historically, the output of that work has been a stack of human-to-human artifacts: product requirements docs, Figma mocks, and Jira tickets. They exist for one reason: to communicate intent to human engineers.
YC argues that it’s now outdated.
Teams may use AI in isolated moments, but there’s still no system that supports the full end-to-end product discovery loop. YC’s vision is a tool that takes in real inputs, not just prompts. You upload customer interviews and product usage data, ask a simple question like “what should we build next?”, and the system responds with more than a brainstorm. It outputs an actual feature outline with a clear explanation, grounded in customer feedback, of why the feature matters.
Then it goes a step further. It proposes concrete changes across the product surface area: UI adjustments, data model changes, workflow updates. And instead of leaving you with a vague plan, it breaks the work into development tasks that can be handed directly to your coding agent of choice.
In YC’s framing, this becomes a “Cursor for product management”: an AI-native system focused on helping teams figure out what to build, not just how to build it.
The timing matters. As agents increasingly take the first pass at implementation, the bottleneck shifts upstream. The way teams define and communicate “what to build” needs to evolve, because the consumer of that “spec” is becoming an agent, not a human engineer.
What this signals:
YC is saying the next big advantage won’t be writing code faster. It’ll turn real-world user signals into accurate product decisions faster and translate those decisions into agent-ready work without losing the thread.
AI-Native Hedge Funds
YC frames this as a familiar story repeating itself.
In the 1980s, a small group of hedge funds began using computers to analyze markets. At the time, it looked unnecessary and even naïve. Today, quantitative trading is table stakes. YC believes we’re at a similar inflection point again — but this time, the shift isn’t just from human intuition to math. It’s from human research to autonomous intelligence.
The core idea isn’t “hedge funds using AI.” It’s hedge funds built entirely around AI from day one.
YC points out that the world’s largest funds have been slow to adapt. Internal compliance friction, legacy workflows, and cultural inertia make it difficult to deploy modern AI systems at scale. In one example, a former quant researcher describes asking compliance for permission to use ChatGPT and never receiving a response.
That hesitation, YC argues, creates the opening.
An AI-native hedge fund would not bolt models onto existing strategies. Instead, it would use swarms of agents to do what analysts and traders do today: reading 10-Ks, parsing earnings calls, reviewing SEC filings, synthesizing analyst research, spotting patterns, and generating trades. The advantage isn’t speed alone. It’s the ability to form entirely new strategies that humans would never arrive at on their own.
YC’s bet is that the next Renaissance, Bridgewater, or D.E. Shaw won’t look like today’s incumbents — and won’t come from inside them either.
What this signals:
YC believes some of the most powerful AI companies will be invisible to consumers and quietly embedded inside financial systems. Alpha, not attention, is the product.
AI-Native Agencies
YC starts from a long-standing truth in the services world: agencies are notoriously hard to scale.
Margins are thin, work is slow and manual, and growth usually means hiring more people. Headcount becomes the ceiling. Even the best agencies struggle to escape this dynamic because their output is tightly coupled to human labor.
AI breaks that equation.
YC’s insight is that, instead of selling software that helps customers do the work themselves, agencies can now use AI internally and sell the finished output directly—and charge dramatically more for it. In other words, the value isn’t in the tool. It’s in the result.
YC points to several concrete examples. A design firm could use AI to generate high-quality, custom design work before a contract is signed, using finished output to win deals rather than pitch decks. An advertising agency could create polished video ads without the time, cost, or logistics of a physical shoot. A law firm could generate legal documents in minutes instead of weeks, compressing timelines that once defined the business.
In each case, the business structure changes. These firms don’t behave like traditional agencies anymore. They start to look like software companies that happen to sell services — with repeatable processes, AI-driven production, and far higher margins.
YC’s bet is that AI doesn’t just improve agencies. It enables a new category of AI-native service companies that can scale far beyond today’s fragmented markets, because their growth is no longer tied directly to headcount.
For founders, this opens a different path. You don’t have to convince customers to adopt a new tool or change how they work. You deliver the outcome faster, cheaper, and better — and capture the value yourself.
