6. Applications
The User Interface
The top layer where end-users interact with AI capabilities. These are the tools that solve specific problems, abstracting away the underlying complexity of the models and infrastructure.
5. Harnesses & Tooling
The Integration Layer
Bridges the gap between raw models and functional apps. Provides frameworks, libraries, and APIs developers need to build, fine-tune, deploy, and orchestrate complex agentic workflows.
4. Models
The Intelligence
The core "brain" of the stack. Large neural networks trained to recognize patterns, reason, make predictions, and generate multi-modal content based on massive computational ingestion.
3. Data
The Fuel
The critical layer that models ingest to learn. Involves massive infrastructure for collection, storage, cleaning, vector embeddings, and increasingly, generating high-fidelity synthetic data.
2. Infrastructure
The Foundation
The massive computing environments required to orchestrate thousands of chips. Includes hyperscale cloud clusters, advanced cooling, and incredibly fast networking fabrics (InfiniBand).
1. Chips
The Hardware Engine
The fundamental physical layer. Highly specialized microprocessors designed to handle the massive parallel computations (matrix math) necessary for training and running AI models.
Deployment Archetypes
How organizations traverse the AI stack: from building the raw infrastructure to consuming high-level APIs.
Self-Hosted
Running open-weights models (e.g., Llama 3) on your own metal or dedicated cloud compute instances.
- Absolute Privacy: Data never leaves your security perimeter.
- Total Control: Deep fine-tuning and hardware optimization capability.
- High CapEx: Expensive compute scaling and ongoing maintenance.
- Requires Talent: Needs MLOps and specialized infrastructure engineers.
Platform / Harness
Managed platforms (e.g., Azure OpenAI, Databricks) that bundle model hosting with orchestration, RAG, and vector tools.
- Faster Deployment: Built-in pipelines for embeddings, retrieval, and agents.
- Enterprise SLAs: Guaranteed uptime, security boundaries, and regulatory compliance.
- Balanced Cost: Pay for usage and premium tooling, avoid underlying hardware costs.
- Ecosystem Lock-in: Often tightly coupled to the cloud provider's specific toolchain.
Model API
Direct API consumption of proprietary frontier models (e.g., OpenAI API, Anthropic API) billed per token.
- Zero Ops: Completely serverless. No infrastructure or model hosting to manage.
- Frontier Access: Instant access to the most capable models on the market.
- Low Barrier to Entry: Extremely cheap and fast to prototype and validate ideas.
- Data Constraints: You are entirely reliant on the provider's privacy terms and rate limits.
Cost Architecture Calculator
Estimate relative monthly costs based on token volume and model capability tier.