Workloads
Workload requirements drive the design. This layer describes what the enterprise is trying to run and the demand imposed on every downstream layer. Build, customization, evaluation, optimization, serving, and promotion are lifecycle capabilities in the AI Platform, not workload families.
Explore ecosystem solutions for workload requirements.
Workload Requirements Record
This record is the input to software composition, capacity planning, hardware selection, support verification, and operational acceptance. Complete it before selecting products, then update it when later decisions expose a constraint.
Record the following dimensions before you select products:
Representative Workload Families
These examples help teams characterize demand. They are not an exhaustive list of enterprise AI use cases.
Agentic AI
Enterprise Agentic AI uses models, tools, memory, and enterprise context to plan and perform multi-step work under bounded authority, observable execution, and human-governed controls. Examples include customer-service assistants and workflow agents. Record plan and tool-call traces, persistent state, delegated authority, connector access, long-running failure modes, human review, revocation, and outcome evaluation. Use Secure Agent Workspace when autonomous work requires a governed execution environment.
Inference Applications
Interactive and batch applications consume production model endpoints. Record time to first token, inter-token latency, request and token throughput, context distribution, concurrency, availability, admission policy, routing, model mix, and cost objectives. Use Inference Hub when inference is offered as a shared enterprise service.
Model Development Workloads
Workloads that create or adapt model candidates characterize dataset throughput, model and optimizer state, checkpoint frequency, parallelism, recovery, artifact outputs, and experiment isolation. The AI Platform provides the build, customization, continuous evaluation, and promotion lifecycle around these workloads.
Physical AI
Robotics, digital-twin, video, and edge applications combine physical-world signals, simulation or perception, AI inference, and control loops. Record sensor and data rates, timing and safety constraints, edge or facility boundaries, simulation needs, connectivity, update cadence, and recovery behavior.
Each workload family is selective. Include only the families required by the approved business outcomes. Knowledge grounding, retrieval, and enterprise data access remain workload requirements and platform-service decisions, not a separate mandatory family. These examples do not make every associated product or lifecycle capability mandatory.
Output
The workload phase produces an approved workload requirements record. It identifies business and service owners, measurable objectives, representative test scenarios, known constraints, and inputs for selecting the software architecture and Enterprise Hardware Reference Architecture.
Example: An inference workload record can specify model and context ranges, peak demand, latency objectives, availability objectives, data residency, tenant isolation, and approved regions. It can also record growth assumptions and the owners who approve changes. The final design records which architecture decisions satisfy each input.