What this signals:
YC believes some of the most profitable AI businesses won’t sell software at all. They’ll sell results, with AI quietly doing the work behind the scenes.
Stablecoin Financial Services
YC frames stablecoins as something bigger than a crypto trend. In its view, they’re becoming critical infrastructure for global finance, even as much of the financial services layer that should sit on top of them hasn’t yet been built.
What’s changed is regulation. The GENIUS and CLARITY Acts are pushing stablecoins into a rare middle position: compliant enough to work within traditional financial frameworks, yet still crypto-native in how they move, settle, and integrate with programmable systems. That regulatory positioning is new — and it matters.
YC argues this opens the door to an entire class of financial services that previously couldn’t exist. On one side, you have regulated financial products that feel safe but offer limited upside. On the other hand, unregulated crypto products promise higher returns but carry real risks. Stablecoins now sit between those worlds, creating room for services that combine the strengths of both.
That could mean yield-bearing accounts built on stablecoins, access to tokenized real-world assets, or infrastructure that enables money to move faster and more cheaply across borders without stepping outside compliance boundaries. Instead of forcing businesses and individuals to choose between safety and opportunity, stablecoins can bridge the gap.
YC’s emphasis here isn’t speculative. It’s structural. The rails are being laid right now, and the regulatory window is open. Founders who move early have the chance to define how money flows in this new middle ground before incumbents fully wake up to it.
What this signals:
YC believes the next wave of fintech won’t come from flashy consumer apps, but from quiet infrastructure that makes capital more efficient, compliant, and globally accessible — with stablecoins as the connective tissue.
AI for Government
YC points out a growing mismatch in how information flows.
The first wave of AI companies has dramatically improved how businesses and individuals fill out forms, submit applications, and generate documents. Speed and accuracy have increased on the front end. But on the back end, many of those same forms still end up in local, state, and federal agencies, where they’re printed and processed by hand.
That gap is widening.
As AI makes it easier for people to generate and submit paperwork, governments are about to face a surge of volume they’re structurally unprepared to handle. YC’s view is simple: the government desperately needs AI tools just to keep up. And if built correctly, those tools don’t just prevent collapse — they make public services far more cost-effective and responsive.
YC points to places like Estonia as early proof that digital government can work. But those examples remain isolated. The opportunity is to take that model and scale it across jurisdictions that still rely on fragmented systems and manual workflows.
YC is also clear-eyed about the difficulty. Building for government isn’t glamorous, and it isn’t easy. Sales cycles are long, procurement is complex, and early traction can be painfully slow. This category is not for the faint of heart.
But there’s a tradeoff. Once a startup successfully lands its first government customer, the relationship tends to be sticky. Systems become embedded. Contracts expand. What starts as a small deployment can grow into a large, long-term agreement.
YC’s message is that founders willing to endure the front-loaded friction have a chance to build companies with enormous scale and lasting impact.
What this signals:
YC believes AI’s next big productivity gains won’t just show up in private companies. They’ll show up in governments that finally modernize how they process information — quietly reshaping how public services operate.
Modern Metal Mills
YC challenges the common narrative around reindustrializing America.
Most discussions focus on labor costs or geopolitics. YC argues that the real problem is more fundamental: American metal mills are slow by design.
If you buy rolled aluminum or steel tube in the U.S., lead times of eight to thirty weeks are routine. Many buyers can’t even purchase directly from mills. And despite high prices, margins remain thin. YC is explicit that this isn’t due to weak demand or unskilled workers. It’s because the operating systems running these mills were designed decades ago.
Production planning, scheduling, quoting, and execution are fragmented across disconnected tools and processes. Mills optimize for tonnage and utilization, not speed, flexibility, or margin. Short production runs and specification changes are treated as disruptions rather than opportunities, even when customers are willing to pay for faster turnaround.
Automation hasn’t filled the gap. At the very moment the workforce is shrinking, critical functions such as material handling, changeovers, inspection, and quality control still depend on tribal knowledge held by a small number of experienced operators. Where automation exists, it’s largely used to increase throughput in inherently slow systems, rather than to eliminate setup time, variability, or coordination bottlenecks.
