App on Top vs Model Itself
The most important distinction in AI: are you building an application that uses a model, or building the model itself?
The core difference
When people say "I'm building an AI product," they almost always mean one of two very different things:
Building on Top
You call an AI model via API (like Claude, GPT, or Gemini) and build your product around it. The model is a tool you use.
Codmir's chat, agents, and code review
GitHub Copilot, Cursor, Notion AI
Any chatbot, AI search, or recommendation engine
Cost: $0 to start (API pay-per-use)
Skills: Software engineering, prompt design, UX
Time: Hours to weeks
Building the Model
You collect data, design an architecture, and train a model from scratch (or fine-tune one). The model is the product you create.
OpenAI training GPT, Anthropic training Claude
Meta training Llama, Google training Gemini
Research labs, AI companies, specialized domains
Cost: $10K–$100M+ in compute
Skills: ML engineering, math, data science
Time: Months to years
The key insight
99% of AI products are built on top of existing models. You don't need to train GPT to build a customer support chatbot — you just need to call it with the right prompts and surround it with good software.
Training a model from scratch is like building a car engine from raw metal. Using an API is like buying a car and customizing it. Both are valid — but most people should start with the car.
The AI stack
Every AI product sits on a stack of layers. Understanding where you're working helps you decide what to build vs. what to buy.
The AI Stack — 5 Layers
Your Application
← You are hereThe product users interact with. Chat UI, code editor, mobile app.
Prompt Layer
← You are hereSystem prompts, instructions, templates that guide the model's behavior.
Fine-Tune Layer
Custom training on your specific data. Specializes the model for your domain.
Base Model
The pre-trained foundation. GPT, Claude, Llama, Gemini. Trained on the entire internet.
Infrastructure
GPUs, data centers, training frameworks. The physical compute layer.
Most developers work in the top two layers: building applications and writing prompts. That's where the value is for most products. The model and infrastructure layers are handled by companies like Anthropic, Google, Meta, and OpenAI.
When to actually build your own model
There are legitimate reasons to train or fine-tune your own model:
Proprietary data advantage
You have domain-specific data (medical records, legal documents, internal code) that no public model was trained on. Fine-tuning on this data gives you an edge.
Cost at scale
At millions of API calls per month, hosting your own model becomes cheaper than paying per-token. The breakeven is typically around $50K–$100K/month in API costs.
Latency requirements
You need sub-100ms responses (real-time applications, gaming, robotics). Self-hosted models eliminate network round-trips.
Privacy and compliance
Your data absolutely cannot leave your servers (HIPAA, SOC2, government). Running a model locally keeps everything on-premises.
Novel capabilities
You need something no existing model can do — a new type of output, a specialized reasoning pattern, or multimodal capabilities specific to your domain.
The middle ground: Fine-tuning
Between "just use the API" and "train from scratch," there's a powerful middle option: fine-tuning.
Fine-tuning takes an existing model (like Llama 3 or Mistral) and trains it further on your specific data. The model keeps all its general knowledge but learns your domain, your tone, your patterns.
$5–50
Typical fine-tune cost
1–4 hrs
Training time
100+
Examples needed
Fine-tuning is covered in detail in the Fine-Tuning Explained guide.
Decision framework
Use this quick decision tree:
Can an existing model (Claude, GPT, Gemini) do what you need with good prompts?
→ Yes: Build on top. Write great prompts, build great UX.
Does it need to know domain-specific things it wasn't trained on?
→ Try RAG first (feeding context at query time). If that's not enough: Fine-tune.
Do you need complete control, extreme speed, or a novel architecture?
→ Train your own (or fine-tune an open-source base model).
What's next?
The next article covers Fine-Tuning Explained — the practical middle ground that most teams should try before training from scratch.