YC RFS AI Infrastructure — What YC Wants to Fund

This page breaks down exactly what YC means by AI infrastructure across its RFS editions, what specific problems they want solved, and what a fundable AI infrastructure application looks like.

YC's RFS draws a consistent line between three layers:

Layer 1 — Foundation Models: OpenAI, Anthropic, Google, Meta. YC funds few companies competing directly at this layer. The capital requirements and talent concentration make it structurally difficult for a new startup to compete.

Layer 2 — AI Infrastructure (what YC wants): The tooling, platforms, and physical infrastructure that sits between foundation models and end-user applications. This is the layer YC is most actively funding and most consistently calling for in the RFS.

Layer 3 — AI Applications: Products built on top of foundation models for specific use cases. YC funds many of these (the "full-stack AI company" thesis), but they are distinct from infrastructure.

AI infrastructure in YC's framing includes:

• Evaluation and testing tools for LLM applications
• Observability and monitoring for production AI systems
• Data pipelines and data management for AI training
• Model fine-tuning and deployment infrastructure
• Physical compute infrastructure (data centers, cooling, power)
• Security and compliance for AI systems
• Cost optimization for AI API usage
• Agent orchestration and multi-agent infrastructure

Datacenter and Physical Compute Infrastructure

Across the 2025 RFS editions, YC repeatedly called for companies addressing the physical compute bottleneck. The specific framing from the Spring 2025 RFS: demand for data centers is growing faster than the traditional development pipeline can supply, creating a structural shortage. YC wants startups working on:

• Software for data center construction and management: Planning, project management, materials procurement, and operational management tools that compress the multi-year timeline for new data center builds
• Power infrastructure: Solutions that address the electrical grid capacity constraints limiting data center expansion — demand response systems, on-site generation, efficiency improvements
• Cooling innovation: New cooling approaches for the density of compute required by AI workloads, where traditional air cooling is increasingly inadequate
• Modular and portable compute: The Fall 2026 RFS introduced "Compute at Sea" — offshore compute infrastructure using the ocean as a natural cooling medium and avoiding land permitting constraints. This reflects a broader YC interest in non-traditional physical compute deployments

LLM Observability and Evaluation

YC has explicitly called for and funded multiple companies building observability and evaluation infrastructure for production LLM applications. The specific problems:

• Evaluation frameworks: Systematic methods for measuring LLM application quality, accuracy, and regression — comparable to unit testing for traditional software
• Production monitoring: Tools that detect when an LLM application is producing wrong, hallucinated, or degraded outputs in real time
• Dataset management: Infrastructure for building, versioning, and managing the datasets used to evaluate and fine-tune LLM applications
• Cost tracking and optimization: Tools that give developers and enterprises visibility into LLM API costs at the application and feature level

Multi-Agent Infrastructure

The Summer 2025 and Fall 2025 RFS editions both explicitly called for multi-agent system infrastructure. The specific problems:

• Agent orchestration: Systems for coordinating multiple AI agents working on shared tasks — task routing, state management, conflict resolution, and progress tracking across agent fleets
• Persistent memory for agents: Infrastructure that gives AI agents durable, queryable memory across sessions — beyond the context window limitations of individual model calls
• Agent security: Access control, audit logging, and sandboxing for agents that take real-world actions on behalf of users or organizations
• Error handling and recovery: Frameworks for detecting when an agent has failed or gone off-task and recovering gracefully without human intervention

Fine-Tuning and Model Specialization

YC wants companies building infrastructure that makes it economical for enterprises to fine-tune foundation models on their proprietary data. The specific problems:

• Fine-tuning pipelines: End-to-end infrastructure for collecting training data, running fine-tuning jobs, evaluating the resulting model, and deploying it safely
• Synthetic data generation: Tools for generating high-quality synthetic training data in domains where real labeled data is scarce or sensitive
• Model compression and optimization: Tools that reduce the cost and latency of running fine-tuned models in production

AI Security and Compliance Infrastructure

The AI-native compliance infrastructure category appeared explicitly in the Fall 2026 RFS (written by a YC-backed founder) and reflects a broader YC interest across 2025 editions:

• AI governance platforms: Tools that give enterprises visibility and control over which AI models are being used, what data they are accessing, and what outputs they are producing
• Prompt injection and adversarial defense: Security infrastructure that protects AI applications from prompt injection attacks, data exfiltration via AI, and adversarial inputs
• Regulatory compliance for AI: Tools that help companies comply with emerging AI regulations (EU AI Act, NIST AI RMF, sector-specific AI rules in healthcare and finance)

YC's own portfolio provides the clearest signal of what fundable AI infrastructure looks like. Notable YC-funded companies in AI infrastructure across recent batches:

The pattern: W24 and W25 were particularly strong batches for AI infrastructure, correlating directly with the RFS emphasis on these categories.

The picks-and-shovels moment

Every major platform shift produces a period where infrastructure companies — picks and shovels — outperform application companies in the early years. This happened with cloud infrastructure (AWS, Cloudflare, HashiCorp) during the cloud transition. YC believes the same dynamic is playing out in the AI transition: as developers and enterprises build AI applications, demand for the underlying infrastructure compounds.

Foundation model commoditization creates infrastructure demand

As foundation model capabilities become more similar across providers — and as pricing falls — the differentiation in AI applications moves to deployment quality, reliability, cost efficiency, and security. All of these are infrastructure problems. The better foundation models become, the more important the infrastructure layer becomes for building production-quality applications on top of them.

Production AI is harder than prototype AI

YC partners observed across multiple batches that many AI applications that worked impressively in demos failed in production because of reliability, latency, cost, and safety issues. Infrastructure that solves production AI challenges is therefore addressing a pain point that becomes more acute as AI adoption matures.

Physical compute is a genuine bottleneck

YC's repeated call for data center and compute infrastructure reflects a real supply constraint that multiple YC partners believe will persist for 5-10 years. The demand for AI compute is growing faster than the traditional electricity-grid-to-data-center-to-rack supply chain can support — creating a structural opportunity for startups that improve any part of that supply chain.