Energy compounds the problem. Aluminum and steel production are extremely energy-intensive, yet most mills operate on legacy power contracts tied to inflexible grids. YC points out that newer energy models — on-site generation, smarter power management, and even next-generation nuclear — could dramatically reduce costs. But energy is rarely designed into mill operations from the ground up.
What’s changed is that the enabling technology is finally ready.
AI-driven planning systems, real-time manufacturing execution systems, and modern automation can now be integrated into a single, coherent operating model. Instead of optimizing one part of the process at a time, software can coordinate planning, execution, and energy usage together — compressing lead times while improving margins.
YC believes this creates an opening to build modern, software-defined American mills, particularly in aluminum rolling and steel tube, where long lead times and energy costs are most entrenched.
The goal isn’t just speed. It’s making domestic metal production cheaper, more flexible, and more profitable — and in the process, rebuilding a core piece of the U.S. industrial foundation.
What this signals:
YC is betting that some of the most important startups of the next decade won’t look like tech companies at all. They’ll look like industrial operators rebuilt from first principles, with software and energy systems as the real competitive advantage.
AI Guidance for Physical Work
YC opens this idea with a striking image.
In The Matrix, Neo plugs a cable into the back of his head, wakes up, and calmly says, “I know kung fu.” YC isn’t talking about brain implants — but it believes physical work is about to get surprisingly close to something through real-time AI guidance.
Most AI debate centers on which desk jobs will be automated. YC thinks that misses the largest opportunity. In physical fields like field services, manufacturing, and healthcare, AI can’t yet act directly in the world. What it can do is see, reason, and guide the human who does.
YC’s vision is concrete. A worker wears a small camera while an AI system sees what they see and talks them through the task step by step:
-
“Turn off that valve.”
-
“Use the ⅜-inch wrench.”
-
“That part looks worn — replace it.”
Instead of months or years of training, workers become productive almost immediately, with AI acting as a constant coach and unlocking new skills on demand.
YC explains why this is suddenly feasible now. Three forces have converged. First, multimodal models have crossed a threshold at which they can reliably perceive and reason about real-world environments. Second, the hardware already exists at scale — smartphones, earbuds, and smart glasses. And third, skilled labor shortages have turned this into an economic necessity, not a novelty, with the potential to create higher-wage jobs for millions of people.
There are multiple paths YC sees founders taking. One is building the guidance system itself and selling it to companies with existing workforces. Another is choosing a specific vertical, such as HVAC repair or nursing, and building a full-stack, AI-augmented workforce from the ground up. A third is creating a platform that lets anyone sign up, receive AI guidance, and immediately start working or launch a small business.
YC frames this as giving physical workers the same kind of leverage that tools like Claude Code give software developers — not replacing them, but radically increasing what they’re capable of.
What this signals:
YC believes AI’s most profound economic impact may come not from replacing jobs, but from compressing skill and training timelines in the physical world — turning labor shortages into an opportunity rather than a constraint.
Large Spatial Models
YC draws a clear line between where AI has made progress and where it remains fundamentally limited.
Large language models have powered most of the recent breakthroughs in AI, but their impact has largely been confined to domains that can be expressed through text. Language is a powerful abstraction, but it breaks down when systems need to reason about space, geometry, and physical structure.
Today’s AI systems can handle narrow spatial tasks — identifying basic relationships, estimating depth, or recognizing objects in images. What they cannot reliably do is reason about spatial manipulation, 2D and 3D features, how those features relate to one another, or transformations such as mental rotation. As a result, AI still struggles to understand and interact meaningfully with the physical world.
YC argues that unlocking the next wave of AI capabilities—and moving closer to artificial general intelligence—requires a different foundation. The opportunity is to build large-scale spatial reasoning models that treat geometry and physical structure as first-class primitives, rather than approximations layered on top of language.
Such models wouldn’t just describe the world. They could reason about it. They could design real-world objects, understand physical environments, and manipulate spatial relationships in ways that today’s systems simply can’t.
YC is explicit about the stakes. A company that succeeds here wouldn’t just build another application layer. It could define the next generation of AI foundation models, on the scale of OpenAI or Anthropic.
