Base44 training its own model is a live test of a question every applied AI company is asking: is the wrapper business defensible, or does owning the model become necessary once you have enough data and scale?
- Base44, the Wix-owned vibe coding platform, is rolling out its own AI model, Base1, trained on tens of millions of real user interactions from its platform.
- Founder Maor Shlomo says owning the model gives Base44 more control over latency, cost, and efficiency than relying on frontier models like Opus.
- Base44 passed $100 million in annual recurring revenue a few months ago. Competitor Lovable, which relies on external LLMs, hit $500 million ARR earlier this month.
- Jonathan Userovici, a general partner at Headline VC, says data, distribution, and tech stack are the three pillars of AI startup defensibility, and warns against underestimating frontier models.
- Wix laid off 20% of its workforce in May 2026, citing AI and exchange rate pressures, making margin improvement at Base44 strategically important for its parent company.
What Happened
Base44, the vibe coding platform Wix acquired for $80 million in 2025 when the company was six months old with a team of eight, has started rolling out its own AI model, Base1, to support natural-language app creation, TechCrunch reported.
Base1 was trained on a dataset built from tens of millions of real user interactions on the platform. Founder Maor Shlomo said owning the model as part of the full stack gives Base44 more room to optimize for latency, cost, and efficiency than relying on third-party frontier models. He expects other AI startups with enough scale and data will follow the same path.
The move puts Base44 in contrast with Swedish competitor Lovable, which relies entirely on external LLMs and recently hit $500 million in annualized revenue. Base44 has passed $100 million in ARR. Jonathan Userovici, general partner at VC firm Headline, told TechCrunch that data, distribution, and tech stack are the three core ingredients of AI startup defensibility, and that companies with strong brands are now leaning into data and infrastructure ownership to build that moat.
Userovici cautioned against assuming this works for everyone, pointing to legal tech startup Harvey, which abandoned its own model training plans. He framed Base44’s move inside a broader cost pressure trend: enterprise customers increasingly demand orchestration and model-selection infrastructure so they are not paying frontier model prices for every task.
Why It Matters
Base44’s bet tests a real strategic question for any applied AI company: at what point does enough proprietary usage data justify the cost of training your own model instead of staying a frontier model wrapper.
The competitive pressure is intensifying from two directions at once. Frontier labs are moving into vibe coding’s territory directly, with Cursor and xAI now both under SpaceX and Claude Code becoming a vibe coding competitor in its own right. That gives Anthropic and other foundation labs access to the same kind of app-creation feedback loops Base44 is using to train Base1.
The economics are unproven. Base44’s own press release frames the benefit as a “structurally stronger margin profile over time,” language that signals the payoff is not immediate. Wix’s 20% layoff in May makes margin improvement at Base44 a real strategic priority for its parent company, which adds urgency but does not guarantee the model investment pays off on the timeline Wix needs.
Relve, an AI trends intelligence platform, is tracking whether vertically integrated applied AI companies can sustain a cost advantage against frontier labs entering the same product categories.
“Models are progressing, but they’ll stay very general in what they can do.” Maor Shlomo, Founder, Base44
Bottom Line
Watch whether Base1 actually beats frontier model performance on Base44’s core use cases within the next two quarters. Shlomo’s bet is that specialization beats generality for app creation specifically. If Base1 underperforms Opus or GPT-5.6 on real user tasks, the defensibility argument collapses and Base44 becomes a cautionary tale rather than a template.
For SaaS founders evaluating build-vs-buy on AI infrastructure, the Userovici framework is the practical takeaway here: defensibility comes from data, distribution, and tech stack together, not from any one alone. If you do not have Base44’s scale of proprietary usage data, training your own model is unlikely to produce a similar cost advantage. Orchestration and model selection across existing frontier APIs is the more realistic cost lever for most teams.
