Data gravity never disappeared; however, AI has made it impossible to ignore. In earlier cloud analytics eras, by contrast, the cost and latency of moving large datasets were more tolerable. As a result, reports could run overnight, dashboards could load slowly, and systems still functioned. Today, AI workloads have removed that margin for error.
Specifically, modern AI systems require continuous, high-frequency access to large and often distributed datasets. At the same time, training, inference, evaluation, and retraining frequently occur across different environments and regions. Consequently, when data sits far from compute, performance degrades, costs rise, and failures become visible. In this sense, AI did not create a new infrastructure problem; rather, it exposed an existing one at scale.
AI intensifies data gravity in several ways. Models depend on fresh, near-real-time data, making delays immediately harmful to accuracy. As a result, pipelines repeatedly access the same datasets rather than pulling them once. Meanwhile, AI systems generate new data which then become future inputs and compound dependency over time. The result is persistent pressure on data movement across clouds, regions, and platforms.
Enterprises are now feeling the operational impact. Duplicating data to run models closer to compute increases storage footprints and network costs. Distributed pipelines introduce fragility, where delays in one system cascade downstream. Governance becomes harder as data spreads across environments with different policies, controls, and audit requirements. Data teams spend more time managing movement than improving models, slowing delivery velocity.
In response, organizations are shifting toward compute-on-data architectures. Instead of moving data to centralized pipelines, they push compute closer to where data already resides. Training and inference increasingly run inside lakehouse and warehouse platforms, with regional or edge inference reducing latency. Federated access layers allow models to operate across distributed datasets without full consolidation.
This data-centric approach treats locality as a design principle. Models adapt to data location, not the reverse. As analytics, storage, and AI platforms converge, compute-on-data is emerging as the most practical way for enterprises to scale AI while controlling cost, reliability, and complexity.
Source:
