Related Models#
The nine-layer model is a synthesis of established ideas, not a new theory. This page lists other public models of the AI stack and how this one relates to each. Every idea here exists somewhere already; what this model adds is the combination and its use as a co-design method.
| Model | Its layers or scope | How this model differs |
|---|---|---|
| AI is a five-layer cake (Jensen Huang, NVIDIA blog, March 2026) | Energy, chips, infrastructure, models, applications | An industry and economic view of where value is created. This model is an engineering view of how the middle of the cake works: energy maps to the power-and-cooling concern, chips to L8–L9, infrastructure to L6–L8, storage and the control plane, models to L2–L5, and applications to L1. |
| The 5 Layers of the AI Infrastructure Stack (James Fahey) | Energy and physical infrastructure, compute, model and training frameworks, platform and agents, applications | A business and market view. This model is an engineering view; energy appears here as the cross-cutting power-and-cooling concern. |
| AI infrastructure stack (Mirantis) | Foundation, data and storage, software and frameworks, orchestration and deployment | Closest in spirit (platform engineering). This model treats orchestration as a control plane beside the stack rather than a layer on top, and breaks the foundation into transport, fabric, interconnect and accelerator. |
| MLSYSIM architecture (MLSysBook) | Workloads, hardware, infrastructure, systems/topology, analytical solvers | A modeling framework for estimating performance. It has no separate software layer or control plane; this model is a design and implementation guide. |
| NVIDIA Enterprise Reference Architectures | Prescriptive hardware and software designs built on NVIDIA products | Product-specific blueprints. This model is vendor-neutral and organizes the concepts those blueprints implement. |
| Meta MTIA communication co-design (HOTI 2026 talk) | Silicon, network, collective library, compute kernels, co-designed for communication | The closest match to this model's co-design thinking, focused on communication. This model applies the same approach to the whole stack. |
| CNCF Kubernetes AI Conformance (summary by Luca Berton) | Platform requirements: accelerator device plugins, DRA, gang and topology-aware scheduling | Covers what this model calls the control plane. Its requirements map onto the control plane section here. |
What this model combines#
No single public model found during review (October 2026) combines all of these:
- Parallelism as an explicit design layer that sets the requirements for every layer below it.
- Compute runtime and kernels separated from communication libraries, because they fail and are tuned differently.
- Storage as a parallel data path, designed with the fabric rather than after it.
- Orchestration as a control plane beside the stack, not a layer inside the data path.
- A Kubernetes mapping of where each layer runs, and a workload-by-layer matrix showing which layers each workload stresses.
The review was not exhaustive. Suggestions of other models to compare against are welcome.