Glossary
AI data architecture
AI data architecture is the set of design decisions that fix where data for AI workloads lives, which stages read it, where derived data is kept and how metadata and security follow it. It is a plan rather than a system, and it outlasts the models, frameworks and products chosen to carry it out.
Decisions an architecture settles
An architecture is written before any product is bought, and it answers a small number of questions that are expensive to change later:
- Authoritative copy. Where the system of record for each dataset sits, and which copies are allowed to exist beside it.
- Derived data. Where cleaned training sets, embeddings, indexes and model checkpoints are stored, and who owns each one.
- Scaling boundaries. Whether data capacity grows independently of compute, or is bought in the same units.
- Cross-cutting rules. How identity, encryption, retention and lineage travel with data from stage to stage, rather than being configured again in each tool.
Separating data capacity from compute is usually the decision with the longest reach. GPU fleets are resized every year or two, while datasets only grow. A data layer built as scale-out storage grows by adding nodes and leaves compute purchases out of the question of how much data can be kept.
Three common patterns
| Pattern | How data reaches each stage | Typical consequence |
|---|---|---|
| Copy per stage | Training, retrieval and analytics each receive their own extract | Fast to start; copies drift apart and lineage becomes hard to prove |
| Shared data layer | Every stage reads from one storage layer, often a data lake | Fewer copies; the shared layer has to serve mixed access patterns at once |
| Federated | Data stays in its source systems and a catalog and policy layer spans them | Least movement; depends on consistent metadata in every source |
Most enterprises run a mix. The pattern matters because it sets how many copies exist when a regulator asks what trained a model, or when one person's records must be deleted. Large datasets are also costly to move, a pull described as data gravity, so designs that bring compute to the data tend to age better than designs that ship data to each tool.
How does architecture differ from infrastructure and platform?
The architecture is the drawing. AI data infrastructure is the storage, network and compute that physically carry it out. An AI data platform is a product category that packages part of that infrastructure with shared governance. The AI data pipeline is the sequence of stages data passes through, and AI data management is the ongoing governance that keeps records of it. A modern platform with no architecture behind it tends to reproduce the copy-per-stage pattern inside the product.
Architecture is also the term with the longest life. Infrastructure is refreshed on hardware cycles, platforms are replaced on contract cycles, and pipelines are rewritten when models change, while the decisions about authoritative copies and lineage carry across all of them.
Frequently asked questions
Who owns AI data architecture in an enterprise?
It usually sits across data engineering, infrastructure, security and the AI teams. Without a named owner and a shared diagram of where authoritative copies live, it defaults to whatever the first large project built.
Where do vector indexes sit in an AI data architecture?
A vector index is derived data built from source documents. It sits beside retrieval rather than at the storage foundation, and it can be rebuilt from the sources, which remain the system of record.