Y Combinator• Dec 22, 2025• 0:46:38Interview

The Truth About The AI Bubble

From Y Combinator

Y Combinator

Executive Summary

This episode reviews surprising AI trends in 2025, highlighting a major shift in the LLM landscape where Anthropic has surpassed OpenAI in usage among Y Combinator startups.
The speakers discuss the stabilization of the AI economy into distinct infrastructure, model, and application layers, creating a clearer playbook for new companies.
They also explore the intense infrastructure build-out, which is facing terrestrial constraints in power and land, leading to novel solutions like space-based data centers and fusion energy.
The conversation concludes that despite "bubble" concerns, the over-investment in infrastructure creates immense opportunities for application-layer startups, similar to how the dot-com bubble paved the way for Web 2.0 giants.

12 quotes

Concerns Raised

Potential for an AI investment bubble
Physical infrastructure constraints (power, land) for data centers
The difficulty of building a moat for application-layer AI companies

Opportunities Identified

Commoditization of foundation models creates opportunities for application-layer startups
Over-investment in compute infrastructure will lower costs for new companies
Development of novel energy and data center solutions (e.g., space-based)

Key Themes

LLM Market Share Disruption

The podcast highlights a significant shift in the foundation model market, with Anthropic overtaking OpenAI as the preferred API for Y Combinator's latest batch of startups. This change is attributed to Anthropic's superior performance in coding tasks and a potential 'bleed-through' effect into other use cases.

This indicates a maturing and competitive market where no single model provider is guaranteed long-term dominance, requiring developers to stay agile.

Stabilization of the AI Economy

The speakers observe that the AI industry has matured into a more stable, three-tiered structure: infrastructure (NVIDIA, AMD), models (OpenAI, Anthropic, Google), and applications. This has created a more predictable environment and a clearer 'playbook' for building AI-native companies.

This stability reduces existential risk for startups and allows for more focused value creation at the application layer.

Infrastructure Bottlenecks and Novel Solutions

The massive demand for AI compute is running into physical limits on Earth, specifically power generation and land availability. This is driving ambitious solutions like space-based data centers (StarCloud, Google) and new energy sources like fusion (Helion, Zephyr Fusion).

This highlights critical long-term investment opportunities and risks in the energy and real estate sectors that support the AI industry.

The 'Productive Bubble' Hypothesis

The speakers frame the current high level of AI investment not as a destructive bubble, but as a 'productive' one, similar to the telecom bubble of the late 90s. The over-investment in infrastructure (compute, data centers) will ultimately lower barriers to entry and enable a new generation of applications, like YouTube did for video.

This perspective suggests that even if some infrastructure investments fail, the resulting abundance of resources will be a net positive for innovation and entrepreneurs.

The Rise of Model Orchestration

Sophisticated AI companies are moving away from loyalty to a single model provider. Instead, they are building orchestration layers to dynamically route tasks to the best-performing or most cost-effective model, such as using Gemini for context engineering and OpenAI for execution.

This trend emphasizes the importance of building proprietary evaluation frameworks (evals) and avoiding vendor lock-in to maintain a competitive edge.

Get started free

Topics

AI market share LLMs Anthropic OpenAI Google Gemini Y Combinator AI startups AI economy Model orchestration AI infrastructure Data centers Space-based data centers Fusion energy NVIDIA AI bubble Vibe coding Carlota Perez

Processed Feb 24, 2026 yt-dlp + mlx-whisper + Gemini