YC's Latest Startup Barometer: AI Is Moving from Models to the Real Economy—These 11 Frontier Sectors Are Worth the Bet

Author: Y Combinator team (Garry Tan, Brad Flora, Diana Hu, Jared Friedman, Gustaf Alströmer, Ankit Gupta, Jon Xu, and others)

Translated by TechFlow

TechFlow Editor's Note:Throughout the evolution of Silicon Valley and the global venture ecosystem, Y Combinator's Request for Startups (RFS) has long been regarded as both a barometer and the eye of the storm at the frontier of technology waves. Looking back, from early cloud computing, mobile internet, and Stripe-style financial infrastructure to later vertical SaaS, YC's RFS has often pointed entrepreneurs and capital precisely to untapped areas that are not yet monopolized by incumbents but are in urgent need of technological reinvention.

In this latest RFS, YC makes a clear and consequential macro judgment: the biggest bottleneck in AI development is no longer confined to the parameters and algorithms of large foundation models themselves, but is accelerating into the real world. Thin software-layer wrappers are losing their premium, while truly disruptive opportunities are shifting comprehensively into capital- and physics-intensive deep waters such as physical manufacturing, energy and grid infrastructure, siloed public administration, high-security isolation systems, and commercial space.

This long-form article summarizes 11 frontier directions on which YC's core partner team is placing key bets. It covers not only $15-per-hour democratized AI experts and writing large models with genuinely literary-quality prose, but also domestic robot manufacturing in the United States, data center power and cooling supply chains, Wall Street financial instruments for trillion-dollar infrastructure, and enterprise-exclusive private model systems. TechFlow has provided a complete, in-depth translation, offering entrepreneurs, researchers, and investors at the front lines of the AI wave a rare industry landscape map and navigational guide.

Publishing the Request for Startups (RFS) has long been a proud tradition at Y Combinator, aimed at sharing with entrepreneurs around the world the key problems we most want to see solved.

However, these directions represent only a small portion of our investment landscape. If one of these topics ignites your passion, consider it a strong endorsement to pursue it; but even if your startup is not among them, you are still welcome to apply to YC.

A major trend today is that AI's biggest bottleneck is rapidly spreading beyond the model layer itself.

We are excited about the next wave of startups. They are building the hardware, infrastructure, financial systems, and specialized vertical software needed to bring frontier intelligence into the real economy, across a broad range of areas from factory manufacturing and data centers to public administration and even outer space.

Startups for the Era of Cheap Space Launches

Written by:Ankit Gupta (YC Group Partner)

SpaceX is painting a clear near-future picture: in the near term, Starship will reach 10000 launches per year.

Today, even a non-recoverable, expendable Starship already has launch costs nearly comparable to the reusable Falcon 9, with cost per kilogram to orbit significantly lower.

As Starship technology scales and achieves full reusability, we will see launch cost per kilogram of payload drop by several orders of magnitude.

This means now is the golden moment to systematically think about startup opportunities in an era of abundant space access.

Which business models are today constrained by launch costs and difficult to make viable, but would become highly feasible once launch costs fall by 100 times or more and the orbital economy fully takes off?

We believe companies of this kind should be founded now. If you are a builder eager to explore this space, YC would love to receive your application.

$15-per-Hour Experts: AI-Driven Accessible Professional Services

Written by:Brad Flora (YC Group Partner)

Most discussions about AI today focus on how it will replace existing workers. What truly excites us, however, is the opposite path: using AI to turn an ordinary worker directly into a professional-level expert.

We are witnessing more and more companies moving beyond degrees and credentials to hire ordinary people with exceptional insight and execution. Fully equipped with advanced AI, they can handle complex work that previously required senior experts.

Imagine a customer service representative with excellent communication skills and deep AI proficiency suddenly able to advise on complex M&A, supply chain optimization, or global trade compliance.

In fields such as public benefits, tax planning, insurance claims, medical billing, and immigration and visa paperwork, there are millions of similar jobs.

In short, wherever professional credentialing creates a bottleneck to service supply, this model can excel.

The previous industrial revolution created a vast modern middle class through factory blue-collar jobs; this technological transformation is poised to reshape a new middle-class system through democratized expert roles.

We want to fully fund startups committed to realizing this vision.

Made-in-America Robots: Hardware Comes Home

Written by:Diana Hu (YC Group Partner)

Today, leading robot foundation models are beginning to achieve general skill transfer across embodiments and robot types. In my view, the competitive bottleneck in robotics is rapidly shifting to physical hardware.

