YC's latest entrepreneurial trend indicator: AI is moving from models to the real economy, and these 11 cutting-edge tracks are the most worth betting on

Author: Y Combinator Team (Garry Tan, Brad Flora, Diana Hu, Jared Friedman, Gustaf Alströmer, Ankit Gupta, Jon Xu, etc.)

Compiled by: Deep Tide TechFlow

Deep Tide Introduction: In the evolution of the venture capital ecosystem in Silicon Valley and globally, Y Combinator's "Request for Startups" (RFS) has always been regarded as a barometer and eye of the storm at the forefront of the tech wave. Looking back at history, from the early days of cloud computing, mobile internet, and Stripe-like financial infrastructure to the later vertical SaaS, YC's RFS often points out virgin territories that have not yet been monopolized by giants but urgently need technological reconstruction, right before a qualitative change in technology.

In this latest RFS, YC has provided a very clear and significant macro judgment: the biggest bottleneck in AI development is no longer just the parameters and algorithms of the foundational large models themselves, but is accelerating towards the real world. Shallow packaging at the pure software level is losing its premium, and real disruptive opportunities are fully migrating to deep waters of capital-intensive, physically constrained areas such as manufacturing, energy and grid infrastructure, public administration administrative islands, high-security isolation systems, and commercial space.

This lengthy article summarizes the 11 cutting-edge directions that the core partner team at YC is focusing on, covering everything from $15 hourly AI experts and writing large models with truly literary-level writing skills to American-made robotics, data center power cooling supply chains, trillion-dollar infrastructure Wall Street financial instruments, and enterprise-specific privatized model systems. Deep Tide TechFlow has conducted a complete in-depth compilation of this, providing an invaluable industrial panorama and navigation guide for every entrepreneur, researcher, and investor at the forefront of the AI wave.

Publishing the "Request for Startups" (RFS) has always been a long-standing tradition at Y Combinator, aimed at sharing with global entrepreneurs the key propositions we most desire to see tackled.

However, these directions only represent a small part of our investment landscape. If any of these topics ignite your passion, consider it a strong endorsement to dive in; but even if your startup project is not listed here, it does not hinder you from submitting an application to YC.

A significant current trend is that the biggest bottleneck facing AI is rapidly spreading from the model layer itself outward.

We are very excited about the new wave of startups. They are working hard to build the necessary hardware, infrastructure, financial systems, and dedicated vertical software to truly inject cutting-edge intelligence into the real economy, covering a wide range of fields from factory manufacturing and data centers to government governance and even outer space.

Startups in the Era of Cheap Space Launches

Written by: Ankit Gupta (YC Group Partner)

SpaceX is painting a clear picture of the near future: in the not-so-distant future, Starship will achieve a launch frequency of 10,000 times per year.

Currently, even a non-reusable disposable Starship has launch costs that are almost on par with the reusable Falcon 9, with the cost per kilogram to orbit significantly reduced.

As Starship technology scales and achieves full reusability, we will witness a several orders of magnitude drop in the launch cost per kilogram of payload.

This means that now is the golden moment to think systemically about entrepreneurial opportunities in the "abundant space access era."

What business models today may be constrained by launch costs but will become highly feasible once launch costs plummet by 100 times or more, and the orbital economy fully explodes?

We believe that such companies must start now. If you are a builder eager to explore this, YC is very eager to receive your application.

$15 Hourly Experts: AI-Driven Inclusive Professional Services

Written by: Brad Flora (YC Group Partner)

Regarding AI, most discussions today focus on how it will replace existing workers. However, what truly excites us is the completely opposite path: using AI to directly transform an ordinary worker into a professional-level expert.

We are witnessing more and more companies breaking free from the constraints of diplomas and qualifications, recruiting ordinary people with high insight and execution. Armed with advanced AI, they can handle complex tasks that only senior experts could manage in the past.

Imagine a customer service representative with excellent communication skills who is also highly skilled in using AI suddenly gaining the ability to provide consulting advice on complex mergers and acquisitions (M&A), supply chain optimization, or global trade compliance.

In fields such as public welfare, tax planning, insurance claims, medical billing, and immigration visa documentation, there are millions of similar job positions.

In short, as long as professional qualification certification constitutes a bottleneck in service supply, this model can shine.

