Y CombinatorWhy the AI Bubble Misses Where Startups Actually Win
With models swapping in and out, the startup edge lies in the application layer; vibe coding and infrastructure bets are the actual durable advantages.
CHAPTERS
- 0:00 – 0:42
AI economy feels stabilized: clearer layers and a startup playbook
The hosts open by reflecting on 2025 and how the AI market feels more “settled” than prior years. They describe a clearer separation between model, application, and infrastructure layers, and a more repeatable playbook for AI-native startups.
- •AI stack has crystallized into model/app/infrastructure layers
- •Perception that multiple layers can make significant money
- •Less feeling of “ground shifting” compared to late 2024
- •A more standard playbook is emerging for founders building on models
- 0:42 – 2:06
Shift in YC founders’ preferred LLMs: Anthropic overtakes OpenAI
YC’s Winter ’26 applicant data shows a major change: Anthropic becomes the most selected API, surpassing OpenAI for the first time. The group discusses the speed of this shift and what it signals about developer preferences.
- •YC asks founders which model/API they use in their stack
- •Anthropic edges out OpenAI in this batch after years of OpenAI dominance
- •The change accelerated in the last 3–6 months
- •Anthropic’s share is described as “hockey stick” growth
- 2:06 – 3:18
Why Claude is winning: coding agents, “vibe coding,” and model taste
They attribute Anthropic’s surge to strong performance in coding and agentic developer workflows, which became a high-value category. They also propose that personal tool familiarity (using Claude for coding) influences founders’ default choice even for non-coding products.
- •Coding agents/vibe coding became a bigger value driver than expected
- •Anthropic models are perceived best for coding workflows
- •Anthropic appears to have intentionally optimized internal evals for coding
- •Founder familiarity and “personality” of a model can drive adoption
- 3:18 – 5:35
Gemini’s rise and the consumer UX moat: grounding, accuracy, and memory
The conversation turns to Google’s Gemini climbing rapidly in YC usage and in personal workflows. They compare why people stick with ChatGPT (memory/personalization) versus Gemini (grounded, fresh information) and Perplexity (speed but sometimes accuracy gaps).
- •Gemini’s share rises substantially compared to prior year
- •Gemini is favored for grounded, near-real-time answers using Google index
- •Perplexity can be fast but not always accurate
- •ChatGPT’s memory/personalization is described as a sticky consumer moat
- 5:35 – 6:38
Why there still aren’t more AI consumer apps (and why trust is the blocker)
Harj is surprised that more verticalized consumer apps haven’t formed around common high-stakes life workflows. He explains that model reliability still requires heavy prompting/context engineering, which makes it hard to productize into a “just works” app for critical decisions.
- •User behavior: long-running conversations stuffed with documents and context
- •High-value transactions amplify the need for accuracy
- •Lack of trust pushes users toward manual prompting and verification
- •Opportunity exists for apps that package workflows end-to-end
- 6:38 – 9:08
The new normal: model arbitrage and orchestration layers
They describe a growing pattern among startups—especially later-stage ones—of abstracting the model layer behind an orchestration system. Companies swap models in/out by task, using proprietary evals tied to their vertical data and regulatory constraints.
- •Startups increasingly use multiple models rather than a single provider
- •Orchestration layers enable quick swapping as new models release
- •Task specialization: different models for different agent steps (e.g., context vs execution)
- •Proprietary evals become a competitive asset, especially in regulated verticals
- 9:08 – 12:35
The AI bubble question: why ‘overbuild’ can be good for startups
They tackle the “AI bubble” discourse, arguing that even if infrastructure is overbuilt, it can create abundance that enables breakout applications—like telecom overbuild helped make YouTube possible. The key distinction: bubbles hurt capex-heavy incumbents more than dorm-room startups.
- •Bubble framing matters more for infrastructure players than app startups
- •Telecom bubble analogy: excess capacity lowers costs and unlocks new products
- •Compute competition (NVIDIA vs AMD/TPUs) increases supply and drives commoditization
- •More competing AI labs generally benefits application-layer founders
- 12:35 – 14:33
Installation vs deployment: Carlota Perez framework for AI’s ‘frenzy’ phase
Diana introduces the idea that tech revolutions have an investment-heavy installation phase followed by a deployment phase where applications proliferate. AI today looks like it’s transitioning between these phases—good news for startups building products rather than data centers.
