What are the five layers of the AI stack?

Energy, chips, infrastructure, models and applications. Electricity and cooling sit at the bottom; GPUs and accelerators sit on top of that; clouds and orchestration rent the silicon by the hour; models and their tooling run on that infrastructure; and applications are the software people actually open. Every AI cost you pay is one of these five layers billed back to you.

Which layers does Macrostack cover?

Layers three, four and five — infrastructure, models and applications — with 84 tools and 382 verified alternatives. We do not cover energy or chips, because there is no decision a reader can act on at those layers today. They are shown on the map anyway, because a map that only draws what we sell would be a brochure.

The map

The AI stack, in five layers

Q: Why does the layer matter when choosing a tool?

Because the substitutes are different at every level. At the application layer you are usually replacing a subscription with a self-hosted equivalent. At the infrastructure layer you are usually paying a convenience premium over an open engine you could run yourself. At the model layer the open weights have closed most of the gap. Knowing which layer a bill comes from tells you which kind of fix applies.

AI is a layered cake. Electricity at the bottom, chips above it, the infrastructure that rents those chips by the hour, the models that run on it, and — at the top — the software people actually open. Every bill you pay is one of these five layers charged back to you.

Almost every buyer's guide describes one slice of one layer. The few maps of the whole stack are published by companies that sell one of the layers. This one is drawn by nobody who sells any of them, and it is wired to 84 real decisions with 382 verified alternatives underneath.

L5 Applications

42 decisions · 183 alternatives

The software people actually open — the layer where AI stops being infrastructure and becomes a product.

The largest layer by headcount and the most crowded. It is also where subscription creep happens fastest, because every app is now $10 to $30 a seat a month and nobody audits the total.

Who dominates it: Microsoft 365 · Google Workspace · Notion · Adobe · Salesforce

Cursor alternatives ElevenLabs alternatives Amplitude alternatives Figma alternatives Shopify alternatives Retool alternatives all of layer 5 →

L4 Models & tooling

14 decisions · 70 alternatives

The models themselves, plus the gateways, vector stores, frameworks and observability that make them usable.

The fastest-moving layer, and the one where open weights have most closed the gap on closed ones. A decision made here in 2024 is almost certainly wrong now.

Who dominates it: OpenAI · Anthropic · Google DeepMind · Meta Llama · Mistral · Qwen

Local & Sovereign AI LLM & Agent Frameworks AI agent frameworks LLM Gateways & Routing LLM Observability & Evaluation Vector Databases & AI Search Embedding Models Fine-Tuning & Model Training RAG & Retrieval Platforms LLM Evaluation & Testing AI Guardrails & Content Safety Speech Recognition & Transcription Experiment Tracking & ML Ops Document AI & OCR

OpenAI API (ChatGPT) alternatives LangChain alternatives OpenAI AgentKit alternatives Portkey alternatives LangSmith alternatives Pinecone alternatives all of layer 4 →

L3 Infrastructure

28 decisions · 129 alternatives

The clouds, orchestration and plumbing that turn silicon into something you can rent by the hour.

This is where the money leaks. Nearly every recurring bill a technical team complains about — GPU hours, CI minutes, log ingestion, cluster management — is billed at this layer, and nearly every one of them has a free engine underneath the paid convenience.

Who dominates it: AWS · Google Cloud · Azure · RunPod · Vast.ai · GitHub

AWS GPU Instances alternatives Modal alternatives Replicate alternatives AWS (EC2 compute) alternatives Red Hat OpenShift alternatives Docker Desktop alternatives all of layer 3 →

L2.5 Compute

covered — dedicated guide

The companies that actually rent you the accelerators — neoclouds, GPU marketplaces, serverless inference and the hyperscalers.

Layer 2 tells you which chip. Layer 3 tells you which platform. Neither answers the question with the money attached: from whom do you rent it, and what happens to your workload if they are not there in two years. A multi-year GPU commitment is a credit decision wearing a cloud contract.

Who dominates it: CoreWeave · Lambda · Nebius · RunPod · Vast.ai · Together AI

Open the Computeguide →

all of layer 2.5 →

L2 Chips & silicon

covered — dedicated guide

The physical accelerators — GPUs, TPUs, custom inference silicon — and the fabs that make them.

The single scarcest input in the industry. Which chip you can actually get, at what price, decides what you can afford to run.

Who dominates it: NVIDIA · AMD · Google TPU · AWS Trainium · TSMC

Open the Chips & siliconguide →

all of layer 2 →

L1 Energy

covered — dedicated guide

The electricity and cooling that everything above it consumes.

A datacentre's real constraint in 2026 is not chips, it is megawatts and the grid connection to deliver them. Every price further up the stack is downstream of this number.

Who dominates it: Utility grids · Nuclear PPAs · Gas turbines · Solar + storage

Open the Energyguide →

all of layer 1 →

How to read this map

The substitution is different at every layer, and that is the whole point of drawing it. At the application layer you are usually replacing a per-seat subscription with software you host. At the infrastructure layer you are usually paying a convenience premium over an open engine that is already inside the product you rent — a managed Airflow is Airflow, a managed inference platform is very often vLLM. At the model layer the open weights have closed most of the gap and the decision is now about where you run them, not whether they are good enough.

Two layers are blank here and stay blank. We show them because leaving them out would make the map flattering rather than true, and because the shape of the whole cake explains prices further up it: a GPU hour costs what it costs because of megawatts and fab capacity, neither of which you can shop for.

Every figure on the pages below is verified against primary sources and dated. Rankings are merit-only — affiliate income never changes an order or a pick. See our methodology.

L5 Applications

L4 Models & tooling

L3 Infrastructure

L2.5 Compute

L2 Chips & silicon

L1 Energy

How to read this map

New swaps, worth your inbox.