Base44 Builds Its Own AI Model to Cut Costs and Escape Vendor Lock-In

Source: TechCrunch.

Base44, the natural language app builder that Wix acquired for $80 million last year, just rolled out its own large language model instead of relying on third-party AI providers. The move reflects a broader tension in the AI startup world: companies built entirely on someone else's infrastructure are vulnerable to pricing changes, model deprecations, and competitive pressure from the very providers they depend on.

The company's founder, Maor Shlomo, says owning the full stack allows Base44 to optimize for latency, cost, and efficiency in ways that using frontier models does not. Base44's first model, called Base1, was trained on tens of millions of user interactions collected through the platform. That means the company now controls its own compute and inference spending, which it expects will improve margins over time.

Why Vertical Integration Matters Now

Base44's decision arrives as AI companies face mounting questions about defensibility. If your entire product runs on Claude or GPT-4, what happens when your supplier raises prices or launches a competing feature? Anthropic already entered the vibe coding space with Claude Code. xAI now owns both Cursor and Grok through its SpaceX parent company. The providers are moving downstream fast.

Jonathan Userovici, a general partner at venture firm Headline, says AI startup defensibility comes down to three elements: distribution, data, and infrastructure. Companies with strong brands are now leveraging their user data and building their own models to control more of that stack. Base44 fits that pattern. It has brand recognition, a growing dataset from real usage, and now its own model trained specifically for app generation.

Other vibe coding platforms still rely on external models. Swedish competitor Lovable, which reached unicorn status during its Series A, uses third-party LLMs. Shlomo expects that will change as companies reach enough scale to justify the engineering investment, but he believes specialization gives Base44 an edge over general-purpose models.

The Cost Problem Driving Change

Inference costs have become a real line item for AI product companies. Enterprise customers, who represent a growing share of Base44's revenue, are pushing back on the expense of running frontier models for every task. Many are demanding orchestration systems that route queries to cheaper models when high performance is not critical.

Userovici points to legal tech startup Harvey, which abandoned plans to train its own model, as a cautionary example. Not every applied AI company will become a model lab. But the broader shift toward cost optimization is real. Companies are building infrastructure to select the right model for each use case rather than defaulting to the most expensive option.

Base44's move addresses that pressure directly. Shlomo says the company wants a model that aligns with what users actually prefer in results, runs faster, and costs less than using frontier models like Claude Opus. For Base44's parent company, improved margins would be welcome news. Wix recently laid off 20% of its workforce, while Base44 has been hiring and growing. The platform passed $150 million in annual recurring revenue in May, two months after crossing $100 million.

Scale Remains the Open Question

Base44's ARR still trails Lovable, which reported $500 million in annual recurring revenue earlier this year. The gap raises a question about whether owning your own model matters if competitors achieve faster distribution using off-the-shelf infrastructure. Shlomo is betting that vertical integration will prove more durable, but the experiment is just beginning.

The first iteration of Base1 is rolling out now. Whether it actually outperforms frontier models for app generation will depend on how quickly Base44 can iterate using its proprietary dataset. The company says it is the only vertically integrated vibe coding platform, controlling distribution, data, and infrastructure simultaneously. That claim will be tested as frontier labs continue building better coding agents and competitors scale their own operations.

Bottom Line

If you run a product on rented AI infrastructure, you are exposed to pricing changes and competitive encroachment from your own suppliers. Base44's bet on a custom model shows one path toward reducing that risk, but it requires significant engineering resources and enough user data to make training worthwhile. For smaller operators, the takeaway is simpler: understand your inference costs now, and build contingency plans for when your model provider enters your market.