In this latest edition of the Request for Startups (RFS), Y Combinator delivers a profoundly clear and pivotal macro-level assessment: the greatest bottleneck in AI development is no longer confined to the parameters and algorithms of foundational large models, but is rapidly diffusing into the real world. Superficial software-layer packaging is losing its premium value; the true disruptive opportunities are now migrating comprehensively into deep-water domains characterized by heavy capital investment and physical constraints—namely, physical manufacturing, energy and grid infrastructure, fragmented public administration silos, high-security isolation systems, and commercial space ventures.

This comprehensive article summarizes 11 cutting-edge directions prioritized by YC’s core partner team. It spans from AI experts priced at just $15/hour, democratizing access to professional-grade intelligence, to large language models capable of literary-quality writing. It dives deeper into domestic U.S. robotics manufacturing, data center power and cooling supply chains, trillion-dollar infrastructure finance tools on Wall Street, and enterprise-specific private model ecosystems. DeepTide TechFlow has conducted a full, in-depth translation of this piece, offering every founder, researcher, and investor on the front lines of the AI revolution an invaluable industry-wide panorama and navigational guide.

Issuing a "Request for Startups (RFS)" has long been a hallmark tradition at Y Combinator, designed to share with global founders the most critical challenges we are eager to see them solve.

However, these directions represent only a small fraction of our investment landscape. If any of these topics ignite your passion, consider it strong endorsement to dive in; but even if your venture doesn’t fall within this list, that absolutely does not preclude you from submitting an application to YC.

A major current trend is this: the primary bottleneck facing AI is rapidly expanding outward from the model layer itself.

We are excited about the next wave of startups actively building the necessary hardware, infrastructure, financial systems, and vertical-specific software—bringing frontier intelligence into the real economy across factories, data centers, public governance, and even outer space.

Startups in the Era of Cheap Space Launch

Author: Ankit Gupta (YC General Partner)

SpaceX is painting a clear near-future vision: in the not-so-distant future, Starship could achieve up to 10,000 launches per year.

Even today, the launch cost of a single-use Starship is already nearly on par with reusable Falcon 9, with orbital payload costs drastically reduced per kilogram.

As Starship technology scales and achieves full reusability, we will witness a multi-order-of-magnitude collapse in the cost per kilogram of payload delivery.

This means now is the golden moment to systematically contemplate entrepreneurial opportunities in the era of abundant space access.

What business models today remain unviable due to prohibitive launch costs—but would become highly feasible once those costs drop by 100x or more, triggering a full-scale orbital economy boom?

We believe such companies must be founded now. If you're a builder passionate about exploring this frontier, YC eagerly awaits your application.

$15/Hour Experts: AI-Powered Democratization of Professional Services

Author: Brad Flora (YC General Partner)

Most conversations about AI today revolve around how it will displace existing workers. What truly excites us, however, is the opposite path: leveraging AI to transform ordinary workers directly into professionals of expert caliber.

We are witnessing more and more enterprises breaking free from credentialism, hiring individuals with exceptional intuition and execution ability. Armed with advanced AI tools, they can now handle complex tasks previously reserved for senior specialists.

Imagine a customer service representative with outstanding communication skills and deep AI fluency suddenly capable of advising on complex mergers & acquisitions (M&A), supply chain optimization, or global trade compliance.

Countless roles exist in public welfare, tax planning, insurance claims processing, medical billing reconciliation, and immigration visa documentation.

In short, whenever professional certification acts as a bottleneck to service delivery, this model shines.

The last industrial revolution built a vast modern middle class through factory blue-collar jobs; this technological shift holds the potential to reshape an entirely new middle-class structure through universally accessible expert roles.

We are committed to funding startups dedicated to realizing this vision.

Robotics Made in America: Hardware Reasserting Domestic Sovereignty

Author: Diana Hu (YC General Partner)

Today, top-tier robotic foundation models are achieving cross-architecture, cross-platform general skill transfer. In my view, the competitive bottleneck in robotics is rapidly shifting toward physical hardware.

