On October 19, 2026, NVIDIA announced a 104-qubit simulation of fault-tolerant quantum computing on JUPITER, a supercomputer at the Jülich Supercomputing Centre in Germany. The run used 8,192 NVIDIA GH200 GPUs to model magic-state cultivation under realistic hardware noise. It was a classical GPU simulation, not a computation carried out on a physical 104-qubit quantum processor.
What NVIDIA simulated
Magic-state cultivation is a method for preparing special quantum states used in fault-tolerant computing. NVIDIA said the simulated circuit was expressed with CUDA-Q’s fault-tolerant logical layer and run using cuQuantum. The distinction matters: the 104-qubit figure describes the modeled quantum workload, while JUPITER’s GPUs performed the simulation.
Technical coverage described two modeled noise effects: coherent phase errors and leakage. Leakage occurs when a superconducting qubit moves beyond its usual two computational levels into a third energy level. Including such effects makes the simulation a model of noisy hardware behavior rather than an ideal, noiseless circuit.
The reported computing footprint
Quantum Computing Report described a 45-bit encoded state representing about 35.2 trillion complex amplitudes. For the 8,192-GPU configuration, it reported 512 TiB of distributed GPU memory. These figures convey the scale of the classical computation used to represent the workload; they do not describe memory or qubits in a physical quantum processor.
How the result compares
NVIDIA called the JUPITER run the largest comprehensive simulation of a fault-tolerant quantum processor to date. The company compared it with an earlier 50-qubit Jülich simulation that used ideal, noiseless conditions. The newer simulation modeled realistic hardware noise, so the two qubit counts do not make the runs a like-for-like performance comparison.
Unite.AI reported separate checks on small packed circuits. In those cases, state-vector and tensor-network calculations produced measurement probabilities that agreed to about 10⁻¹³ when the tensor-network truncation threshold was tightened. Those checks covered small circuits, not the complete 104-qubit JUPITER simulation.