This is a long-horizon bet — technically demanding, capital-intensive, and research-heavy — but the upside is correspondingly large.
What this signals:
YC believes the next true leap in AI won’t come from better prompts or bigger language models, but from systems that can reason about the physical world itself — turning space, structure, and geometry into native intelligence rather than afterthoughts.
Infra for Government Fraud Hunters
YC is explicit about the problem it wants founders to tackle: government fraud investigation is still operating in a pre-AI era.
The government is the largest customer on Earth, spending trillions of dollars every year across the federal, state, and local levels. It also loses staggering amounts to fraud. YC points to Medicare alone, which loses tens of billions annually to improper payments.
One of the most effective tools for recovering that money already exists. Under the qui tam provision of the False Claims Act, private citizens can file lawsuits on behalf of the government against companies that defraud public programs. If the case succeeds, the whistleblower receives a percentage of whatever is recovered.
The bottleneck isn’t law. It’s a process.
Today, the typical flow looks like this: an insider tips off a law firm, which then spends months or even years manually pulling documents, organizing evidence, tracing ownership structures, and assembling a case strong enough to file. The work is slow, expensive, and poorly matched to the scale of modern fraud.
YC argues this entire process should be accelerated with software — not dashboards or surface-level analytics, but intelligent systems that can do the heavy lifting. That means parsing messy PDFs, following opaque corporate structures across shell entities, correlating records, and packaging everything into complaint-ready filings.
Some startups have already begun filing False Claims Act cases themselves. YC sees a much larger opportunity in building tools that dramatically speed up the people already doing this work: whistleblower law firms, state attorneys general, and inspectors general.
The founder’s background is critical here. YC is clear that strong teams will include at least one founder who has lived this problem — former FCA counsel, compliance leaders, or auditors who understand where fraud hides and how cases are built.
Timing matters. YC believes the necessary AI capabilities have finally arrived, and there’s bipartisan political momentum to recover wasted public funds. A company that can make fraud recovery ten times faster wouldn’t just build a large business. It would return billions of dollars to taxpayers.
What this signals:
YC is actively encouraging founders to build high-leverage companies where success is measured not by user growth, but by recovered dollars — and where AI turns slow, manual civic processes into scalable systems.
Make LLMs Easy to Train
YC closes its request with a problem that’s far less visible than flashy AI demos, but deeply familiar to anyone actually training models.
Despite the explosion of interest and investment in AI, training large language models remains surprisingly painful. One YC partner describes spending the past three years training diffusion and language models at Can of Soup, only to find that the tooling has barely kept pace with the hype.
On a typical day, progress is slowed by broken SDKs, SSH sessions into GPU instances that turn out to be faulty only after half an hour of setup, and major bugs buried deep in open-source tooling. These issues aren’t edge cases. They’re routine friction.
Then there’s the data. Managing, sourcing, processing, and visualizing terabytes of training data is a job unto itself, often stitched together with fragile pipelines and one-off scripts. The complexity compounds quickly, especially as teams move beyond pretraining into fine-tuning, post-training, and specialization.
YC’s ask here is straightforward: make this easy.
That could take several forms. APIs that abstract away the mechanics of training. Databases are purpose-built to manage massive datasets cleanly and reliably. Developer environments designed specifically for ML research, rather than retrofitted from general software tooling.
The deeper point YC is making is about timing. As post-training and model specialization become more important than training giant models from scratch, the bottleneck shifts. Teams that can iterate quickly, manage data sanely, and avoid infrastructure landmines gain a decisive advantage.
YC sees these tools not as supporting infrastructure, but as potential foundations. The companies that make model training simple and reliable could end up shaping how future software is built, just as cloud platforms and dev frameworks did in earlier eras.
What this signals:
YC believes the next wave of AI leverage won’t come from bigger models alone, but from tooling that lets smaller teams train, adapt, and specialize models without burning time on infrastructure chaos.
Summary of YC request for startups 2026
1) Cursor for product managers
-
Opportunity to help teams figure out what to build, not just how to build it
-
Upload customer interviews and product usage data, then ask, “What should we build next?”
