Skip to content
#AI

The Truth About The AI Bubble (AI column)

#AI #Engineering #ML #Architecture #Software #Economics #Future

Another one. episode The Lightcone podcast from the guys at Y Combinator was about the AI bubble, so I watched it with great interest. The participants discussed the following topics:

1Anthropic became No1 among YC startups Startups from Winter 26 Batch YC is more likely to use Anthropic models than OpenAI. ​- Anthropic Claude: 52% (was ~20-25percent 2024)

  • OpenAI: fell from 90%+ to <50%
  • Google Gemini: 23% (was single-digit)

The authors' hypotheses about why Claude is ahead

  • Best model for coding ​- Enterprise market share: 32% vs OpenAI 25% Focus on safety and reliability for corporations
  • Targeted optimization for coding (northstar eval by Tom Brown, co-founder Anthropic)

2Vibe Coding has become mainstream It looks like Development through task description in natural language LLM Generate code without detailed review Focus on iterations and results rather than code structure

Popular tools: ​- Cursor (VS Code + GPT-4o/Claude) Claude Code by Anthropic ​- GitHub Copilot, Lovable, Replit, Bolt

3AI economy stabilized According to partner Jared Friedman, The most surprising thing for me is how much the AI economy has stabilized. We have model layer, application layer and infrastructure layer companies. Everyone seems to be making a lot of money, and there’s a relatively straightforward playbook to build an AI-native company on top of models.

What's changed: Previously, every few months new model releases made completely new ideas possible. Now the search for startup ideas has returned to “normal complexity”

4Models turn each other into commodities Startups build an orchestration layer and switch between models:

  • Use Gemini. 2.0 context engineering Transmitted to OpenAI for execution Model selection based on proprietary evals for specific tasks Intel/AMD is a competition between architectures, but users can replace them.

What does that mean? Value shifts from models to application layer Model companies commoditize each other Application-layer startups get an advantage

5AI Bubble is good news for startups The guys remember the dot-com bubble, which led to investments in infra, and then on top of that came the conditional YouTube and could exist - cheap bandwidth was the result of a bubble. Now:

  • Big companies. (Meta, Google, OpenAI) Invest capex in GPUs and data centers If demand falls, it’s their capex, not startups. The infrastructure will remain and will be cheap. Guys remember about Carlota Perez frameworkThere are two phases: installation phase. (now), deployment phase (next x years). In the first phase of CAPEX spending, and in the second phase of creating economic value and the emergence of new bigtech companies ala Google.

6Space as an energy bottleneck solution As it turned out, there is not enough electricity on Earth for the AI bubble - there is nothing to power the AI data centers. And then the story. Summer. 2024StarCloud has proposed data centers in space → people laughed

  • 18 Months later, Google and Elon Musk do it

7More startups are making specialized models Harj notes the growing interest in creating smaller, specialized models in the latest YC batches:

  • Edge device models Voice models for specific languages
  • Domain-specific models Analogy: how in the early days of YC, knowledge about startups became common → explosion SaaS companies. Knowledge of training models becomes common knowledge.

#AI #Engineering #ML #Architecture #Software #Economics #Future