The Five-Layer AI Stack Is an Inverted Pyramid

TL;DR

• Jensen Huang’s five layers are energy, chips, infrastructure, models, and applications. The dependency chain is clear, but it is not a map of market power.
• Participation expands upward: a few physical platforms support many models and far more applications.
• Power concentrates at bottlenecks below and distribution gateways above. TSMC, ASML, three global clouds, and a short list of digital gatekeepers make that concentration visible.
• Applications create the customer value. That does not mean application vendors keep most of the profit.

AI looks like software at the top and heavy industry at the bottom.

Jensen Huang describes AI as a five-layer cake: energy, chips, infrastructure, models, and applications. Each layer depends on the one below it. An application calls a model. The model runs on computing infrastructure. The infrastructure contains chips. The chips consume electricity.

That is the right dependency map. It is the wrong shape.

My read is that the stack behaves like an inverted production pyramid. Participation expands as you move upward. A small number of physical platforms support more models, which support far more applications and industry workflows.

Then the pyramid meets a second bottleneck. Thousands of applications still have to pass through a small number of distribution systems to reach customers.

The stack widens in solution variety. It narrows around production capacity. It narrows again around customer access.

Width Is Not Power

An inverted pyramid only works after defining what its width represents.

These measures do not move together.

A layer can contain thousands of products while a few suppliers control its inputs. An application can create the customer outcome while model, cloud, chip, data, and distribution providers collect part of the profit. More companies at the top does not automatically mean more competition or more economic power.

This distinction is the key to reading the five layers.

Energy Is Broad Globally and Narrow Locally

Energy looks like the broadest physical market. Countries have thousands of utilities, generators, fuel suppliers, and grid operators. It is not a global oligopoly.

But AI infrastructure does not need electricity in the abstract. It needs deliverable power at a specific site, on a specific schedule, with transmission, transformers, cooling, permits, and financing already aligned.

The International Energy Agency estimates that data centers consumed about 415 terawatt-hours of electricity in 2024, or 1.5% of global consumption. Its base case reaches roughly 945 TWh in 2030. That global share can look manageable while individual data-center regions face severe local constraints.

The hidden bottleneck can be an ordinary grid component. The US Department of Energy reports that distribution-transformer lead times expanded from three to six months in 2019 to 12 to 30 months in 2023.

So the bottom of the pyramid is not “a few energy companies.” The narrow point is the combination of powered land, grid access, equipment, and permission to build.

Large companies such as NextEra Energy, Constellation, and GE Vernova supply generation and grid technology. Startups including Fervo Energy and Oklo are pursuing new generation paths. But a power announcement is not operating capacity. Licensing, construction, fuel, transmission, and interconnection still determine when a project becomes real.

Chips Contain the Narrowest Global Chokepoints

The chip layer contains many products but only a few qualified suppliers at critical stages.

NVIDIA and AMD sell merchant accelerators. Google, Amazon, and Microsoft design custom chips for their own platforms. Startups such as Cerebras, Groq, and Lightmatter attack different parts of compute and interconnect performance.

That design activity still funnels into a concentrated manufacturing chain.

The US Department of Commerce stated in April 2024 that TSMC manufactures more than 90% of the world’s leading-edge logic chips. ASML describes extreme-ultraviolet lithography as technology unique to ASML. High-bandwidth memory comes from a small supplier group led by SK hynix, Samsung, and Micron.

This is the inverted pyramid at its clearest. A large application market can rest on one to three scaled suppliers at several manufacturing steps.

Custom silicon can weaken one accelerator vendor’s position. It does not remove the dependence on foundries, lithography, memory, packaging, and materials.

Infrastructure Is Several Markets Compressed Into One

Huang’s infrastructure layer includes land, power delivery, cooling, construction, networking, and orchestration. Economically, that combines at least three different markets:

1. Physical facilities: powered land, buildings, substations, cooling, and colocation.
2. Compute systems: servers, racks, storage, networking, schedulers, and inference software.
3. Cloud services: capacity sold through APIs, managed platforms, and long-term contracts.

The first two contain many specialists. Vertiv and Schneider Electric connect power and cooling. Equinix and Digital Realty operate data centers. CoreWeave, Crusoe, and Nebius sell specialized AI capacity.

Global cloud is much more concentrated. Synergy Research Group reported that Amazon, Microsoft, and Google held 63% of worldwide cloud infrastructure spending in the first quarter of 2026.