The last industrial revolution created a large modern middle class through blue-collar factory jobs; this time, the technological transformation is expected to reshape a new middle class system through inclusive expert positions.

We hope to fully fund those startups dedicated to realizing this vision.

American-Made Robots: Hardware Returning to the Homeland

Written by: Diana Hu (YC Group Partner)

Today, top robotic foundational models have begun to achieve cross-form and cross-model general skill transfer. In my view, the competitive bottleneck in the robotics field is rapidly shifting towards physical hardware.

In 2025, 38,400 industrial robots were installed in the U.S., a year-on-year increase of 12%. However, the vast majority of robots still rely heavily on imports from Japan and Europe.

Even those robots assembled in the U.S. often deeply depend on magnets sourced from China.

In 2024, China produced 94% of the world's sintered rare earth permanent magnets. When export restriction policies were introduced in April 2025, some automotive manufacturing giants were even forced to directly cut production capacity.

By July, the Federal Communications Commission (FCC) suspended the entry certification for new connected mobile robots unless these robots were entirely manufactured in the U.S. or obtained very few exemption permits.

According to current compliance standards, a robot must have over 65% of its component costs sourced from domestic manufacturing, and this ratio will rise to 75% by 2029. This means that simply importing components for local assembly is no longer a viable path.

We hope to fully support entrepreneurs who are independently developing robots and their complete upstream supply chains in the U.S.

This includes rare earth magnets, motors, strain-wave gears, encoders, torque sensors, end effectors, dedicated chips, battery packs, and high-precision industrial mother machines and production equipment used to manufacture these components.

If you are deeply engaged in this field, we look forward to in-depth discussions with you.

Truly High-Level Writing and Depth Writing Capabilities in Large Language Models

Written by: Jared Friedman (YC Group Partner)

As large language models have made rapid advances in programming and mathematical reasoning capabilities, they have quietly regressed in writing and expression.

This is not your illusion: compared to models from 2024 and 2025, today's mainstream large models do indeed appear more rigid and mechanical in tone.

This does not seem to be an occasional temporary technical flaw, but rather a deep structural issue.

The reinforcement learning reward signals set up to give the models top programming capabilities actively flatten the model's tonal color, emotional fluctuations, and unique writing style during actual training.

When faced with trade-offs in R&D resources, leading AI labs will undoubtedly prioritize computational power for code generation and agentic workflows, sacrificing the demands for textual quality and writing style.

This precisely leaves a huge market vacuum for startups: to create a model that people genuinely love to interact with, characterized by warmth and depth in writing.

If you are eager to make breakthroughs in this direction, we very much look forward to receiving your message.

"Plaid" for Government Welfare and Public Assistance: Breaking Down Administrative Data Silos

Written by: Gustaf Alströmer (YC Group Partner)

If you have ever applied for public welfare from government agencies, you must have detested this cumbersome process: obtaining income verification, verifying household members, providing proof of residence. Then, when switching to the next assistance program, you have to resubmit all the materials; when it comes time for annual eligibility review, the whole ordeal has to be repeated.

Every government administrative department has its own independent portal, with no connectivity between them, resulting in data being completely siloed.

Countless applicants are directly rejected simply for missing a non-critical document, while public officials have to rely on manual verification of pay stubs day after day.

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

The U.S. Government Accountability Office (GAO) estimates that various forms of welfare fraud cost the U.S. federal government nearly $500 billion annually. To make matters worse, generative AI can now produce a fake pay stub that looks convincingly real in just seconds.

Before the financial technology infrastructure Plaid was born, every financial application had to go through great lengths to connect with various banks; after Plaid emerged, users could complete seamless connections with just a click. The public administration sector urgently needs its own layer of this infrastructure.

With user authorization, it can directly pull verified real income, tax, and residence records from the most authoritative official sources and synchronize them to various government approval systems.

"One verification, universal access." This not only allows truly needy and trustworthy citizens to obtain approvals at lightning speed but also completely cuts off the space for fraudsters to fabricate materials at the source.

Starting in 2027, all states in the U.S. will be required to re-verify the eligibility of millions of Medicaid beneficiaries every six months.

In the near future, AI agents will fully take over government applications and eligibility approval work, but the prerequisite for all this is that AI must have completely trustworthy, tamper-proof underlying real data sources.