- •Two phases: installation (capex, hype) then deployment (broad adoption)
- •ChatGPT moment triggered excitement; GPU/datacenter buildout followed
- •Overbuild (e.g., dark fiber) can coexist with long-term economic transformation
- •Founders are positioned to win during deployment by building apps
- 14:33 – 16:20
Space data centers and energy constraints: when “crazy” becomes mainstream
They revisit the once-ridiculed idea of data centers in space and note how quickly it moved into the mainstream conversation. The drivers are practical constraints—power generation, land, supply chains, and regulation—pushing big players to extreme solutions.
- •StarCloud-style space datacenters went from mocked to copied by major companies
- •Power generation is a bottleneck for AI datacenter expansion
- •Land/regulatory limits (e.g., building constraints in the US/California) push alternatives
- •Supply chain constraints (even jet engines as generators) shape strategic timelines
- 16:20 – 17:28
YC’s ‘trifecta’ of infrastructure bets: space, fusion, and power for AI
The hosts connect multiple YC companies to the same macro constraint: scaling compute requires new land and massive energy. They discuss fusion efforts and a new investment thesis—fusion in space—as a potential path to gigawatt-scale energy.
- •Companies tackling land constraints (space) and energy constraints (fusion/other)
- •Boom and Helion referenced as part of the solution space
- •Zephyr Fusion: fusion-in-space concept driven by physics/economics penciling out
- •Compute scaling increasingly couples AI progress to infrastructure breakthroughs
- 17:28 – 19:31
Rising interest in starting model companies—especially small, specialized models
Harj notes increased founder interest in creating model companies, not just apps. While few can compete head-on with frontier labs, many teams aim for smaller edge models, language-specific voice models, and other targeted capabilities as know-how spreads.
- •More founders exploring training/building models vs only building on APIs
- •Frontier competitors are rare; specialized/smaller models are more common
- •Analogy to early SaaS: knowledge diffusion increases the number of builders
- •Model-building skill sets are becoming less rare and more teachable
- 19:31 – 21:01
Domain models via fine-tuning + RL: potential—and the treadmill problem
They discuss how reinforcement learning and fine-tuning on open-source bases can produce domain-leading performance with far fewer parameters, especially with proprietary datasets. But they also warn that frontier releases can quickly invalidate an advantage unless teams keep investing in post-training.
- •RL and fine-tuning make strong domain-specific models feasible
- •Proprietary datasets can create real performance edges in verticals like healthcare
- •Small models can beat general models on targeted benchmarks
- •Frontier model releases can leapfrog fine-tuned systems, forcing continuous iteration
- 21:01 – 23:04
Vibe coding becomes a major category, but not a full replacement for engineers
They reflect on how vibe coding shifted from an observed founder behavior to a major startup category, with multiple winners. At the same time, they caution that fully trustworthy production code generation isn’t here yet, even if tooling is rapidly improving.
- •Vibe coding moved from meme/behavior to major product category
- •Multiple companies cited as succeeding in the space
- •Even in late 2025, AI coding isn’t 100% reliable for production code
- •Big tech marketing and demos amplify the trend, but practical limits remain
- 23:04 – 30:22
AI progress feels more incremental; founder realities: teams still matter
They return to the theme that 2025 brought fewer destabilizing leaps, making ideation harder again and execution more important. They also address AI 2027 “fast takeoff” skepticism and conclude that despite earlier tiny-team stories, companies still hire—because customer expectations and competition rise.
- •Model progress in 2025 felt incremental, not paradigm-shifting
- •Startup idea-finding returns to ‘normal’ difficulty as surprise announcements slow
- •Skepticism of doomer timelines: scaling laws and organizational inertia act as brakes
- •Post-Series A scaling still requires hiring; efficiency gains raise the competitive bar
