📰 Key Takeaways

Wix acquired vibe coding platform Base44 for $80 million in just six months with an eight-person team. Now Base44 is launching its own large language model Base1, enabling users to build applications with natural language.

Base1’s training data comes from “tens of millions of real user interactions” on the platform—this proprietary dataset is its core competitive advantage. Founder Maor Shlomo pointed out that owning and training their own model allows deeper optimization in latency, cost, and efficiency, with the goal of eventually outperforming frontier general-purpose models.

There’s clear strategic logic behind this move: whether AI startups relying on external LLMs (like Swedish unicorn Lovable) have lasting moats is an increasingly heated industry debate. Headline general partner Jonathan Userovici believes AI startup defensibility depends on three factors: data, distribution channels, and tech stack—Base44 is strengthening its position in this direction.

But the threat doesn’t just come from competitors—frontier AI labs are also making moves into vibe coding. Both Cursor and xAI have joined SpaceX, and Claude Code itself has become a vibe coding player, giving Anthropic access to application scenario data and feedback loops. Userovici also notes that legal tech startup Harvey planned to train its own model before giving up—don’t underestimate how fast frontier general-purpose models are evolving.

💬 JudyAI Lab’s Take

After being acquired by Wix for $80 million, Base44 immediately launched its own trained model Base1. This timeline sends a clear signal: the core question for AI startups isn’t just “how good are the features,” but long-term survival.

Base1’s底气 (confidence/strength) comes from the platform’s tens of millions of real user interaction data—an asset that external models can’t directly replicate. The VC community’s framework for judging AI startup defensibility is becoming clearer: data, distribution channels, and tech stack, all three are essential. But training your own model isn’t a silver bullet: legal tech startup Harvey planned to train its own model before ultimately giving up, showing that the evolution speed of frontier general-purpose models is a variable you can’t underestimate. Cursor, xAI, and Claude Code have all entered the vibe coding space—so even if Base44 builds its own model, platform-level competition pressure still exists. What we’re watching in this case isn’t just the trained model itself, but that proprietary dataset behind it—that’s the part that’s truly hard to replicate.

If you’re building an AI product, it’s worth asking yourself now: do you have systematic accumulation of user interaction data? Can this data create an advantage that’s hard for others to replicate directly?

📅 Source Info

🔗 Further Reading