industrial ai

Jensen Huang’s Five-Layer AI Stack—And the Missing Sixth: The Data Layer

Inkwelldata

10 Apr 2026 • 3 min read

Jensen Huang's five-layer AI framework — and the data layer that completes it.

Late last year, at a fireside discussion hosted by the Center for Strategic and International Studies in Washington, Jensen Huang outlined a clear framework for understanding artificial intelligence.

Speaking with Dr John Hamre, the discussion focused on national capability, industrial policy and the systems required to support AI at scale. In that context, Huang described AI not as software, but as infrastructure.

He introduced what he termed a five-layer AI cake:

Energy → Chips → Infrastructure → Models → Applications

The five-layer system

Each layer within this framework represents a domain that has been engineered and industrialised.

Energy is generated, distributed and managed. Chips are designed and manufactured through specialised supply chains. Infrastructure is built to support computation and connectivity. Models are trained and refined using that infrastructure. Applications deliver outcomes across industries.

Each layer operates with defined processes, standards and investment structures.

The omitted layer

Alongside these five layers, there is another layer that remains implicit.

The data layer.

Without data, none of the other layers can function. Models require data to learn. Applications depend on data to operate. Infrastructure exists to move and process it.

Its absence from the formal framework reflects how AI has developed to date.

In IT environments, data has largely been taken for granted. It has been collected, aggregated and often scraped from the web at scale. This has enabled rapid progress in model development, with limited emphasis on how data is sourced and governed.

In operational and industrial environments, data has evolved differently. It has been built up over decades through a wide range of acquisition processes, technologies and standards. As a result, it does not yet exist as a coherent, unified infrastructure.

The nature of data in operational systems

In environments such as energy, manufacturing, water and transport, data is generated and managed within defined contexts.

It is:

Produced by assets such as meters, sensors and control systems

Distributed across protocols including SCADA, DLMS, M-Bus and PLC environments

Organised across multiple systems, vendors and operators

Subject to ownership, regulatory and operational requirements

Dependent on timely access for operational use

This data is not immediately usable. It requires structuring, securing and governing before it can support AI systems.

Data as infrastructure

Within Huang's framework, each layer is engineered and operated with discipline.

Energy systems are designed and managed. Semiconductor supply chains are structured and optimised. Infrastructure is deployed and maintained. Models are developed and refined.

In real-world applications, data requires the same level of treatment.

It must be structured to ensure consistency. It must be secured to protect integrity and access. It must be governed to reflect ownership and compliance. It must be delivered reliably and at scale.

This places the data layer alongside the other layers of the system as a form of infrastructure in its own right.

Enabling the data layer

Inkwell Data's Platform-as-a-Service, Altior, is designed to enable this layer.

It provides a distributed execution model in which each device is represented as an independent, state-aware process, allowing data to be handled consistently at source.

Data is validated in real time for structure, timing and consistency before it is ingested, ensuring that downstream systems receive reliable inputs.

From an infrastructure perspective, this introduces several key characteristics.

Security

Data is controlled at source, with strict access policies and encryption applied throughout acquisition, processing and delivery. Identity and key management remain within the operator's control, ensuring full ownership of data and security operations.

Usability

Telemetry is translated into structured, machine-readable formats through configurable digital twin models. This allows data from different devices and protocols to be normalised and made directly usable by AI, analytics and digital systems.

Scalability

The platform operates as a cluster of distributed nodes. Capacity can be increased by adding nodes as required, enabling linear scaling without re-engineering or service disruption.

Reliability

Processing is fault-isolated at the device level, preventing failures from propagating across the system. Persistent queues, backpressure mechanisms and time-series storage ensure continuity of data flows and protection against data loss.

Through this architecture, data is not only collected but actively managed as part of the infrastructure, ensuring it is consistent, secure and available for use.

Completing the system

The five-layer cake provides a clear structure for understanding AI as infrastructure.

For this system to function effectively however, data must be handled with the same discipline as every other component. It is structured, secured, governed and delivered in a way that supports operational use.

Recognising the role of a data layer completes the framework. It reflects how AI systems are implemented in practice, particularly in environments where data must be actively managed as part of the infrastructure itself.

AI infrastructure extends beyond compute and models. It includes the systems that ensure data can be accessed, controlled and used with precision.

This is the role of the sixth layer.