The Latency Gradient Across Layers
From sub-10-microsecond actuation at L1 to multi-hour Monte Carlo at L0, the stack spans nine orders of magnitude in time — deliberately.
Nine orders of magnitude
The defining quantitative feature of the architecture is its latency gradient. Each layer operates in a characteristic time band, and the bands span from microseconds at the control plane to hours at the compute foundation. The stack is organized so that time budgets tighten monotonically toward the hardware.
- L1 Control Plane — sub-10 microseconds, hard deterministic.
- L2 Data Fabric — sub-microsecond sampling, aggregated at milliseconds.
- L3 Twin & AI — 50-100 ms predictive shadow; MPC within the control cycle.
- L4 Orchestration — milliseconds to seconds for routing and rule checks.
- L5 Copilots — seconds for interactive reasoning.
- L6 Experience — human-interactive, tens to hundreds of milliseconds for display.
- L7 Ecosystem — minutes to hours for dispatch and scheduling.
- L0 Foundation — minutes to hours for batch training and Monte Carlo.
Why the gradient must be monotone toward hardware
The closer a function sits to the plasma and magnets, the tighter its time budget, because the physics it guards evolves fast. A breeder disruption develops in milliseconds; magnet protection must respond in microseconds. Placing the tightest budgets at L1 and relaxing upward means the layers that can afford to think slowly are the ones farthest from irreversible harm.
The edge-to-cloud shape
The gradient is physically realized as an edge-to-cloud continuum: FPGAs at the edge for microseconds, on-site servers for the twin, and cloud HPC for batch work. The edge-to-cloud continuum page details how compute placement follows the time budget.
The gradient is enforced by the line semantics and underlies the real-time versus offline boundary.