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

Even robots assembled domestically often depend deeply on Chinese-sourced 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 automotive giants were forced to cut production capacity directly.

By July, the Federal Communications Commission (FCC) suspended certification for new network-connected mobile robots unless they were fully manufactured in the U.S. or granted rare exemptions.

Under current compliance standards, a robot must have over 65% of its component cost sourced domestically, rising to 75% by 2029. This renders the traditional “import parts, assemble locally” branding strategy obsolete.

We are committed to supporting entrepreneurs developing robots and their entire upstream supply chains in the U.S.

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

If you’re working in this domain, we look forward to engaging with you deeply.

Truly High-Quality Writing LLMs: Literary-Level Proficiency and Depth

Author: Jared Friedman (YC General Partner)

As large language models advance dramatically in coding and mathematical reasoning, their writing quality and expressive depth have quietly declined.

This is not your imagination: compared to models from 2024 and 2025, today’s mainstream LLMs indeed sound more mechanical, rigid, and less nuanced in tone.

This isn’t a temporary technical glitch—it’s a fundamental structural challenge.

Reinforcement learning reward signals engineered to maximize coding proficiency actively flatten the model’s tonal color, emotional range, and unique writing style during training.

When faced with resource allocation trade-offs, top-tier AI labs will inevitably prioritize computational power for code generation and autonomous agent workflows (Agentic Workflows), sacrificing meticulous attention to text texture and stylistic elegance.

This creates a massive market vacuum: a chance for startups to build a truly distinctive model—deeply human, emotionally resonant, and loved for its authentic voice.

If you’re driven to break through here, we welcome your message.

"Plaid for Government Benefits": Breaking Down Administrative Data Silos

Author: Gustaf Alströmer (YC General Partner)

If you’ve ever applied for public benefits, you know how cumbersome the process is: providing income verification, confirming household members, submitting proof of residence. Then, switching to another aid program? You start over. Annual eligibility reviews? The same nightmare repeats.

Each government agency operates a separate portal—no interoperability, complete data fragmentation into isolated silos.

Countless applicants get rejected simply for missing a non-critical document, while civil servants manually verify pay stubs day after day.

Ironically, this bureaucratic system that excludes honest citizens is riddled with fraud vulnerabilities.

The U.S. Government Accountability Office (GAO) estimates that benefit fraud costs the federal government nearly $500 billion annually. To make matters worse, generative AI can now fabricate convincing fake pay stubs in seconds.

Before Plaid emerged in fintech, every financial app had to individually integrate with banks. After Plaid, users connect seamlessly with one click. Public administration desperately needs its own foundational layer.

With user consent, such a system could pull verified, authoritative records—from official sources—of income, taxes, and residence, then distribute them instantly to all relevant government approval systems.

“One verification, universal access.” This not only accelerates approvals for genuine beneficiaries but also eliminates room for fraudsters to forge documents at the source.

Starting in 2027, all U.S. states will be mandated to revalidate eligibility for millions of Medicaid recipients every six months.

Soon, AI agents will fully automate application processing and eligibility determination—but only if they have access to completely trustworthy, tamper-proof underlying data.

Whoever builds this unified data pipeline will effectively control the core digital infrastructure of modern societal safety nets. If you’re constructing this vital conduit, we look forward to partnering with you.

The "Wall Street" of Physical Infrastructure Boom: New Capital Mechanics & Financing Structures

Author: Brad Flora (YC General Partner)

Software startups face minimal initial costs, but machinery, GPU clusters, and ultra-large-scale data centers are entirely different.

Someone must pay real money upfront for thousands of tons of specialty steel and high-end silicon chips from day one. Yet, in today’s financial markets, mature financing tools for such assets are virtually nonexistent.