In 2025, 38400 industrial robots were installed in the United States, a 12% year-over-year increase. But most robots still rely heavily on imports from Japan and Europe.

Even robots assembled in the United States often depend deeply on Chinese-made magnets at their core.

In 2024, China produced 94% of the world's sintered rare-earth permanent magnets. When export restrictions were introduced in April 2025, some major automakers were even forced to directly cut production capacity.

By July, the U.S. Federal Communications Commission (FCC) had also suspended certification for new connected mobile robots unless they were entirely made in the United States or received a very limited number of exemptions.

Under current compliance standards, more than 65% of a robot's component costs must come from domestic manufacturing, and this share will rise to 75% by 2029. This means that the route of simply importing components for domestic assembly under a local brand is no longer viable.

We want to fully support entrepreneurs who develop robots and their complete upstream supply chains in the United States.

This covers rare-earth magnets, motors, harmonic drive reducers (Strain-Wave Gears), encoders, torque sensors, end grippers, specialized chips, battery packs, and the high-precision machine tools and production equipment used to manufacture these components.

If you are deeply engaged in this field, we would love to have an in-depth conversation with you.

Large Language Models with Truly High-Quality Writing and Deep Composition Ability

Written by:Jared Friedman (YC Group Partner)

As large language models have advanced rapidly in programming and mathematical reasoning, their writing quality and expressive prose have quietly declined.

This is not your imagination: compared with models from 2024 and 2025, today's mainstream large models are indeed more rigid and mechanical in tone.

This does not appear to be a sporadic, temporary technical flaw, but a deep structural problem.

The reinforcement-learning reward signals (Reward Signals) designed to give models top-tier programming ability actively flatten tone, emotional variation, and distinctive writing style during actual training.

When forced to make R&D resource trade-offs, leading AI labs are undoubtedly prioritizing compute for code generation and Agentic Workflows, while abandoning rigorous attention to prose texture and writing style.

This leaves a huge market gap for startups: to build a distinctive model that people genuinely love to converse with and that writes with warmth and depth.

If you want to make breakthroughs in this direction, we would love to hear from you.

The “Plaid” for Government Benefits and Public Assistance: Breaking Down Administrative Data Silos

Written by:Gustaf Alströmer (YC Group Partner)

If you have ever applied for public benefits from a government agency, you have likely grown to hate this cumbersome process: providing proof of income, verifying household members, and submitting proof of residency. Then, when moving to the next assistance program, you must resubmit all the documents; at annual eligibility reviews, the entire ordeal repeats itself.

Each government department has its own separate portal, with no interoperability among them and data completely siloed.

Numerous applicants are denied outright simply because they omitted one nonessential document, while public servants are forced to manually review pay stubs by eye day after day.

Ironically, this bureaucratic system that shuts out honest citizens is riddled with loopholes in fraud prevention.

The U.S. Government Accountability Office (GAO) estimates that benefit fraud costs the federal government nearly $500 billion per year. To make matters worse, with generative AI today, a convincing fake pay stub can be forged in seconds.

Before the fintech infrastructure company Plaid emerged, every financial app had to go through painstaking efforts to connect with each bank individually; after Plaid arrived, users could connect seamlessly with a single click. The public sector urgently needs its own infrastructure layer.

With user authorization, it can directly retrieve verified real income, tax, and residency records from the most authoritative official sources and distribute them synchronously to government approval systems.

'Verify once, use everywhere.' This can not only help trustworthy citizens who truly need assistance get approved extremely quickly, but also eliminate fraudsters' ability to forge documents at the source.

Starting in 2027, states across the United States will be required to reverify eligibility for millions of Medicaid (medical assistance program) enrollees every 6 months.

In the near future, AI agents will take over government applications and eligibility determinations, but this is contingent on AI having fully trusted, tamper-proof underlying real data sources.

Whoever builds this government-wide data pipeline will effectively control the core digital infrastructure of the modern social safety net. If you are building this critical pipeline, we would welcome the opportunity to work with you.

The 'Wall Street' of the Physical Infrastructure Buildout Era: New Capital Operations and Financing Structures

Written by:Brad Flora (YC Group Partner)

The startup cost for developing software is very low, but machinery, GPU compute clusters, and hyperscale data centers are entirely different.

Someone must pay real cash on day one of a project for thousands of tons of specialty steel and high-end silicon-based chips. In today's financial markets, mature financing instruments for these assets are virtually nonexistent.

