Y Combinator

Harshil Mathur: AI Is Compressing Every Moat

Harshil Mathur, Jon Xu

Founded by Harshil Mathur in 2015, Razorpay became the first Indian Y Combinator company to pivot from a failed educational focus to a payment infrastructure provider for startups after navigating a twelve-month regulatory gestation period. The firm secured a competitive moat by rejecting early global acquisition offers and becoming the first gateway to support UPI, subsequently processing over $180 billion in gross merchandise value while maintaining capital efficiency that inadvertently generated profits. Today, Mathur continues to lead the company through a strategic reinvention using AI, emphasizing that founder-led conviction in long-term problem-solving remains essential despite advancements in software development speed.

Inference Chips for Agent Workflows

Current AI inference hardware fails to support agentic workflows, which rely on complex loops and tool calling, resulting in low GPU utilization and a critical market gap for specialized silicon. Strategic acquisitions by NVIDIA and innovations like Google's TPU v7 highlight the industry's urgent recognition that successful agentic AI requires native features for fast context switching, speculative decoding, and persistent memory caching across execution graphs. To address this, the speaker is actively seeking collaborators with deep expertise in both chip architecture and agent execution flows to develop the necessary compiler and hardware optimizations for this emerging paradigm.

AI-Native Discovery Engines

Frontier AI models are transforming the traditional scientific discovery loop by advancing from co-pilot tools to autonomous systems capable of executing full design-make-test-analyze cycles. These AI-native engines are already driving breakthroughs in drug discovery, material science, and protein engineering by generating hypotheses and validating them through automated laboratory workflows. As these closed-loop systems mature, they are poised to become the primary drivers of scientific progress alongside human researchers rather than remaining mere assistants.

The AI Operating System for Companies

Advanced AI-native organizations are adopting a closed-loop architecture that renders entire organizations queryable by AI, capturing data from meetings, tickets, and customer interactions into a unified reasoning engine. This methodology delivers a 50% reduction in sprint times and a 10x velocity increase by enabling immediate system monitoring and automatic adjustments, addressing a critical market gap where no existing product successfully aggregates context from disparate tools like Slack, Linear, and GitHub. A new initiative aims to build this essential connective layer to make companies legible to AI by default and is actively seeking collaboration with entities developing similar infrastructure.

SaaS Challengers

AI coding tools have slashed software production costs by up to 100 times, dismantling the legacy code moats that once protected incumbent SaaS providers from competition. While investors have slashed billions from established market valuations, the analysis outlines four disruptive strategies for startups to target massive, untapped sectors like chip design and ERP systems. This shift marks a generational transition where new entrants can replace decades-old codebases of millions of lines with integrated, AI-native alternatives.

Industrial Capabilities in Space

Adi Oltjan

StarCloud co-founder Adi Oltjan outlines a strategy to build space-based data centers by extracting lunar raw materials like silicon, aluminum, iron, and titanium. The company plans to achieve superior efficiency through electrolysis and 3D printing technologies that process molten regolith without Earth's gravity constraints. Y Combinator is currently soliciting applications from founders developing similar industrial capabilities in the Moon and deep space sectors.

Software for Agents

Projections indicate that the next trillion internet users will be AI agents rather than humans, yet current human-centric infrastructure creates slow and brittle interactions due to its reliance on visual interfaces. To address this, the market requires a new software foundation built on machine-readable protocols like APIs and CLIs, with successful startups emerging by designing "agent-first" architectures instead of retrofitting legacy systems. This shift creates a primary investment opportunity in building the foundational software infrastructure that enables autonomous agent discovery and usage, inviting founders to initiate contact if they are developing solutions for this specific need.

Hardware Supply Chain

The firm is accelerating investment in US hardware startups across medical devices, robotics, and space sectors to build a comprehensive stack for rapid iteration. This strategy directly addresses the significant speed disparity between US production cycles, which take several weeks, and the one-day turnaround available in Shenzhen. By prioritizing companies that tightly integrate design, manufacturing, and logistics, the initiative aims to close the competitive gap with early players like HLabs and Prototyping IO.

