The Truth About The AI Bubble

Diana says Anthropic became the most popular API among founders applying to YC’s Winter 2026 batch, edging out OpenAI. Gemini also rose to about 23%. The speakers point to coding performance, reasoning, and reliable web grounding as reasons founders and users choose different models.

The speakers describe relying on ChatGPT’s memory, Perplexity for research, and Gemini for grounded answers. Some startups are building orchestration layers to switch models or assign them different tasks. They say proprietary evaluations help companies identify which models work best for their own users and industries.

Jared compares AI infrastructure spending to the telecom buildout that helped make YouTube possible. The speakers argue that overbuilding can leave abundant compute for application companies, even if infrastructure investors face losses. They describe the current period as a transition from heavy installation spending toward wider deployment of AI applications.

The speakers see more founders building smaller models for particular devices, languages, and industries. They say reinforcement learning and fine-tuning can help specialized models outperform larger ones on domain tasks, citing a healthcare model with 8 billion parameters. They also caution that advances in frontier models can quickly erase a fine-tuning advantage.

The hosts say early AI startups reached $1 million in ARR with very small teams, but that pattern had not yet carried through to $10 million in ARR without hiring. They suggest faster production raises customer expectations and keeps companies hiring to execute. Gamma is cited as reaching $100 million in ARR with 50 employees.