When underwriting credit, traditional commercial banks still assess depreciation for expensive advanced intelligent robots as if they were ordinary agricultural forklifts.

For this reason, a group of sharp entrepreneurs working with YC has already begun personally inventing entirely new financial derivatives. For example, they are creating new credit lines specifically for robots that generate stable annual income of $500,000, or designing bespoke put option contracts for future GPU compute capacity.

Looking back, every great physical-world infrastructure buildout—whether transcontinental railroads, national power grids, or modern communications infrastructure across cities and rural areas—gave rise to financial giants of comparable scale as the physical assets were built.

We are looking for financial-innovation founders who can build dedicated lending institutions, asset leasing platforms, asset-backed securities (ABS) structures, and risk-hedging tools for the physical AI economy.

If you deeply understand both the physical and industrial principles of real-world machines and the underlying logic of capital operations, we would welcome an in-depth conversation with you.

Vertical-Domain Foundation Large Models

Written by:Diana Hu (YC Group Partner)

I believe Jev precisely addresses industry pain points because it provides developers with strongly typed structured decisions and explicit confidence probabilities, rather than open-ended generative text, while completely eliminating the heavy burden of training separate classifiers for each narrow task.

I would like to see more frontier research teams build specialized vertical foundation models tailored to physical industries such as manufacturing, energy and power, and life sciences. These models should be able to operate accurately with zero-shot (Zero-Shot) or only a few examples.

The bar for this track is extremely high: it must match or even surpass the business accuracy of general-purpose large models at inference compute costs several orders of magnitude lower, enabling these models to be seamlessly deployed on all endpoint devices at extremely low cost.

Google’s TimesFM-3 is a highly representative example of time-series forecasting.

This compact model has only 330 million parameters (330M). Without task-specific fine-tuning, it can generate complete forecasts for multiple related time-series datasets in a single forward pass. The model only needs to be trained once; each new user can call it directly by supplying their own historical data as context.

The real challenge for vertical large models lies in obtaining high-quality, industry-specific reward signals that can be rigorously verified mathematically (Verifiable Rewards).

Startup teams must not only have legitimate data ownership, but also demonstrate an overwhelming competitive advantage over existing expert models in forecast accuracy and compute consumption.

If you have sufficiently deep knowledge of a specific vertical real-world industry and can build a dedicated model for it, we would love to hear your proposal.

The AI Revolution for City Halls and Local Governments

Written by:Gustaf Alströmer (YC Group Partner)

In the United States, if you plan to build a house or open a store, the first step is to obtain a permit from the local city government.

Endless application forms, lengthy construction drawing reviews, and cumbersome site inspections often mean waiting for months or even longer. In some cities, the time spent waiting for a permit to be approved is longer than the actual construction schedule.

Public employees at city halls are buried in large volumes of highly repetitive administrative work: reviewing engineering drawings line by line against detailed building codes, comparing planned uses with zoning regulations word by word, and repeatedly exchanging revision comments among departments.

Modern AI models are fully capable of completing all of this independently and efficiently.

Lengthy delays in administrative approvals carry an extremely heavy economic cost. Every additional month of waiting causes development costs to surge, and many planned projects are abandoned midway as a result.

Keep in mind that every newly constructed building will generate property tax revenue for the city for decades to come.

Our rough calculation shows that if the processing time for all building permits across the United States were reduced to 10% of the current duration, U.S. cities would collect an additional $16 billion in tax revenue each year.

We are eager to fund technology companies dedicated to using AI to empower and reshape the day-to-day operations of city halls. The entry point can be building and construction permits, then gradually expand into broader areas such as 311 municipal service hotlines, public records management, business licensing, and on-site safety inspections.

The United States has more than 19000 municipalities at various levels, and the vast majority still rely entirely on extremely primitive manual processes to handle all of this.

A traditional belief has long circulated in tech venture capital: “Never sell software to government agencies.”

This belief may have held true in the past, but it is now completely outdated. YC has invested in several startups that sell AI solutions specifically to government agencies, and some of these teams have grown faster than almost any traditional enterprise SaaS company we have seen.

If you are building in this space, we would love to receive your message.

AI Data Center Supply Chain Innovation: Power, Cooling, and Grid Connection

Written by:Jared Friedman (YC Group Partner)

The massive expansion of AI data centers currently underway across the United States is the largest single capital expenditure project in U.S. business history.