Traditional banks still assess the depreciation of advanced intelligent robots like ordinary farm tractors.

That’s why a group of sharp entrepreneurs collaborating with YC have begun inventing novel financial derivatives themselves. For example, issuing new credit lines specifically for robots generating $500,000 in annual revenue, or designing bespoke put-option contracts tied to future GPU compute capacity.

Historically, every great physical infrastructure boom—the transcontinental rail network, the national power grid, nationwide communications—has given rise to financial titans of equivalent scale.

We seek founders who can build dedicated lending institutions, asset leasing platforms, asset-backed securities (ABS) structures, and risk hedging tools tailored to the physical AI economy.

If you deeply understand both the physics and industrial dynamics of physical machines and master the fundamentals of capital operations, we look forward to a deep conversation.

Vertical-Specific Foundation Models

Author: Diana Hu (YC General Partner)

I believe Jev succeeds so precisely because it delivers strongly-typed, structured decisions with explicit confidence scores—not endless generative text—and completely eliminates the burden of training individual classifiers for each niche task.

I want to see more pioneering teams develop purpose-built vertical foundation models for industries like manufacturing, energy & power, and life sciences. These models should run accurately in zero-shot or few-shot scenarios.

The benchmark is extremely rigorous: must achieve or surpass the business accuracy of general-purpose models using orders of magnitude less inference compute cost—enabling low-cost, seamless deployment across all edge devices.

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

This compact 330M-parameter model requires no fine-tuning for specific tasks. With a single forward pass, it predicts multiple correlated time series simultaneously. Trained once, each new user simply provides their historical data as context and invokes the model directly.

The real difficulty lies in acquiring high-quality, domain-specific, mathematically verifiable reward signals (Verifiable Rewards).

Teams must not only hold legal ownership of data but demonstrate overwhelming competitive advantage over existing expert models in both prediction accuracy and compute efficiency.

If you possess deep expertise in a particular vertical industry and can build a proprietary model, we are eager to hear your proposal.

AI Revolution in City Halls and Local Governments

Author: Gustaf Alströmer (YC General Partner)

In the U.S., if you want to build a house or open a business, your first step is obtaining municipal administrative permits.

Endless forms, lengthy plan reviews, tedious site inspections—often requiring months of endless waiting. In some cities, the permit approval period exceeds the actual construction timeline.

City hall staff are buried under repetitive, menial administrative work: meticulously cross-checking blueprints against dense building codes, comparing land use zoning rules line by line, shuttling comments between departments.

Modern AI models, however, are fully capable of performing these tasks independently and efficiently.

The prolonged delays in approvals carry severe economic costs. Every extra month adds significantly to development expenses, causing many planned projects to stall mid-way.

Consider this: every newly constructed building contributes property tax revenue to the city for decades to come.

We rough-calculated: if the average U.S. city could reduce building permit approval times to just 10% of current duration, the country would gain an additional $16 billion in annual municipal tax revenue.

We are eager to fund tech companies transforming city hall operations with AI. Starting points could include construction permits, then expanding into 311 civic service hotlines, public records management, business licensing, and on-site safety inspections.

The U.S. has over 19,000 municipalities, most of which still rely entirely on primitive manual processes.

The tech venture community has long held a conventional belief: “Never sell software to government agencies.”

That rule may have held in the past, but it’s utterly obsolete today. YC has invested in several startups selling AI solutions directly to government bodies, with some achieving growth rates surpassing any traditional enterprise SaaS company we’ve seen.

If you’re expanding in this space, we very much look forward to hearing from you.

AI Data Center Supply Chain Innovation: Power, Cooling & Grid Integration

Author: Jared Friedman (YC General Partner)

The massive expansion of AI data centers unfolding across the U.S. represents the largest single capital expenditure project in American commercial history.

The scale of investment has surpassed the combined total of the transcontinental railroad, interstate highway system, and Manhattan Project.

