Nvidia says it completed a 104-qubit simulation on JUPITER using 8,192 GPUs, modelling a fault-tolerant quantum computer. The company calls it the largest full simulation of such a machine so far. The run took place on conventional computers, not a quantum processor.
What happened before this latest run?
Qubits are the building blocks of quantum computers, and the source says they are sensitive to interference. Even small amounts of noise from control equipment can create errors. A fault-tolerant machine is designed to detect and correct those errors while calculations are running.
Researchers at Jülich had previously simulated 50 qubits on JUPITER in November 2025. That earlier work used ideal, noise-free conditions, according to the source. Nvidia says the newer simulation also accounted for noise affecting real quantum hardware.
What did Nvidia simulate?
The company modelled a 104-qubit fault-tolerant quantum computer using JUPITER, the supercomputer at Jülich Supercomputing Centre in Germany. It said the run was the largest full simulation of this type so far. The simulation was carried out on conventional computers with Nvidia GH200 chips spread across 8,192 GPUs.
The point of this kind of simulation is to let researchers test how a proposed design could perform before building the actual hardware. Nvidia said its team developed and tested the model first on a smaller system at its own quantum lab. The company said that preparation cut the work from weeks to days.
What does magic state cultivation mean here?
The team tested a technique called magic state cultivation. Special quantum states are needed for many operations, but making them can require many additional qubits, the source explains. Nvidia says its approach requires 90% fewer qubits than earlier approaches.
The source does not provide detailed measurements behind that comparison. It also does not describe the earlier approaches by name. The reduction is therefore a claim attributed to Nvidia, alongside the company’s statement that the run is the largest full simulation of this kind.
Why was JUPITER used?
JUPITER opened in September 2025 and was Europe’s first exascale supercomputer, according to the source. It can perform more than a billion billion calculations per second. Its role in this work was to run the large conventional-computing simulation, rather than to operate as quantum hardware.
The prior Jülich simulation and Nvidia’s new run represent different conditions: the earlier run was noise-free, while the latter included noise. That distinction matters because the source identifies hardware interference as a cause of quantum errors. The latest model therefore examines a design under conditions closer to those affecting real quantum equipment, as Nvidia describes it.
What remains to be established?
Nvidia CEO Jensen Huang said useful quantum computing will require complex fault-tolerant systems. Kristel Michielsen, director of Jülich Supercomputing Centre, said JUPITER gives researchers the ability to investigate questions previously beyond reach. The source gives no date for building a physical machine based on the simulated design.
It also provides no full technical results or independent assessment of the comparison with earlier approaches. The announcement came as Huang prepared to deliver a keynote at Nvidia GTC in Berlin on Wednesday. Further detail about the model and its assumptions would help researchers assess what the simulation establishes.
The immediate result is a computer simulation, not a working quantum processor. Nvidia’s stated figures describe the scale of the model and the qubit reduction it claims for its method. Any assessment of practical performance will depend on subsequent research and hardware work.