The five-layer AI cake, twice
Fable 5 · data curated through July 2026
Jensen Huang described a five-layer AI cake at Davos in January 2026. This atlas has used the name for a different cut of the same industry. Here is exactly where the two disagree — and the companies that settle it.
Two cakes, same number of layers
Speaking at the World Economic Forum in Davos in January 2026, NVIDIA's Jensen Huang described artificial intelligence as a five-layer cake: energy, then chips and computing infrastructure, then cloud, then AI models, and applications on top. NVIDIA published the framing on its own blog and it was reported widely.
This atlas has used the same name for a different cut of the industry. Both have five layers and three of them match. The two that do not are the interesting part, because each disagreement is a claim about how the industry actually works.
| # | Huang (Davos, Jan 2026) | This atlas | Same? |
|---|---|---|---|
| 1 | Energy | Silicon — chips, GPUs, TPUs, fabs | differs |
| 2 | Chips & computing infrastructure | Cloud & Compute | shifted |
| 3 | Cloud infrastructure | Foundation Models | shifted |
| 4 | AI models | Tooling & Orchestration | differs |
| 5 | Applications | Applications | same |
Read as sets rather than as rows: both contain chips, cloud, models and applications. Huang adds energy at the bottom. This atlas adds tooling between models and applications.
Difference one: energy
Huang puts energy at the foundation, and for an infrastructure argument that is the right place for it — a data centre is an electrical machine before it is a software one.
This atlas sorts companies by where they capture value, and an energy company does not capture AI value; it sells electricity, at electricity margins, to whoever pays. So energy is tracked here as infrastructure, beside data centre operators, networking silicon and submarine cables, rather than as a layer of the cake. The companies are in the atlas either way.
| Company | Influence | Why it is here |
|---|---|---|
| Constellation Energy | 6/10 | Restarting Three Mile Island for Microsoft. America's largest nuclear operator. |
| Kairos Power | 4/10 | Small modular reactors (SMRs). Google signed for 500MW. Deployment 2030. |
| X-energy | 4/10 | High-temp gas-cooled SMRs. Amazon backing + $500M Series C-1 2024. |
Difference two: tooling
The connective tissue — how developers wire models into real software. Huang's cut goes straight from models to applications. This one does not, because 24 companies in this dataset capture their value in the gap between the two — the SDKs, agent frameworks, orchestration and evaluation that turn a model endpoint into a product. A taxonomy that skips that layer cannot explain where those companies make money.
| Company | Also on | Influence | Why it is here |
|---|---|---|---|
| OpenAI | Foundation Models · Applications | 10/10 | ChatGPT, GPT-4, Sora, o-series. Defined the category. |
| Anthropic | Foundation Models · Applications | 9/10 | Claude, MCP, constitutional AI. Safety-first frontier lab. |
| Meta AI | Foundation Models | 8/10 | Llama, PyTorch — the open-weight anchor. |
| Hugging Face | Foundation Models | 8/10 | The GitHub of ML. 1M+ models hosted. |
| Vercel | Cloud & Compute | 7/10 | AI SDK, v0, AI Gateway. Frontend-first infra. |
| Claude Code | Applications | 7/10 | Agentic CLI. Killer app for coding agents. |
| Databricks | Cloud & Compute · Foundation Models | 7/10 | Acquired MosaicML ($1.3B). Enterprise AI plumbing. |
| LangChain | — | 6/10 | Agent/chain orchestration — loved and loathed. |
| Scale AI | — | 6/10 | Data labeling → RLHF vendor. $870M Meta investment (2025). |
| Qualcomm | Silicon | 6/10 | Bought Modular to fight CUDA with software. Silicon without a dev stack does not sell. |
| LlamaIndex | — | 5/10 | RAG-first data framework. |
| Pinecone | — | 5/10 | Managed vector DB pioneer. |
Which one should you use
They answer different questions. Huang's cake answers what does AI physically require, which is why it starts at energy and why it is an argument for infrastructure investment. This one answers who gets paid, and how much, which is why it starts at silicon and why the capital pyramid is its second claim: cloud pays silicon roughly thirty times what tooling pays models.
If you want the physical dependency chain, use his. If you want to know which company captures value where, this atlas maps 103 of them, with 155 dated events and 33 people, and it is CC BY 4.0, so you can take the data and check the argument yourself.
The five layers here, in full
- 1 · Silicon — Raw compute. Who makes the metal that every layer above depends on.
- 2 · Cloud & Compute — Where the chips live. Data centers, GPU clusters, and the contracts that route demand.
- 3 · Foundation Models — The model itself — architecture, weights, capabilities. The narrow funnel the whole stack aims at.
- 4 · Tooling & Orchestration — The connective tissue — how developers wire models into real software.
- 5 · Applications — The surface users actually touch. Where attention, subscriptions, and enterprise contracts accumulate.
Plus two companion dimensions: infrastructure (where energy lives) and physical AI.
Sources
- NVIDIA — Jensen Huang at Davos, World Economic Forum
- Forbes — Davos 2026: Jensen Huang on the five layer AI cake (22 January 2026)
- This atlas, in full — 103 companies, 155 events, 33 people
The Five Layer AI Cake is published by ConeApex, the trading name of Nirukta Digital XR Technologies Private Limited (CIN U62090TS2025PTC198275), Hyderabad, Telangana, India. The dataset is licensed CC BY 4.0 — cite it freely. See llms.txt for how to attribute it.