Whoever builds this data pipeline connecting government affairs will substantially grasp the core digital infrastructure of modern social security networks. If you are building this critical pipeline, we look forward to working with you.

"Wall Street" in the Era of Major Infrastructure Construction: New Capital Operations and Financing Structures

Written by: Brad Flora (YC Group Partner)

The startup costs for developing software are very low, but the costs for machinery, GPU computing clusters, and large-scale data centers are entirely different.

There must be someone who pays real money for thousands of tons of specialty steel and high-precision silicon-based chips on the very first day of the project launch. However, in the current financial market, there is almost a blank slate for mature financing tools targeting these types of assets.

Traditional commercial banks still assess the depreciation of expensive advanced intelligent robots as if they were ordinary agricultural forklifts during credit approval.

Because of this, a group of sharp entrepreneurs collaborating with YC has begun to invent entirely new financial derivatives. For example, providing new types of credit for robots that can generate a stable income of $500,000 annually, or designing exclusive put option contracts for future GPU computing capacity.

Looking back at history, every great infrastructure project in the physical world, whether it be the transcontinental railway network, the national power grid, or the modern communication infrastructure spread across urban and rural areas, has given birth to financial giants of equal scale as they rose from the ground.

We are looking for financial innovative 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 have a deep understanding of the physical and industrial laws of tangible machines, as well as the underlying logic of capital operations, we look forward to having an in-depth conversation with you.

Vertical Domain Foundation Models

Author: Diana Hu (YC Group Partner)

I believe that Jev can accurately hit the industry's pain points because it directly outputs strongly typed structured decisions and clear confidence probabilities to developers, rather than generating endless text, while completely eliminating the heavy burden of training classifiers separately for each sub-task.

I hope to see more cutting-edge R&D teams create dedicated vertical foundation models for industries such as manufacturing, energy, and life sciences. These models should be able to operate accurately under zero-shot conditions or with only a few examples.

The measurement criteria for this track are very strict: it must achieve or exceed the business accuracy of general large models at a reasoning power cost several orders of magnitude lower, allowing these models to be seamlessly deployed to all terminal devices at a very low cost.

Google's TimesFM-3 is a highly representative case of time series prediction.

This compact model, with only 330 million parameters, can make complete predictions for multiple related time series data in a single forward pass without any specific task fine-tuning. The model only needs to be trained once, and each new user can directly call it by providing their historical data as context.

The real challenge of vertical large models lies in obtaining high-quality, industry-specific proprietary reward signals that can be mathematically verified.

Startup teams not only need to have legal data ownership but also need to demonstrate overwhelming competitive advantages over existing expert models in terms of prediction accuracy and power consumption.

If you have a deep understanding of a specific vertical industry and the ability to build its dedicated model, we look forward to hearing your proposal.

The AI Revolution in City Halls and Local Governments

Author: 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 administrative permission from the local city government.

Endless application forms, lengthy construction drawing reviews, and tedious on-site inspections often require months or even longer of endless waiting. In some cities, the waiting time for permit approval can even be longer than the actual construction period.

City hall officials are buried under a mountain of highly repetitive administrative tasks: meticulously reviewing engineering drawings against strict building code regulations, comparing planning uses and zoning regulations word by word, and repeatedly sending and receiving modification opinions between various departments.

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

The long delays in administrative approvals bring extremely heavy economic costs. For every additional month of waiting, development costs soar, and many planned projects may even be aborted midway.

It is important to note that every new building that rises will continuously contribute property tax revenue to the city for decades to come.

We roughly calculated an economic account: if the approval time for all building permits in the U.S. could be reduced to 10% of the current time, cities across America would gain an additional $16 billion in fiscal revenue each year.

We are eager to fund technology companies dedicated to empowering and reshaping the daily operations of city halls with AI. The entry point could be building project permits, gradually extending to the 311 municipal hotline, public records management, business license issuance, and on-site safety inspections.

There are over 19,000 municipalities of various levels across the U.S., the vast majority of which still rely entirely on extremely primitive manual processes to handle all of this.

A traditional belief in the tech venture capital community has been: "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 departments, some of which have seen business growth rates that surpass any traditional enterprise-level SaaS we have ever seen.

If you are expanding in this field, we are very eager to hear from you.

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

Author: Jared Friedman (YC Group Partner)

The large-scale expansion of AI data centers being rolled out across the U.S. is the largest single capital expenditure project in American business history.

