Kronos Fusion Energy · Computing & ControlThe software that flies the machine.
A fusion plant is a control problem as much as a physics problem. Kronos runs on an AI-native control architecture — eight layers, a full stack of compute features, all on owned hardware. Two machines, one architecture: the breeder makes the fuel, the burner burns it, and the same stack runs both.
Phase 1 is complete. The de-risking computing campaign is run and closed — the design is proven and every frozen number held. What follows is the build.
A four-phase build
Where the control work lands — from a completed de-risking campaign to a fleet.
Phase 1 · 2026
COMPLETE
The de-risking campaign is run and closed; the core control features are built or de-risked. The design is proven — no frozen number moved.
Phase 2 · 2027–29
Build-out
The AI-control build-out — closed-loop on the twin, hardened onto the machine as the first plant is built.
Phase 3 · 2030–31
First plant
First plant — first tritium, commissioning, the live twin.
Phase 4 · 2036+
Fleet
Fleet — federate and orchestrate a growing fleet; the lunar helium-3 fuel loop.
Eight layers, one stack
Every diagram is drawn the same way — eight numbered layers from owned compute to the markets the plant serves.
L0
Foundation Layer
Kronos-owned GPU clusters & bare-metal (NVIDIA A100/H100-class) — the ground everything stands on, in-house.
L1
Infrastructure Layer
Edge FPGA and deterministic real-time compute — the hard loop.
L2
Data & Context Layer
Feature store and vector database — the plant’s working memory.
L3
AI Capabilities & Models
Graph neural nets, physics-informed nets, anomaly ensembles, MPC — the model layer.
L4
AI Workflows & Orchestration
Event streaming, workflow engine, rules, safety bounds, task routing.
L5
Applications & Copilots
Agentic AI — plasma, engineering, ops and finance copilots.
L6
Experience Layer
Control room, spatial twin, mobile — the human surfaces.
L7
Ecosystem Layer
Outcomes, partner APIs, markets — the plant as a node in the world.
Built to be trusted
A 50–100 ms predictive shadow
KRONOS-CTRL runs coupled digital twins ahead of the real plant, so a correction leaves before the disturbance arrives — the difference between riding a disruption and preventing it.
A hardware-enforced safety clamp
No AI output reaches an actuator without passing a deterministic, hardware-enforced clamp. The AI advises, the boundary enforces, the human commands — auditable from the plasma to the metal.
On owned compute
Every layer runs on Kronos-owned hardware — an in-house NVIDIA A100/H100-class GPU HPC platform and deterministic real-time edge compute. Nothing critical is rented, and nothing is a black box.
The platform behind both machines is the S.M.A.R.T. generator — a Superconducting, Minimum-Aspect-Ratio Torus — a granted invention (US 12,009,112 B2), with its control architecture, KRONOS-CTRL, filed.