Cursor Hit $2B ARR With a Fine-Tuned Open-Source Model. Here's How to Build Your Own Vertical AI Business.

Cursor just revealed something most AI startups don't want you to know: you don't need to build a frontier model to compete with one.

Last week, Cursor released Composer 2 — a coding-specific AI model that matches Claude Opus 4.6 and GPT-5.4 on coding benchmarks. The pricing? $0.50 per million input tokens vs $5.00 for Opus. That's a 10x cost advantage.

Then the real story dropped. Cursor didn't build Composer 2 from scratch. They fine-tuned it on top of Kimi K2.5, an open-source model from Chinese AI lab Moonshot AI. Roughly 25% of the pretraining came from the base model. Cursor did the rest through domain-specific fine-tuning and continued training on code data.

The reaction was split. Some criticized Cursor for not disclosing the base model. But the business lesson is far more important than the PR misstep.

The $2 billion proof of concept

Cursor crossed $2 billion in annualized revenue in February 2026. The company is used by 67% of the Fortune 500. It generates 150 million lines of enterprise code daily. And its upcoming funding round reportedly values it at $50 billion.

All of this was built on top of models Cursor didn't create from scratch.

Why vertical beats general

OpenAI spends billions training GPT-5 to be good at everything. Anthropic does the same with Claude. This is rational for their business model — they're platforms selling general intelligence.

But general intelligence is expensive to run, expensive to access, and mediocre at domain-specific tasks. A general model processing a legal contract doesn't know what a "change of control provision" means in context. A general model analyzing medical images hasn't been trained on the specific pathology patterns that matter.

Cursor figured this out early. By training exclusively on code data, they built a smaller model that runs cheaper and performs better on the one task their customers care about. As co-founder Aman Sanger told Bloomberg: "It won't help you do your taxes. It won't tell you a bedtime story. But for code, it's competitive with models that cost 10x more to run."

This is the vertical AI thesis: sacrifice breadth for depth, and win on both performance and price.

The playbook for any vertical

Step 1: Choose your domain. The best verticals have specialized language, high accuracy requirements, and restricted data access. Legal, healthcare, financial services, construction, insurance, logistics, and agriculture all qualify. The key question: does a domain expert using your model outperform a generalist using GPT-5? If yes, there's a business.

Step 2: Secure domain data. Cursor has 150 million lines of code generated daily as training data. You need the equivalent for your vertical. Partner with companies. Offer free tools. Process public domain data. A legal AI company might start by processing every publicly available court filing. A financial AI might begin with SEC filings and earnings transcripts. The dataset compounds over time.

Step 3: Pick your base model. The open-source ecosystem in 2026 is remarkably strong. Kimi K2.5, Llama 4, DeepSeek V3, and Qwen 2.5 all perform near frontier levels. Evaluate each on tasks closest to your domain. Cursor chose Kimi K2.5 because it scored highest on code benchmarks among open-source options.

Step 4: Fine-tune with domain experts. Generic RLHF annotators won't cut it. Hire actual lawyers to evaluate legal AI output. Hire doctors to review medical AI suggestions. Hire financial analysts to validate compliance recommendations. Domain expertise in your training loop is what separates a toy demo from a production product.

Step 5: Build the product, not just the model. Cursor doesn't sell API access — it sells an IDE. The model is one component of a product designed for developers. Your vertical AI needs the same product thinking. A legal AI should be a contract review workspace. A medical AI should integrate into clinical workflows. The product creates switching costs that an API never will.

Step 6: Price to win. Your fine-tuned model runs at a fraction of the cost of frontier APIs. Pass those savings on to customers. Cursor charges $0.50/M input tokens vs $5.00 for Opus. This pricing advantage lets you undercut anyone building on top of frontier APIs while maintaining healthy margins.

The market opportunity

Enterprise buyers are shifting from general AI platforms to specialized solutions that understand their industry. Legal tech alone is projected at $35 billion. Healthcare AI at $45 billion. Financial services AI at $50 billion. In each vertical, the current solutions are either general-purpose models that underperform or legacy software that pre-dates the AI era.

Why frontier labs won't crush you

The natural fear is that OpenAI or Anthropic will release a legal-specific or medical-specific model. But their business model works against this. Frontier labs generate revenue by being platforms — selling general intelligence to millions of customers. Building vertical-specific models requires domain expertise, specialized data, and purpose-built products that don't scale across their entire customer base.

The real competition will come from other startups doing the same thing in your vertical. The winner will be whoever builds the best data flywheel fastest.

The Cursor lesson

Cursor's story has an important coda. When the Kimi K2.5 base was revealed, co-founder Aman Sanger acknowledged: "It was a miss to not mention the Kimi base in our blog from the start. We'll fix that for the next model."

But the business isn't suffering. Users don't care what base model powers their coding assistant. They care that it works, it's fast, and it's cheap. The same will be true for your vertical AI product.

The opportunity is clear: take the best open-source model available, fine-tune it on domain-specific data, wrap it in a purpose-built product, and sell it for less than frontier API access costs. Cursor built a $50 billion company this way. The playbook is now public.