The scale of investment has already surpassed the total of the transcontinental railroads, interstate highway networks, and the Manhattan Project combined.

Such an unprecedented scale of construction is generating immeasurable business opportunities throughout the entire industry chain from top to bottom.

Some of the tracks are already well-known: the market demand for cutting-edge computing GPU chips, basic power supply, and highly skilled electricians has reached an unprecedented level of intensity.

However, hidden bottlenecks that are little known but capable of stalling the entire industry’s progress lurk in the shadows.

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

Even the core raw materials at the upstream level have sounded the alarm for shortages: the grain-oriented electrical steel essential for manufacturing transformer cores is only produced by one factory in the U.S., and its production capacity has long been severely overdrawn.

We are extremely eager to invest in startups that can empower and break through the entire upstream and downstream supply chain for data centers. If you are engaged in this battle, we look forward to receiving your application.

AI in a Physically Isolated Economic System: Defense, Confidentiality, and High-Security Environments

Author: Brad Flora (YC Group Partner)

The most valuable business secrets and strategic data in the world are never stored in public clouds.

Defense contractors, large commercial banks, top medical institutions, and nuclear power plants keep their vital data locked away behind physically isolated high walls. Due to the strictest confidentiality and compliance considerations, they would never upload this data to any third-party AI API interface.

This is why there remains such a vast trillion-dollar pillar industry in the real economy that is still struggling outside the door of the AI wave.

Many founders we work with repeatedly hit the same wall: target customers rave about the powerful capabilities of AI products, but the only requirement is that the entire system must be 100% deployed within their internal buildings and must not connect to any public network.

We hope to fully fund those who have the ability to truly bring cutting-edge AI technology into these physically isolated high walls.

This includes purely privatized local deployment of large models, full-stack deployment and operation toolchains, authoritative security compliance certification systems, and highly customized special hardware infrastructure.

The sales and expansion cycle for this closed market is indeed slower, and the barriers are extremely high.

But the flip side of the coin is that there is the most abundant procurement budget in the entire industry, which will inevitably give birth to great companies with the most unbreakable moats. If you are tackling this high wall, we look forward to having an in-depth conversation with you.

Take Control of Your Exclusive Intelligence: Privatized and Open Source Sovereign Models

Author: Jon Xu (YC Visiting Partner)

Today, almost all companies building products based on AI are essentially "renting" intelligence from external sources.

Your system continuously sends token data to several top cutting-edge AI laboratories, receives the generated results in return, and then pays expensive monthly usage bills.

In the past, this transaction seemed taken for granted: the capabilities of top closed-source models have been years ahead of other open-source ecosystems, and no matter how much effort you put into your own private data, you cannot bridge this generational capability gap.

However, the winds of change have fundamentally shifted.

Open-source weight models have already demonstrated hard power that can compete with the best cutting-edge closed-source models in many core benchmark tests with significant practical business value.

As AI accelerates its evolution into the central nervous system of business operations, more and more companies are becoming acutely aware of their high dependence on external model rental services.

We believe that in the future, more and more companies will desire to truly "own" the core models they rely on for survival: continuously iterating with the private data accumulated by the enterprise, autonomously operating on infrastructure of their own choosing, and having absolute control over when to upgrade and change the model.

This is particularly critical in highly regulated industries. Under extremely strict data sovereignty compliance requirements and physically locked network architectures, the hidden costs of continuing to rent external models compared to actual business output become extremely unprofitable.

However, there remains a huge engineering and cognitive gap between downloading a public model from the open-source community and enabling a company to truly own a self-evolving proprietary intelligence system.

Professional forces must help enterprises transform complex business workflows into precise evaluation benchmarks (Evals), distill employees' implicit business experiences into high-quality training data, convert occasional failures in online production environments into nutrients for continuous model iteration, and seamlessly pass on these core assets during each iteration of the underlying base model.

When dealing with the most extreme and complex general unknown problems, enterprises may still choose to continue renting public computing power from top cutting-edge laboratories; but we firmly believe that the vast majority of enterprises must have their own vertical dedicated models.

These models will fully encode the business operation mechanisms of the enterprise itself and will become increasingly valuable as they evolve with the business development of the enterprise.

If you are helping businesses across various industries truly take control of and own their proprietary intelligence, we look forward to hearing from you.