-
The way we define and communicate “what to build” needs to change
2) AI-native hedge funds
-
The next Renaissance, Bridgewater, and D.E. Shaw are going to be built on AI
-
The biggest funds in the world have been slow to adapt
-
Swarms of agents will do what hedge fund traders do now
3) AI-native agencies
-
Agencies have always been crazy hard to scale
-
Now, instead of selling software to customers to help them work, you can use the software
-
You can charge way more by selling them the finished product at 100× the price
-
Agencies of the future will look more like software companies, with margin and scale
4) Stablecoin financial services
-
Much of the financial services layer remains unbuilt
-
Room for financial services that offer DeFi benefits like better yield or tokenized assets
-
The regulatory window is open, the rails are being laid, perfect time to build
5) AI for government
-
People now fill in forms and applications with unprecedented speed and accuracy
-
The government needs AI tools to deal with the huge increase coming down the line
-
Selling to the government is hard, but very sticky, and can expand to huge contracts
6) Modern metal mills
-
American metal mills are slow by design; systems were designed decades ago
-
Automation has lagged at the exact moment the workforce is shrinking
-
Software and energy technology are good enough to rethink the entire system
7) AI guidance for physical work
-
Real-time AI guidance that can see, reason, and guide the human who does the work
-
Imagine wearing a small camera while an AI talks you through the job
-
Three things have converged: multimodal models, hardware everywhere, and skilled labor shortages
8) Large spatial models
-
Large language models are limited to domains expressible in language, leaving physical and spatial reasoning largely unsolved
-
Today’s systems struggle with robust 2D/3D reasoning, spatial manipulation, and understanding physical structure
-
Opportunity to build foundation models where geometry and spatial relationships are first-class primitives
-
Such models could unlock real-world design, simulation, and interaction—and define the next major AI platform
9) Infrastructure for government fraud hunters
-
The government loses tens of billions annually to fraud, with recovery processes that are slow and manual
-
Whistleblower and FCA cases rely on months or years of document review, corporate tracing, and evidence assembly
-
Opportunity to build intelligent systems that turn insider tips into complaint-ready cases automatically
-
Strong founder–market fit matters; teams with legal, compliance, or audit backgrounds have a major edge
10) Make LLMs easy to train
-
Training large models remains brittle, time-consuming, and poorly tooled despite massive AI investment
-
Teams still struggle with broken SDKs, unstable GPU infrastructure, and fragmented open-source tooling
-
Opportunity for APIs, data systems, and dev environments that abstract and simplify model training
-
As post-training and specialization grow, these tools could become core infrastructure for future software
What it means for tech founders
Taken together, YC’s 2026 request sends a clear message:
-
AI is shifting from tools to decision-makers
-
Finance, government, and industry are finally fair game for venture-scale startups
-
The next big companies will blend software, regulation, energy, and human labor
-
Deep, unglamorous problems are back on the table
For founders, this isn’t a checklist — it’s a signal. The frontier is widening, and YC is actively looking for teams willing to build where complexity, regulation, and the real world collide.
If anything, the 2026 request makes one thing clear: the next wave of startups won’t just live on the internet. They’ll reshape how the world actually works.
Closing: The Bet YC Is Really Making
Taken together, YC’s 2026 Request for Startups points to a quiet but profound shift.
The next wave of startups won’t be defined by apps, growth hacks, or surface-level automation. They’ll be defined by systems thinking—companies built to operate within complexity rather than avoid it. AI isn’t being positioned as a feature, but as an operator: deciding what to build, guiding human work, moving capital, uncovering fraud, modernizing industry, and reasoning about the physical world itself.
There’s also a noticeable change in posture. YC is no longer signaling that speed alone wins. It signals that depth wins—domain expertise, regulatory fluency, operational credibility, and a willingness to tackle problems that don’t yield overnight traction.
For founders, the message is both demanding and liberating. You don’t need to chase the obvious idea or the loudest market. You need to understand a system that’s broken, believe it can be rebuilt, and use AI as the leverage to do what was previously impractical.
That’s the real subtext of YC’s 2026 request.
The frontier hasn’t narrowed.
It’s widened — into the real world.
Trending Now
