Google Cuts Magic-State Errors 40-Fold; Nvidia Models 104 Qubits

By Sophie Lindqvist

Google’s Willow superconducting chip cut magic-state errors 40-fold and reached 99.99 percent logical T-state fidelity in the first reported hardware implementation of magic-state cultivation.

Google and Nvidia announced projects tied to the advance on October 15. Nvidia detailed its simulation less than a week later. The Google result came from hardware. Nvidia’s came from simulation. That distinction limits what either company has proved.

Why do magic states constrain scaling?

Qubits are fragile. They quickly lose coherence, and excessive errors make a quantum system useless. Vendors therefore build logical qubits from many physical qubits. These logical units detect, isolate and repair errors without interrupting the underlying quantum state. One logical qubit can require thousands of physical qubits.

Universal quantum systems must support every operation performed by classical servers, plus quantum-specific operations. They need Clifford gates, including S and CNOT, and non-Clifford gates, including T. Most error-correction codes handle Clifford states, but not the non-Clifford operations.

Researchers introduced magic gates in 2004. Computers prepare these gates and inject them into calculations. They enable universal quantum computation, but they are fragile, difficult to create at high fidelity and unusable after one operation. Systems must continuously make more.

Magic-state distillation gathers many noisy states and converts them into fewer high-fidelity states. The method can consume as much as 95 percent of a quantum system’s capacity. The time and capacity burden remains a central scaling constraint.

Google researchers proposed magic-state cultivation in 2024 as a cheaper route. They place a noisy state in a small chip area, then apply special coding and measurements. The measurements act like a filter, identify errors and discard flawed attempts, leaving a lower-error magic state. The method seeks to reduce the hardware overhead of magic-state generation.

What did Google demonstrate?

After two years of moving cultivation from theory toward reality, Google Quantum AI scientists published their work in Science. The paper reports the Willow experiment on a superconducting chip as the first hardware implementation of cultivation.

The demonstration produced T-states with 99.99 percent logical fidelity. It also reduced errors 40-fold relative to uncultivated states. Those are experimental results, not a roadmap for a fault-tolerant system.

The paper’s authors still called for a larger-scale demonstration before cultivation becomes realistic in a practical quantum system. They wrote that “this achievement takes a key theoretical idea and turns it into an experimentally demonstrated reality.”

What did Nvidia simulate?

Nvidia ran a 104-qubit simulation on Jupiter, Europe’s first exascale computer. The system sits at the Jülich Supercomputing Centre in Germany and uses Nvidia GH200 superchips. Nvidia said cultivation uses 90 percent fewer qubits than other methods.

The claim concerns simulated resource requirements, not a completed 104-qubit quantum processor. The company used CUDA-Q and its cuQuantum software development kit to model hardware errors, including coherent errors and leakage.

Stanwyck described coherent errors as small systematic mistakes in quantum operations. They accumulate during a computation and limit what the machine can do. The simulations tested how cultivation responds, which can guide hardware choices, code and the next processor revision.

The work began on a GB200 NVL72-powered system at Nvidia’s Accelerated Quantum Research Center in Boston. Researchers used it to prepare and validate software for a Jupiter-sized system. They scaled tests from four GPUs to hundreds, then moved them to 8,192 GPUs on Jupiter.

Sam Stanwyck, Nvidia’s director of quantum product, said: “This is complementary to Google’s work. What Google did is a small-scale demonstration of this in hardware. It was a proof-of-concept that magic state cultivation, the approach, actually works on real hardware.”

How does simulation support the next processor?

Stanwyck said both projects were critical steps toward fault-tolerant quantum computing. Simulation, he argued, must probe cultivation deeply enough to show how it behaves and handles errors before engineers build larger processors.

He said the physics-to-application model links physical-qubit errors with application runtime. That link can help researchers design better cultivation systems and quantum computers. His claim that simulation can pull in the arrival of useful quantum computing remains unquantified.

Nvidia wants its products to accelerate future hybrid quantum-classical environments. It is not alone in using digital twins for error-correction research. Earlier this year, scientists from Amazon Web Services and Quantum Elements, together with the University of Southern California and Harvard University, wrote about work in this area.

What remains unproven?

Google has shown cultivation on real hardware at small scale. Nvidia has modeled the method at larger computational scale. Neither result establishes that cultivation can support a practical fault-tolerant machine.

The next commercial test is whether both teams can turn the gains into larger systems with lower qubit and compute overhead; if they do, Nvidia’s platform could compete for hybrid-classical workloads, but neither announcement gives a schedule, cost or capacity target.