A 'vibe coding' platform (tools that build apps from plain English prompts) that Wix paid $80 million for a year ago is training its own large language model, betting that owning the model beats renting AI from frontier labs like OpenAI or Anthropic.
Training your own AI model used to mean trillion-dollar infrastructure, tens of thousands of GPUs, and a research team the size of a small university. That bar has quietly dropped. Base44, a Bay Area "vibe coding" platform, meaning a tool that lets users build working apps from plain-English prompts, is rolling out its own proprietary large language model, called Base 1, to power its app-creation engine. The company frames it as the first app-creation platform to launch its own proprietary LLM, a claim that, even as marketing, points at a real structural question: at what point does an AI app-layer company's own user data become a sufficient training asset to justify the cost of building instead of renting from a frontier lab?
That question is bigger than Base44. It is the question every wrapper around OpenAI, Anthropic, or Google is going to have to answer in the next eighteen months.
The bet looks different depending on whose throat you measure from. For founder Maor Shlomo, the reason to own the model is stack-level optimization: latency, cost, and efficiency wins that a company renting a frontier lab's API cannot reach, because it cannot change what sits underneath. Wix, Base44's parent, paid roughly $80 million for the startup about a year ago, when Base44 was six months old and had an eight-person team. CEO Avishai Abrahami has publicly defended the custom-model direction at the Wix level. The acquisition has now grown into an in-house model rollout whose product surface is documented in the Base44 changelog.
The portable framework here comes from Jonathan Userovici, a partner at Headline and an early Mistral investor, who breaks defensibility for AI app-layer companies into three ingredients: data, distribution, and tech stack. In TechCrunch's reporting on Base44's launch, Userovici uses those three as the test for whether a wrapper should keep renting or start building. Base44's pitch is that its own prompt-and-build data, plus the parent company's distribution through Wix, plus its own model stack, clears the bar. Rival Lovable, the Swedish vibe-coding unicorn that closed a Series A last summer, is positioned in the launch conversation as a wrapper-style competitor still understood to rely on external frontier models.
The honest counter, which Shlomo himself surfaces in his founder essay on Base 1, is that this bet fails if the frontier labs simply out-train any specialist. Cursor, Anthropic's Claude Code, and xAI's Grok are the real competitive horizon, not Lovable. A wrapper that spends eighteen months and several hundred million dollars training a custom model only to be lapped by a frontier lab's next release has not built a moat. It has built a slower copy. Shlomo's claim that Base 1 will "outperform frontier models" is, on the record available today, an aspiration rather than a benchmarked outcome. No third-party evaluation of Base 1 against GPT, Claude, or Grok is cited in the launch materials, the wire distribution, or Base44's own product documentation.
So the test is portable, and it applies cleanly to the rest of the AI app layer. The companies most likely to follow Base44's path are the ones whose user behavior produces a training corpus a frontier lab cannot easily replicate, whose distribution already carries them past the cold-start problem, and whose product is bottlenecked by something a general model handles badly enough that the optimization tax is worth paying. The companies that should keep renting are the ones whose defensibility lives in UX, brand, or a workflow a general model already handles well enough that the marginal cost of ownership exceeds the marginal lift.
What to watch next: which other AI app-layer companies cross the user-data threshold where training their own model starts to make sense, and which decide the wrapper tax is still cheaper than the build tax. Base44 has placed the first public marker. The rest of the field now has a frame to argue with.