This unprecedented infrastructure buildout is generating incalculable commercial opportunities across the entire supply chain.

Some segments are already well-known: demand for cutting-edge GPU chips, base power supply, and skilled electricians has reached fever pitch.

Yet beyond the spotlight lie numerous obscure but critical bottlenecks capable of halting entire industries.

Currently, lead times for large high-voltage transformers stretch to three full years. Low- and medium-voltage switchgear, high-voltage circuit breakers, and advanced liquid cooling systems are sold out.

Even raw materials at the top of the supply chain are sounding alarms: grain-oriented electrical steel (GOES)—essential for transformer cores—is produced by only one facility in the U.S., operating far beyond capacity.

We are intensely eager to invest in startups that empower and break through the entire data center supply chain. If you’re fighting this battle, we warmly welcome your application.

AI in Physically Isolated Economies: Defense, Classified & High-Security Environments

Author: Brad Flora (YC General Partner)

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

Defense contractors, major commercial banks, elite healthcare institutions, and nuclear power plants keep their mission-critical data locked behind physical isolation walls. Due to strict confidentiality and compliance requirements, they absolutely cannot upload such data to any third-party AI API interface.

This is precisely why vast, trillion-dollar sectors of the real economy remain outside the AI wave.

Many of our founders repeatedly hit the same wall: clients love the AI product’s capabilities, but insist on one condition—entire system must be deployed 100% inside their internal facilities, with zero connection to any public network.

We are committed to supporting technological pioneers capable of delivering cutting-edge AI into these physically isolated fortresses.

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

Sales cycles in this closed market are slower and barriers higher.

But the flip side is equally powerful: this sector holds the most abundant procurement budgets in the industry and will birth the most defensible, fortress-like companies. If you’re tackling this high barrier, we look forward to an in-depth conversation.

Own Your Own Intelligence: Private & Open-Source Sovereign Models

Author: Jon Xu (YC Visiting Partner)

Today, almost every AI-powered product relies fundamentally on externally “renting” intelligence.

Your system continuously sends token data to a handful of top-tier AI labs, receives generated outputs, and pays expensive usage bills monthly.

Until recently, this seemed inevitable: leading closed-source models were years ahead of open-source ecosystems. No matter how much effort you invested in your private data, you couldn’t close the generational gap in capability.

But the tides have turned fundamentally.

Open-weight models now match or exceed top-tier closed-source models in key benchmarks with tangible business value.

As AI accelerates into the central nervous system of enterprise operations, more companies are increasingly wary of their dependence on external model rental services.

We are certain that in the future, more and more enterprises will desire to truly “own” their core models: trained continuously on proprietary data, running autonomously on self-chosen infrastructure, and retaining absolute control over when and how the model evolves.

This is especially critical in heavily regulated industries. Under stringent data sovereignty compliance and air-gapped network architectures, continuing to rent external models results in hidden costs that far outweigh actual business output.

Yet, going from downloading an open-source model to enabling a company to operate a self-evolving, sovereign intelligent system involves a massive engineering and cognitive chasm.

Professional support is needed to convert complex business workflows into precise evaluation benchmarks (Evals), turn employees’ tacit knowledge into high-quality training corpora, transform sporadic production failures into continuous model improvement feedstock, and ensure these core assets are seamlessly inherited through every underlying model upgrade.

When facing the most extreme, complex, and unknown general problems, enterprises might still choose to rent public compute from top labs. But we firmly believe that the vast majority of enterprises must have their own vertical-specific models.

These models will encode the enterprise’s own operational logic, evolving alongside the business to become increasingly irreplaceably valuable.

If you’re helping organizations across industries truly own and control their own intelligence, we welcome your voice.

Authors: Y Combinator Team (Garry Tan, Brad Flora, Diana Hu, Jared Friedman, Gustaf Alströmer, Ankit Gupta, Jon Xu et al.), Translation: DeepTide TechFlow