NVIDIA Runs 104-Qubit Quantum Simulation on JUPITER

NVIDIA has used JUPITER, Europe’s first exascale supercomputer, to run what it calls a record-breaking simulation of a fault-tolerant quantum processor, the company announced on 19 October 2026. The 104-qubit run, at the Jülich Supercomputing Centre in Germany, tested a technique called magic state cultivation under realistic noise. For anyone betting on when quantum computers become useful, the point is cost: NVIDIA says the method needs 90% fewer qubits than earlier approaches.

This is not a quantum chip result. It is a classical simulation, run on thousands of NVIDIA graphics processors, of how a future error-corrected quantum machine would behave. Its value, if NVIDIA’s account holds up, is that hardware builders can test a design in software before spending years building it.

In brief

NVIDIA says it ran the largest comprehensive simulation of a fault-tolerant quantum processor to date: 104 qubits, on up to 8,192 GPUs on JUPITER.
It modelled magic state cultivation, a way of making scarce “magic states” that NVIDIA says needs 90% fewer qubits, under realistic hardware noise.
No paper or independent check has been published alongside the release, and “record” is NVIDIA’s own description.

Why magic states are the bottleneck

Quantum computers are fragile. As NVIDIA’s release puts it, “the interactions needed to control them introduce unavoidable, error-inducing noise”, so useful machines will need error correction. That in turn depends on magic states, which NVIDIA describes as “a specific arrangement of qubits that is continually consumed during a fault-tolerant quantum computation”.

Because those states are used up as a machine runs, they must be made again and again. “Any optimization in how they are made drastically reduces the difficulty of building a useful quantum computer,” the company says. That is why magic state cultivation has drawn attention: it promises to make the states far more cheaply than older methods.

Last week the idea moved from paper to hardware. As TSN reported, Google Quantum AI said on 15 October that cultivation cut errors about 40-fold on its Willow chip, while stressing that it still had to be shown at larger scale. NVIDIA’s simulation is aimed squarely at that gap.

What NVIDIA says it simulated

The release says the run used NVIDIA GH200 chips at Jülich and modelled “how a realistic quantum processor performs fault-tolerant computation at record scale”. NVIDIA contrasts it with earlier work: “Earlier, groundbreaking research from Jülich simulated 50 qubits under ideal, noiseless conditions. This work simulates magic-state cultivation under realistic hardware noise.”

Simulating qubits is hard because, the company notes, “the memory required can grow exponentially with the number of qubits”. That is why the jump matters more than the raw numbers suggest.

Detail	NVIDIA’s figure
Qubits simulated	104
GPUs on JUPITER	8,192
Earlier Jülich run	50 qubits, noiseless
Qubit saving claimed	90% fewer
Development time	Weeks cut to days

Source: NVIDIA release, 19 October 2026. All figures are NVIDIA’s.

The team first developed its methods on a GB200 NVL72 system at the NVIDIA Accelerated Quantum Research Center, scaling “from four GPUs to hundreds before moving to 8,192 GPUs on JUPITER”. NVIDIA says this “helped reduce a weekslong simulation effort to days”. The software was its own cuQuantum library and the CUDA-Q platform’s fault-tolerant logical layer.

The claims, and the limits

NVIDIA chief executive Jensen Huang framed the work as groundwork for real machines. “Useful quantum computing will require fantastically complex, fault-tolerant systems — and understanding how they behave at scale is essential to building them,” he said.

Kristel Michielsen, director of the Jülich Supercomputing Centre, pointed to Europe’s investment. “Europe’s investment in JUPITER gives researchers the ability to investigate questions that were previously beyond reach,” she said, adding that the work helps scientists choose “the designs most likely to deliver useful applications”.

The bigger claims need care. NVIDIA’s summary says the simulations are “shortening the path to viable quantum computers by years”. That is a forecast, not a measured result. The release links no peer-reviewed paper, and it does not say what error rates the simulation found or how they compare with Google’s hardware results.

The timing is also crowded. On the same day, Quantinuum claimed a record logical fidelity on its Helios machine, and NVIDIA was in Berlin as Germany signed an industrial-AI declaration with the company.

Our read

NVIDIA is positioning its GPUs and software as the test bench every quantum hardware maker will need, whichever qubit technology wins. A noisy 104-qubit simulation is a genuine step beyond noiseless 50-qubit work, but until the methods and results are published, “record” remains NVIDIA’s word. The useful test is whether hardware teams, including Google’s, report that these simulations changed what they built.

The next marker is the technical write-up NVIDIA points to. Watch for a paper or code release, and for any quantum hardware maker naming JUPITER-scale simulation in its own roadmap.