Its investment scale has surpassed the combined totals of the transcontinental railroads, the interstate highway system, and the Manhattan Project.

A construction effort of unprecedented scale is generating immeasurable business opportunities throughout the entire supply chain, from top to bottom.

Some segments of this market are already well known: demand for advanced GPU chips, basic power supply, and highly skilled electricians has reached unprecedentedly intense levels.

However, outside the spotlight, there are many little-known hidden bottlenecks that are enough to stall the progress of the entire industry.

Currently, the scheduled delivery lead time for large high-voltage power transformers has reached a full 3 years; high- and low-voltage distribution switchgear, high-voltage circuit breakers, and advanced liquid cooling system equipment have been sold out in the market.

Even the most upstream core raw materials are flashing shortage alerts: grain-oriented electrical steel (Grain-Oriented Electrical Steel), essential for transformer cores, can now be produced by only one U.S. factory, and its capacity has long been severely overextended.

We are extremely keen to invest in startups that comprehensively empower and break through the entire upstream and downstream supply chain for data centers. If you are tackling this challenge, we would love to receive your application.

AI in the Air-Gapped Economy: Defense, Classified, and High-Security Environments

Written by:Brad Flora (YC Group Partner)

The world's most valuable trade secrets and strategic data have never been stored in the public cloud.

Defense contractors, major commercial banks, top medical institutions, and nuclear power plants all keep their critical data locked behind physically isolated walls. Out of the strictest security and compliance considerations, they would never upload such data to any third-party AI API.

This is why enormous trillion-dollar pillar industries in the real economy are still lingering outside the AI wave.

Many founders we work with repeatedly hit the same wall: target customers praise the powerful capabilities of AI products, but their only requirement is that the entire system must be deployed 100% inside their own facilities and must never connect to the public internet.

We want to fully fund technology pioneers capable of bringing cutting-edge AI technology truly inside these air-gapped walls.

This includes fully private on-premises large models, full-stack deployment and operations toolchains, authoritative security and compliance certification frameworks, and highly customized specialized hardware infrastructure.

Sales and go-to-market cycles in this closed market are indeed slower, and the barriers to entry are extremely high.

But the other side of the coin is that the industry's richest procurement budgets are concentrated here, and the companies with the most durable moats will inevitably be born here. If you are attacking this fortress, we would love to have a deep, extended conversation with you.

Take Control of Your Own Intelligence: Private and Open-Source Sovereign Models

Written by:Jon Xu (YC Visiting Partner)

Today, almost every company building products with AI is essentially renting intelligence from external providers.

Your systems continuously send Token data to a handful of leading frontier AI labs, receive generated outputs, and then pay expensive monthly API usage bills.

In the past, this trade seemed reasonable: top closed-source models were years ahead of the open-source ecosystem, and no amount of work on your proprietary data could close that generational capability gap.

However, the tides have fundamentally shifted.

Open-weight models now have the proven capability to compete head-to-head with the most advanced closed frontier models on many core benchmarks with significant practical business value.

As AI increasingly becomes the central nervous system of business operations, more and more companies are becoming deeply wary of heavy reliance on externally rented model services.

We are convinced that more and more companies will want to truly own the core models on which their survival depends: continuously refining them with their proprietary data, running them on infrastructure of their own choosing, and retaining absolute control over when those models are upgraded or changed.

This is especially critical in heavily regulated industries. Under extremely strict data sovereignty compliance requirements and physically locked network architectures, the hidden cost of continuing to rent external models is grossly disproportionate to actual business output.

However, there remains a vast engineering and conceptual gap between downloading a public model from the open-source community and enabling an enterprise to truly possess a dedicated intelligent system that can evolve autonomously.

Specialized expertise is needed to help enterprises translate complex business workflows into precise evaluation benchmarks (Evals), distill employees' tacit business expertise into high-quality training corpora, turn occasional production incidents into fuel for continuous model iteration, and seamlessly carry forward these core assets whenever the underlying base model is upgraded.

For the most extreme and complex general unknown problems, enterprises may still choose to continue renting public compute from top frontier labs; but we firmly believe that the vast majority of enterprises must have their own vertical, domain-specific models.

These models will fully encode the enterprise's own business operating mechanics and, as the business develops, remain ever more relevant and increasingly possess unmatched proprietary value.

If you are helping enterprises across industries truly take control of and own their dedicated intelligence, we would love to hear from you.

Source: 深潮 TechFlow