The framework

Seven Layers of AI

The complete framework moves from products people use to agents, models, infrastructure, hardware, and AI-enabled scientific discovery.

1 AI Industry System Map Hardware, infrastructure, training, inference, models, agents, applications, and feedback loops

How the industry fits together

The AI Industry: Layers, Supply Chain, and Feedback Loops

AI is not one company or one model. It is a connected industrial system. Physical equipment and energy support chips; chips support data centers; data centers train and serve models; models power agents and applications; use, evaluation, and scientific discovery feed improvements back into the system.

EquipmentASML

EUV lithography systems used to pattern advanced chips.

Accelerator designNVIDIA

GPU and networking designs, CUDA software, and AI systems.

MemoryMicron

High-bandwidth memory and storage keep model data moving.

Fabrication & packagingTSMC

Manufactures leading-edge logic and packages chips into high-performance systems.

Other critical inputsPower, cooling, networking

Electrical grids, data-center construction, optical links, servers, and storage.

Compute systems and cloudAI data centers

Azure, AWS, Google Cloud, Oracle, CoreWeave, sovereign clouds, and private clusters combine accelerators, memory, networks, storage, software, and energy.

Training path

OpenAI training infrastructure

Large distributed jobs process data, update model weights, create checkpoints, run evaluations, and perform post-training. The output is a trained model.

Inference path

Model-serving and inference systems

The trained model is loaded onto serving infrastructure. Each request consumes compute to produce a new answer, prediction, image, action, or tool call.

ModelGPT, Claude, Gemini, Qwen, Mistral, DeepSeek

Learned capabilities encoded in model weights and architecture.

AgentModel + tools + memory + permissions

Plans, calls tools, maintains state, and completes multi-step work.

ApplicationChat, coding, medicine, finance, robots

The user-facing product, workflow, or machine where value is delivered.

Layer 1Applications

Products and interfaces used by people and organizations.

 Layer 2Agents

Planning, tools, memory, state, permissions, and execution loops.

 Layer 3Domain Models

Specialized intelligence for medicine, law, finance, science, and other fields.

 Layer 4Model Architecture

Transformers, multimodal systems, training methods, reasoning, and model design.

 Layer 5Infrastructure

Data, training clusters, inference serving, clouds, networking, storage, and operations.

 Layer 6Hardware

Semiconductor equipment, accelerator design, foundries, memory, packaging, power, and cooling.

 Layer 7Scientific AI

A cross-cutting discovery system that uses every other layer—and produces new knowledge, data, materials, medicines, and engineering methods that feed back into them.

One company may occupy several layers

NVIDIA spans chip design, systems, networking, software, models, and robotics. OpenAI spans training infrastructure, models, inference services, agent platforms, and applications. Google spans custom chips, cloud infrastructure, model research, agents, products, and science. The map shows functions—not rigid company boundaries.

2 Major AI Investment Projects — July 2026 A dated comparison of the factories, compute systems, models, agents, applications, and scientific programs building the industry

This cross-layer snapshot compares major projects using the metrics appropriate to each layer. Layer 5 includes a standardized table of GPU or accelerator counts, gigawatt capacity, and capital expenditure or commitments.

Open Major AI Investment Projects

3 Explore the Seven Layers Open the book-style chapters for Applications through Scientific AI