What Surprised Us Most In 2025

Episode Summary

Executive Summary: The episode argues that the AI market has matured into a more stable, layered economy with clear roles for model, application, and infrastructure companies. It highlights Anthropic overtaking OpenAI as the top YC model choice, Gemini’s rapid rise, the persistence of human and organizational friction, and the view that 2025 marked a shift from chaotic model churn to a more durable startup playbook.

Main Topics: AI economy has stabilized into layers (Priority: 5/5): The hosts say the AI landscape now looks more settled, with model, application, and infrastructure layers forming a recognizable stack and clearer paths for startups to build and monetize on top of foundation models. Anthropic becomes YC founders’ preferred model (Priority: 5/5): Diana notes that Anthropic overtook OpenAI as the most-used API among YC companies in the winter 2026 batch, reflecting a real shift in founder preference driven partly by coding and agent performance. Gemini’s rise and consumer usage patterns (Priority: 4/5): Gemini climbed rapidly in usage, with one host switching to it for reasoning and real-time search due to its grounding and accuracy. The discussion contrasts Gemini with ChatGPT’s memory and Claude’s personality. Model commoditization and orchestration layers (Priority: 5/5): Founders are increasingly using multiple models for different tasks, swapping them based on evals and workflow needs. The conversation suggests applications will abstract away the underlying model wars. AI infrastructure boom and the ‘installation’ phase (Priority: 5/5): The hosts frame current capex-heavy spending on GPUs, data centers, power, and even space-based infrastructure as the installation phase of a broader tech cycle, before widespread deployment creates major application winners. Vibe coding and new startup creation norms (Priority: 4/5): What began as an observed founder behavior became a major category, with vibe coding tools like Replit and others gaining traction. The hosts argue this is real but still not reliable enough for fully production-grade software. Hiring, growth, and the limits of AI efficiency (Priority: 4/5): Despite AI-driven productivity gains, companies still hire as they scale because customer expectations rise and execution remains bottlenecked by people, not ideas. Revenue may come with fewer employees, but not zero-headcount operations.

Key Arguments: AI is no longer in a chaotic, rapidly shifting phase; it has settled into a more legible stack of infrastructure, model, and application companies. Anthropic’s rise at YC suggests founders reward models that perform best on real workflows, especially coding and agentic tasks. Gemini’s growth reflects both model quality and Google’s grounding/search advantages, making it attractive for accurate real-time information. Founders and startups are increasingly model-agnostic, using evals and orchestration to route tasks to the best model for each step. The current AI boom resembles prior infrastructure buildouts: massive capex may look like a bubble, but it can create the conditions for future application-layer winners. Human inertia and organizational change will slow AI adoption, preventing a sudden takeoff and giving society time to adapt. AI raises customer expectations faster than it reduces staffing needs, so many companies still end up hiring more people as they grow. Smaller domain-specific models fine-tuned with RL can outperform large general-purpose models in narrow regulated domains such as healthcare.

Data Points: Anthropic share at YC: a bit more than OpenAI; over 52% growth in the last 3–6 months - Winter 2026 YC batch model-of-choice ranking and recent adoption trend OpenAI share at YC: previously 90%+ in earlier batches; around 20–25% through much of 2024 and early 2025 - Historical YC founder preference compared with current batch Gemini share at YC: about 23% - Winter 2026 batch usage ranking Gemini share previously: single digits, or even 2–3% - Earlier period before its climb in usage Healthcare domain model: 8 billion parameters - A YC startup reportedly beat OpenAI on healthcare benchmarks with a small specialized model Revenue efficiency example: $100 million ARR with 50 employees - Gamma cited as a strong example of the reverse-flex hiring trend AI 2020 revision: originally 2027, later quietly revised - Referenced as an example of long-term AI doom predictions being adjusted Enterprise AI project failure rate: 98% or 90% - Mentioned in reference to an MIT report and enterprise adoption skepticism Capital expenditure scale: tens of billions to hundreds of billions - Used in the telecom/infrastructure bubble analogy to explain AI buildout Model update cadence: 3 to 6 months - The period in which Anthropic’s share surged and model preferences changed rapidly

Pivotal Quotes: "I feel like the AI economy stabilized." — Diana: Describing 2025 as a year when the AI market became more structured and predictable "The number one API is actually Anthropic, came out a bit more than OpenAI, which who would have thought?" — Diana: On YC founders’ current model preference in the winter 2026 batch "It seems like everyone is going to make a lot of money, and there's kind of like a relative playbook for how to build an AI native company on top of the models." — Unnamed host: Summarizing the episode’s thesis that the startup ecosystem now has clearer rules

Implications: For founders, the message is to build with model flexibility, strong evals, and domain expertise. For the industry, the biggest near-term winners may be application and infrastructure companies, while model competition drives down costs and expands opportunity.