What Surprised Us Most In 2025

Analysis of YC Winter 2026 batch data reveals Anthropic Claude has become the #1 API choice (52%+) among applicants, surpassing OpenAI for the first time. This shift is driven by Claude's superior performance in coding tasks and vibe coding tools, with Gemini also climbing to 23% market share. The discussion covers model personalities and switching behaviors among technical users.

• •Anthropic's market share among YC startups grew from ~25% to 52%+ in 2025, overtaking OpenAI's historically dominant 90%+ position
• •Claude's success attributed to superior coding performance - intentionally optimized as their North Star eval metric per Tom Brown
• •Gemini climbed from single digits to 23% share, with partners noting its superior reasoning and real-time information accuracy via Google's Grounding API
• •Model personalities matter: OpenAI has 'black cat energy', Anthropic is the 'helpful golden retriever', Gemini sits in between
• •Consumer stickiness driven by memory features - ChatGPT's personalization creates moat despite competitors' technical advantages

Startups are building orchestration layers to swap models dynamically based on task-specific performance, moving away from single-vendor loyalty. Series B companies now use different models for different stages (e.g., Gemini for context engineering, OpenAI for execution), with proprietary evals determining the winner for each use case. This commoditization of models shifts value to the application layer.

• •Series B AI companies building abstraction layers to swap models in/out as new releases come, treating them like Intel/AMD processors
• •Specific example: startups using Gemini 3.0 for context engineering, then feeding output to OpenAI for execution
• •Proprietary vertical-specific evals (especially in regulated industries) determine best model per task, not generic benchmarks
• •Model commoditization reduces vendor lock-in and increases competition, benefiting application-layer startups with lower costs
• •Consumer behavior mirrors enterprise: power users run same prompts through Claude, Gemini, and ChatGPT tabs, then cross-check outputs

Partners argue the perceived 'AI bubble' in infrastructure spending (NVIDIA, OpenAI CapEx) actually creates opportunities for application-layer startups, similar to how telecom overinvestment in the 90s enabled YouTube. The glut of compute capacity, driven by competition among AI labs and chip makers, means lower costs and more resources for college students and founders building applications.

• •Telecom bubble analogy: billions in 'wasted' 90s fiber investment created cheap bandwidth that made YouTube possible - AI following same pattern
• •NVIDIA stock volatility (down to ~$170s) reflects healthy competition from AMD, TPUs - more compute options benefit startups, not harm them
• •Question 'is it a bubble?' only matters if you're NVIDIA/Comcast equivalent - irrelevant for application-layer founders who are the 'YouTube'
• •Meta's Zuck acknowledged potential CapEx overinvestment, but big companies must build infrastructure - startups benefit without the risk
• •Carlota Perez framework: currently transitioning from 'installation phase' (heavy CapEx, feels like bubble) to 'deployment phase' (application explosion)

Power generation and land constraints are pushing data center innovation to space. StarCloud (YC S24) pioneered the concept, now adopted by Google and Elon Musk. Boom Supersonic pivoted to using jet engines for data center power due to supply chain backlogs. YC portfolio includes the 'trifecta': StarCloud (space data centers), Helion/Boom (energy), and Zephyr Fusion (space-based fusion).

• •StarCloud (YC S24) mocked for space data centers in summer 2024, now Google and Elon Musk's top talking point 18 months later
• •Terrestrial constraints: not enough land (California CEQA blocks building), not enough power generation capacity for AI compute needs
• •Jet engine supply chain for data center power so backed up that orders placed today won't arrive for 2-3 years
• •Zephyr Fusion (YC W26): national lab engineers discovered fusion physics 'pencils' in space, pursuing gigawatt-scale space energy
• •YC trifecta solving data center buildout: StarCloud (land), Helion/Boom (energy), Zephyr (space fusion) - all addressing infrastructure bottlenecks

The skill set for training models is becoming more common, similar to how startup knowledge dispersed over the past decade. Open source models plus fine-tuning and RL enable domain-specific models with 8B parameters to beat OpenAI in vertical tasks. However, companies must maintain post-training infrastructure to keep pace as frontier models improve.

• •Model training expertise transitioning from rare (OpenAI's Ilya/Greg/Sam combination) to commonplace, enabling more applied AI companies
• •Open source models (especially with RL) allow 8B parameter domain-specific models to beat OpenAI in vertical benchmarks (e.g., healthcare)
• •YC healthcare startup collected proprietary dataset, outperformed OpenAI on healthcare benchmarks with only 8B parameters via fine-tuning
• •Critical caveat: must maintain post-training infrastructure - companies that beat GPT-3.5 got 'blown out of the water' by GPT-4.5 and 5.1
• •Trend toward more specialized models for specific tasks rather than relying solely on general-purpose frontier models

What started as observed founder behavior in early 2025 evolved into a major category with companies like Replit, Emergence, and Google's Anti Gravity (led by YC alum Varun Mohan). While vibe coding isn't yet reliable for 100% production code, it's gained significant traction and even Sundar Pichai is publicly discussing it.

• •Vibe coding evolved from YC observation to major category with multiple winners: Replit, Emergence, Google Anti Gravity
• •Google's Anti Gravity launch featured cinematic video with Varun Mohan (YC alum) and Sergey Brin, showing enterprise validation
• •Sundar Pichai publicly discussing vibe coding alongside space data centers as key Google initiatives
• •Reality check: vibe coding not yet 100% reliable for production code shipping in 2025, still requires human oversight
• •Category growth driven by Anthropic's coding-optimized models becoming preferred choice among technical founders

The AI economy has stabilized into clear layers (models, infrastructure, applications) with established playbooks for building AI-native companies. Unlike 2024's rapid ground-shifting changes, 2025 saw incremental model improvements without major disruptions. This maturation means finding startup ideas has returned to normal difficulty levels - the 'wait a few months for new capabilities' strategy no longer works.

• •2025 marked transition from chaotic rapid change to stable AI economy with clear value distribution across model/infrastructure/application layers
• •Incremental model improvements in 2025 vs. 2024's major capability jumps - no more 'other shoe dropping' moments that reshape landscape
• •Pivot-and-wait strategy dead: can no longer survive by waiting months for announcements that unlock new idea spaces
• •Established playbook now exists for building AI-native companies on top of models, reducing uncertainty for founders
• •Finding startup ideas returned to 'normal levels of difficulty' - the brief window of easy pivoting has closed

Fast takeoff predictions (like AI 2027 report) haven't materialized due to log-linear scaling laws and organizational resistance to change. First-wave AI companies like Harvey that raised massive rounds and did victory laps now face serious competition from second-wave startups like Legora and Giga. Despite AI efficiency gains, companies still hire at pre-AI rates because customer expectations rise faster than productivity.

• •AI 2027 'doomer' report quietly revised timeline - fast takeoff skepticism validated by log-linear scaling requiring 10x more compute per improvement
• •90% enterprise AI project 'failure' rate actually reflects general IT incompetence, acts as brake on societal AI adoption (paradoxically good for gradual absorption)
• •First-wave winners like Harvey (legal AI) now facing serious competition from second-wave startups like Legora - early capital moats not insurmountable
• •Harvey's cautionary tale: burned double-digit percentage of capital on fine-tuning that provided no lasting advantage, only diluted founders
• •Hiring hasn't decreased despite AI: customer expectations rise faster than productivity gains, companies still bottlenecked on execution talent not ideas