The hosts say AI is dominating conversation in Silicon Valley, while some students they met in Cambridge were still pursuing conventional startup ideas. They describe an unusual overlap: YC startups are increasingly focused on AI, and gains in the public market are concentrated in the Magnificent Seven, which they link to AI.
Are We In An AI Hype Cycle?
Speakers
Y Combinator partners Garry Tan, Jared Friedman, Harj Taggar, and Diana discuss whether AI is in a hype cycle on the Lightcone podcast.
The panel compares AI infrastructure investment to railroads: building the infrastructure raises the question of whether enough useful services will follow. They discuss how quickly the landscape has changed, from fears that a few foundation models would dominate to more model choice, including competitive open-source models such as Llama. They say it remains unclear whether chips, hosting, model makers, or application companies will capture the most value.
The panel notes that successful businesses such as DoorDash and Instacart emerged years after the iPhone launched. Founders building at the AI application layer can use existing models to solve customer problems without raising the huge sums needed for chips or foundation models. They also point to vertical products such as Photoroom's AI image tools for e-commerce and PermitFlow's permit-form workflows.
The panel distinguishes useful technology from rapidly rising valuations, citing AI startups valued at more than a billion dollars soon after starting and comparisons with crypto's boom. They point to YC companies' aggregate revenue rising from $6 million at application to $20 million by the end of a three-to-four-month batch. They caution that growth must last: renewals and customer retention determine whether a business creates lasting value.
The panel raises concerns that AI workflows may need human review, face enterprise trust barriers, or be commoditized as model wrappers. They argue that domain expertise, private data, fine-tuning, and careful workflow design can make application products valuable, and describe examples of companies automating accounts receivable and call-center work. They close by contrasting short-term popularity with long-term business value built on paying customers and durable cash flows.
Takeaways
- Build around a specific customer problem that people will pay to solve.
- Use existing models to test application ideas before investing in infrastructure.
- Track renewals and retention alongside early revenue growth.
- Assess whether your product's workflow, data, or domain knowledge creates an advantage beyond the underlying model.
- Treat fast-rising valuations as uncertain; focus on whether customers keep getting value.
Mentioned in this talk
Search inside this talk
What was claimed 11
- metric
Y Combinator: Funds startups with $500,000
“we fund you for $500,000”
YouTube Unverified1 source· Talk
- built
“we started the post-training for Llama 2”
1:06inScaling to Long HorizonsYouTube Verified5 sources· Talk, Web· archived
- role atGarry Tan is runs Y Combinator full-time at
Y Combinator
“I'm Garry, president and CEO of Y Combinator.”
0:09inHow to Make Claude Code Your AI Engineering TeamYouTube Verified7 sources· Talk
- feature
Llama: Open source; described as essentially at parity with frontier models
“the four of us were talking about how great it would be if open source could catch up to frontier models”
YouTube Unverified1 source· Talk
- metric
OpenAI: About 80–90% of companies in previous YC batches used its models“90-ish percent, 80% of folks were using OpenAI models”
YouTube Unverified1 source· Talk
- metric
Claude Sonnet: A couple dozen companies in the current YC batch use it, compared with one or two in the previous batch
“now a couple dozen are using it”
YouTube Unverified1 source· Talk
- invested in
Y Combinator invested in
Coinbase
“You two were the founding investors of Coinbase”
YouTube Unverified1 source· Talk
- feature
Coinbase: Marketplace for buying and selling cryptocurrency
“Coinbase was essentially a marketplace”
YouTube Unverified1 source· Talk