Recursion Is The Next Scaling Law In AI

Ankit Gupta, Francois Chaubard

Two 2025 research initiatives, Hierarchical Reasoning Models (HRM) and Tiny Recursive Models (TRM), challenge standard scaling laws by utilizing inference-time recursion to achieve state-of-the-art reasoning with drastically fewer parameters. HRM reaches 27 million parameters using a three-level weighted hierarchy, while the simplified TRM distills the architecture to just 7 million parameters yet achieves 87% accuracy on ARC-Prize benchmarks by treating recurrence as a dynamic latent memory tape. These systems overcome historical RNN limitations through Deep Equilibrium Models and latent recursion, offering a pathway to efficient, deep reasoning that diverges from traditional Chain-of-Thought constraints.

Startups That Sell to the Biggest Companies in the World

Over the past three years, YC-backed startups have secured multi-million dollar deals with Fortune 10 enterprises in their first year, reversing traditional market barriers through AI-driven development velocity and strategic necessity. Coding agents now allow small teams to build nuanced products for major corporations in months, rendering the historical stealth-launch model obsolete while triggering active engagement from enterprise leadership. Consequently, the investment firm is pivoting its strategy to prioritize backing for teams capable of targeting sales deals with the world's largest companies rather than focusing on long-term feature parity.

Supply Chain 2.0 for Semiconductors

The 2021 semiconductor shortage and evolving CHIPS Act investments have exposed critical visibility gaps in multi-tier supply chains that legacy enterprise software cannot resolve. A new market opportunity has emerged to build specialized "Supply Chain 2.0" solutions capable of managing TSMC's advanced packaging bottlenecks, export control volatility, and the unique complexities of wafer allocation. This initiative aims to replace fragmented legacy systems with real-time risk monitoring and compliance tracking tailored specifically for the semiconductor industry.

Dynamic Software Interfaces

Current software interfaces remain largely static and generic, yet the maturation of coding agents now empowers users to function as their own forward-deployed engineers for radical customization. To realize this future where shared primitives support wildly divergent front-end designs, the organization is calling on radical thinkers to redefine the software delivery stack and determine whether source code must replace packaged binaries. This initiative seeks to resolve critical architectural questions regarding how user-controlled agents can safely modify both visual elements and underlying middleware to match specific professional needs.

YC Paper Club

YC is launching the "Paper Club," a community initiative for researchers, engineers, and founders to navigate the rapid pace of AI development through small, biweekly gatherings at its Mountain View office. The program focuses on selecting under 100 participants, specifically targeting builders who implement papers and researchers studying production failures, to discuss new research presentations followed by private dinners. Applications are currently open for the inaugural cohort, with the first session scheduled for May 20th.

Demis Hassabis: Agents, AGI & The Next Big Scientific Breakthrough

Demis Hassabis, Garry Tan

DeepMind CEO Demis Hassabis projects Artificial General Intelligence will emerge around 2030, contingent on inventing specific architectural breakthroughs in continual learning and structured memory rather than merely scaling current methods. While deep tech startups are warned to anticipate AGI disrupting their decade-long cycles, the organization is simultaneously deploying efficient, modular "Flash" and open-source nano models to power edge devices and accelerate scientific discovery in fields like drug design and materials science. Ultimately, DeepMind advocates for interdisciplinary collaboration between AI and physical sciences to overcome current reasoning limitations and drive the next generation of autonomous agents.

Electronics in Space

Philip Johnston

StarCloud co-founder Philip Johnston is addressing the surge in space launch capacity from companies like SpaceX and Astrospace by developing orbital data centers optimized for inference workloads. Y Combinator is currently recruiting chip designers with backgrounds at SpaceX and NVIDIA to create specialized processors that meet critical requirements for reduced mass, thermal efficiency, and radiation hardening. This initiative aims to capture the urgent market demand for computing resources that can operate effectively in the harsh space environment.