Nvidia quantum simulation hits 104 qubits on 8,192 GPUs

The simulation modeled an entire fault-tolerant quantum processor, meaning one that can spot and fix its own errors while it runs. All of it ran on conventional computer chips. No quantum hardware was involved.

Qubits, the basic units of a quantum computer, are extremely sensitive. Even tiny amounts of noise from the equipment used to control them can flip results. Most experts see error correction as the essential step before quantum computers become practically useful, and building it means understanding exactly how errors behave.

That’s what makes this run different. Earlier work at Jülich, announced in November 2025, simulated 50 qubits under ideal, noise-free conditions. Nvidia’s model includes the messy problems real chips have, such as small phase errors that build up over time and leakage, where a qubit drifts out of its two normal states into a third energy level. Modeling that third level made the job far harder.

The technique being tested is called magic state cultivation. Error-corrected quantum computers can do many operations cheaply, but to run any possible program they also need special ingredients called magic states. These get used up constantly during a calculation, so a machine has to keep producing fresh ones.

Making magic states usually eats a lot of extra qubits. Nvidia says the cultivation approach it validated needs 90% fewer physical qubits than earlier methods. Since every qubit adds cost and engineering pain, a cut that size could make a useful machine much smaller and cheaper to build. In this test, the protocol starts with a small error-correcting code, checks the state for faults, then grows it into a larger and sturdier code before reading it out. The 104 qubits include both the qubits holding data and the helpers used along the way.

Simulating qubits on normal computers gets brutally hard, fast. Each extra qubit doubles the computing and memory needed. When Jülich simulated 50 qubits, it took about 2 petabytes of memory. A brute-force simulation of 104 qubits would be far beyond any computer on Earth. For perspective, the 50-qubit run itself beat a 48-qubit record that Jülich researchers set in 2019 on Japan’s K computer, so progress here usually comes a couple of qubits at a time.

So Nvidia used clever shortcuts. Its simulator reorders operations without changing results, recycles memory as qubits are measured and released, and packs the three-level qubits tightly into ordinary binary storage. Even then, the numbers are staggering:

• 8,192 GPUs using Nvidia GH200 chips on JUPITER.
• About 35.2 trillion complex numbers tracked at once to describe the quantum state.
• 512 TiB of GPU memory for that state alone, or 64 GiB on each GPU.

JUPITER itself opened in September 2025 as Europe’s first exascale supercomputer, a machine that can perform more than a billion billion calculations per second. Before moving there, Nvidia developed the method on a smaller system at its own quantum research center, scaling from four GPUs to hundreds. The company says that cut the work from weeks to days.

Don’t expect a quantum computer in your home, or to crack your passwords, any time soon. This was a simulation of a single key process, not a working machine, and even a very detailed model can miss ways real chips fail. Nvidia also hasn’t said which hardware makers plan to build around this exact protocol. Its value is in giving hardware builders data to predict how designs will perform before they commit to building them.

It also shows Nvidia’s strategy clearly. Rather than racing to build qubits itself, it wants every quantum company to depend on its GPUs and software, its cuQuantum libraries and CUDA-Q platform, to design and test their machines. Huang framed it as quantum processors and GPU supercomputers working together. In his words, useful quantum computing will require fantastically complex, fault-tolerant systems.

For Europe, it’s a showcase. Jülich director Kristel Michielsen said the investment in JUPITER lets researchers tackle questions that were previously beyond reach. The timing fits a bigger push: the German government struck an AI partnership deal with Nvidia at a Berlin investment summit just days ago.