The competitive unit is no longer a GPU or a data-center building. It is the ability to assemble powered land, cooling, networking, accelerators, software, financing, and paying workloads into an operating system at high utilization.

That favors companies able to coordinate several sublayers at once.

Models Form Their Own Hourglass

The model layer looks wide when counting model artifacts. A Data Provenance Initiative study examined 851,000 models on the Hugging Face Model Hub, representing 97.6% of downloads in its dataset.

Frontier model development is far narrower. OpenAI, Anthropic, Google DeepMind, and a small set of other laboratories train leading general-purpose systems. Around them sits a much wider ecosystem of open-weight models, fine-tunes, quantizations, routers, and specialized models.

That produces an hourglass inside the pyramid:

Many research projects and model artifacts
          Few frontier laboratories
Many hosted, adapted, and specialized variants

Model commoditization and model concentration can happen at the same time. Routine capabilities become cheaper and easier to substitute. Producing the strongest new general-purpose model can still require more capital, compute, data, and research talent.

Applications Expand, Then Distribution Narrows

Applications create the widest solution surface because industries do not share one universal workflow.

Cursor and GitHub Copilot target software development. Abridge works on clinical documentation. Harvey focuses on professional services. Waymo applies AI to autonomous driving. Each category brings different data, regulation, trust, integrations, and measures of success.

That diversity lowers the value of a generic market map. It also explains why application companies can keep entering even when the lower stack concentrates.

But application creation and application distribution are different markets.

The European Commission currently lists seven Digital Markets Act gatekeepers: Alphabet, Amazon, Apple, Booking, ByteDance, Meta, and Microsoft. The designation covers 23 core platform services, including app stores, search, operating systems, social networks, marketplaces, browsers, advertising, and messaging.

The list is not a measure of AI application share. It shows that the routes through which digital products reach customers already concentrate around a short set of platforms.

An application can call any model and still depend on one operating system, productivity suite, app store, cloud marketplace, identity provider, or social network for distribution. The top of the stack therefore has a gateway, not unlimited width.

The Largest Companies Cross the Layers

The major platforms do not stay inside one layer.

Vertical integration can reduce coordination costs and improve complete-system performance. It also lets one company move value between layers. A platform can subsidize a model to sell cloud capacity, build a chip to lower inference cost, or bundle an application to protect distribution.

The five-layer cake describes the production system. The integrated platforms compete across the whole meal.

Applications Create Value, but Who Captures It?

Huang argues that applications are where the largest economic benefit appears. That claim holds: applications turn computed intelligence into completed code, clinical documentation, legal work, customer service, scientific discovery, or physical action.

It does not follow that application vendors keep most of the profit.

The application may create the customer outcome while cloud, model, data, chip, and distribution providers collect part of the revenue. Infrastructure suppliers can earn exceptional returns during scarcity. Distribution owners can bundle competing features. Model providers can move upward into applications, while application vendors can switch models or build their own.

The durable position is not necessarily an entire layer. It is a control point that other companies cannot easily route around:

• a physical bottleneck that takes years to reproduce;
• a software ecosystem that raises switching costs;
• a workflow with proprietary context and customer trust;
• a distribution surface that controls access, identity, or billing.

What the Five-Layer Model Misses

Jensen’s model leaves two important forces implicit.

Data moves through several layers. It trains models, grounds applications, creates regulatory obligations, and becomes proprietary context inside a workflow.

Distribution sits above applications. It determines whether a product can reach customers, earn attention, establish identity, and collect payment.

Adding those forces does not invalidate the five-layer cake. It explains why the number of participants and the location of market power tell different stories.

So What

The inverted pyramid is a strong map of dependency and participation. It is a weak map of profit until the bottlenecks and gateways are drawn onto it.

For builders and investors, the question is not only, “Which layer are you in?” The sharper question is, “What do you control that the layers above or below cannot easily replace?”

The open question is whether applications eventually retain more of the value they create, or whether model, cloud, chip, energy, and distribution providers continue to collect most of it.

The answer may differ by industry. Coding tools can reach users directly. Healthcare and legal products depend on regulated systems of record. Consumer applications can scale quickly while remaining exposed to mobile and web distribution gates. Physical AI adds manufacturing and safety constraints back into the top layer.

One inverted pyramid may eventually become a different market map for every industry.