When an AI system performs poorly, the model may not be the only constraint. It can also be waiting for data that is slow, fragmented, difficult to govern, or costly to move into the right place. That problem exists at two distinct levels: data traffic inside the computing stack, and data movement among an organization’s lakes, warehouses, and specialized stores. Neither makes model capability irrelevant, but both can limit what an AI system can deliver.
What “data movement” means for AI
Data movement is not just copying files. An organization may need to bring information from application-specific stores into a lake for analytics, or send selected data from a lake into a specialized store for search or another workload. The right path depends on where data originates, where it will be used, and how quickly it must be available. AWS describes the increasing difficulty of moving growing datasets as “data gravity.” AWS explains data movement patterns.
As an Amazon Associate I earn from qualifying purchases.
Separately, data can bottleneck computation itself. During model training, processors may spend time waiting for data to arrive from memory rather than performing calculations. That is not the same problem as moving data between enterprise systems, and solving one does not automatically solve the other.
Where data can get stuck in an organization
AWS groups common enterprise movement into three patterns. These describe where data travels, not a ranking of architectures; a company may use more than one.
Inside-out: from a lake to specialized stores
Data is aggregated in a lake, then selected material is sent to purpose-built stores for workloads such as search or knowledge graphs. This can make information available to a particular application, but requires deciding what to move and maintaining the resulting paths. AWS’s movement-pattern overview.
Outside-in: from application stores to a lake
Application-specific stores are the starting point; data is moved into a lake for broader analysis. The challenge is making that data usable together while accounting for freshness, quality, lineage, and the demands placed on source systems. AWS’s movement-pattern overview.
Around the perimeter: between specialized stores
Data moves directly between purpose-built systems for a particular workload—for example, copying relational data to a NoSQL store to support dashboards. This may suit a focused use case, but creates another integration path to secure and operate. AWS’s movement-pattern overview.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Why pipeline choices affect AI data access
Data pipelines can process accumulated records at intervals or handle events as they arrive. Batch processing can suit workloads that do not need every update immediately; streaming can support fresher data, but requires systems designed to process an ongoing flow. The choice should follow the workload’s freshness and latency needs, not a general assumption that one approach is always better. AWS identifies performance, latency, scale, security, and governance as ingestion design considerations. AWS overview of data ingestion.
Teams also choose where transformation happens:
- ETL (extract, transform, load): data is transformed before it is loaded into the destination. This requires transformation tooling outside the warehouse, while reserving warehouse resources for queries.
- ELT (extract, load, transform): data is loaded first and transformed inside the warehouse. This can simplify the pipeline, but transformation work uses warehouse resources.
Freshness, latency, source-system impact, governance, and available warehouse resources all inform the choice. Neither ETL nor ELT is a universal answer. AWS overview of data ingestion.
Data movement can limit training performance too
Enterprise data integration is only one layer of the problem. In a 2020 study, Andrei Ivanov, Nikoli Dryden, Tal Ben-Nun, Shigang Li, and Torsten Hoefler found that memory-bound operators consumed 37% of a BERT training iteration. In that study, their method reduced data movement by up to 22.91% and improved whole-BERT performance by 1.19×. These are results for the authors’ BERT study, not general figures for all models, training workloads, or enterprise data systems. The 2020 paper on reducing data movement in BERT.
The distinction matters when diagnosing a slow system. A training job waiting on memory traffic calls for a different investigation from an AI application that cannot retrieve current, relevant business records from organizational systems.
Recommended Free Tools
How to identify the constraint in your AI system
Start with the observed failure rather than assuming the model is at fault—or assuming data movement explains everything. Check the path from source to use, and separate retrieval delays from compute-side delays.
Best Value
- Family farms not data design for people against AI server farms, data center expansion, rural land buyouts, corporate agriculture, and industrial tech development replacing farmland and open space. Rural conservation and anti data center message.
- AI protest design for farmers, land conservation supporters, anti AI activists, sustainability groups, environmental advocates, rural communities, and people opposing server farm construction, power grid strain, and farmland destruction.
- Lightweight, Classic fit, Double-needle sleeve and bottom hem
- Define the workload’s freshness and latency needs. Establish how current the data must be and how quickly the system must respond. This determines whether periodic batch delivery is adequate or event-by-event streaming is warranted.
- Trace where the required data lives and where it is used. Map the source applications, lakes, warehouses, and specialized stores involved. Identify each transfer or transformation between them.
- Measure throughput and source-system impact. Determine whether data arrives at the needed rate and whether extraction or queries burden systems that support operational applications.
- Check quality, lineage, security, and governance across each handoff. A fast transfer is not useful if records cannot be trusted, traced, or accessed under the right controls.
- Account for operational complexity and total cost. Include pipeline maintenance, infrastructure and warehouse consumption, and the engineering effort needed to monitor and recover data flows.
- Inspect model-side compute separately. If the issue is training performance, investigate memory and operator behavior rather than treating enterprise ingestion as the cause.
These checks help distinguish a model limitation from a data-access or data-movement constraint. They also reveal when both are contributing.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Why scaling AI can expose infrastructure gaps
In a July 2026 blog, Google Cloud reported survey results in which 83% of organizations said they needed infrastructure upgrades for production-grade agentic AI; 43% of IT leaders cited legacy integration difficulty as a leading infrastructure gap; 81% identified operational complexity and engineering overhead as unforeseen scaling expenses; and 36% cited a lack of specialized high-throughput vector databases. These are Google Cloud’s vendor-reported survey figures, not independently established measures of all organizations or workloads. Google Cloud’s July 2026 discussion of agentic AI infrastructure gaps.
In the same blog, Google Cloud product executive Sirish Chandrasekaran said: “What’s one of the biggest bottlenecks stopping organizations from scaling their AI initiatives? It isn’t the capabilities of today’s models — it’s their access to business context and semantic meaning.” That is a vendor executive’s view, not a universal diagnosis. It captures a practical issue: a capable model cannot use information it cannot retrieve or interpret, but data access is only one part of system performance.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Choose an architecture for the workload, not a slogan
Moving AI “to the data,” centralizing everything, or adopting a particular data architecture is not automatically the right fix. The decision should reflect the workload’s needs and the cost and risk of operating its data paths. Compare candidate designs on:
- Freshness and latency: how current the data must be and how quickly it must reach the application.
- Throughput and scale: how much data must move, and whether the path can sustain that volume.
- Data quality and lineage: whether transformations preserve meaning and make the data’s origin and changes traceable.
- Governance and security: whether access controls and policies hold across stores and transfers.
- Source-system load: whether extraction interferes with systems serving other workloads.
- Operational complexity and total cost: the infrastructure, engineering, monitoring, and recovery work required over time.
The useful outcome is not the architecture with the fewest transfers in the abstract. It is a data path that meets the workload’s needs without creating unacceptable latency, governance risk, source-system impact, or operational burden.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
