The target is active work
The Deploy engineering target is 100,000 concurrent active users. No measured or guaranteed capacity claim exists for that target. Idle sessions, open sockets and sequential requests do not establish concurrent business-work capacity.
Define the workload mix
Separate chat requests, document retrieval, model generations, coding or browser tools, customer voice/text conversations and native record transactions. State how many are active at once, their context sizes and the duration and frequency of work.
Choose the actual infrastructure
Capacity depends on the selected application hosts, GPU and model configuration, storage, network and runtime. A minimum-memory application host or single-GPU fit calculation cannot establish capacity for 100,000 active users. No arbitrary single server is promised to support that number.
Set service levels
Agree request and completion latency, acceptable errors, queueing, fairness between users and record durability. Measure each layer under the declared mix instead of treating a fast model response as proof of a complete task.
Include failure and recovery
Consider lost nodes, model interruption, retry storms, disk pressure and rolling upgrades. Permissions, action identity, budgets and truthful records must survive these conditions. A partially completed action needs reconciliation, not silent replay.
Qualify the deployment profile
The operating profile must identify hardware, model, context limits, workload mix and service levels. Measured evidence is required before a capacity commitment can be made. The target is retained as product scope; it is not a benchmark result or service guarantee.
Capacity overview → / Discuss your organisation’s workload →