Google Cut Magic-State Error 40-Fold on Willow; Nvidia Needed 8,192 GPUs to Simulate It

Google Quantum AI reported on its Willow superconducting processor that a technique called magic state cultivation cut the error in a key quantum resource by a factor of 40, to a fidelity of 0.9999(1). Nvidia then simulated the same protocol on 104 simulated qubits, using all 8,192 GPUs of Europe's JUPITER supercomputer, and said the method can cut physical-qubit overhead by up to 90 percent. The two results measure different things, and the gap between them is the useful part.

The number that matters is the 8 percent

A magic state is a specially prepared quantum state that a fault-tolerant machine consumes to perform the T gate, the operation that makes a quantum computer more than a classical one. The standard way to make them clean is distillation, which burns many noisy copies to produce one good one. According to GCN's report on Nvidia's work, distillation can consume 90 percent to 95 percent of a machine's physical-qubit budget.

Cultivation grows the state in place inside a small error-detecting code instead, and discards the attempts that show an error. The arXiv preprint, submitted Dec. 15, 2025 and carrying about 295 authors including Craig Gidney and Hartmut Neven, reports a 40-fold error reduction, a fidelity of 0.9999(1) and that the protocol retains 8 percent of attempts.

The retention figure is the denominator. Keeping 8 percent means roughly 12.5 attempts for each state that survives, which is a cost in time and qubit-cycles that the headline fidelity hides. The abstract does not give the qubit count, the device name or the other error rates, and the paper body could not be read as text for this report, so those details are unverified here.

A 40-fold reduction to 0.0001 implies the starting error was about 0.004. That is arithmetic on the abstract, not a figure from it.

Where Google's and Nvidia's dates disagree

Google's result appeared in the journal Science, but sources differ on the day. The Next Platform gives Oct. 15; GCN gives Oct. 16. Nvidia announced its simulation on Oct. 19, ahead of its GTC Berlin conference, according to GCN, which is under a week after Google.

The Next Platform quotes Sam Stanwyck, Nvidia's director of quantum product, describing Google's result as a proof of concept on real hardware and Nvidia's work as complementary, since simulation lets researchers study error behaviour before building hardware. Google's authors, the same report says, state that larger demonstrations are still needed.

What 8,192 GPUs bought

Nvidia's run modelled a 104-qubit fault-tolerant processor with realistic hardware noise, including coherent errors and leakage, using its CUDA-Q and cuQuantum software. GCN reports the team started on a GB200 NVL72 rack at its Boston quantum research centre, scaled from 4 GPUs to hundreds, and then ran at full size on JUPITER, at the Jülich Supercomputing Centre in Germany. The system uses GH200 chips.

The cost denominator is worth stating. Simulating a noisy 104-qubit device took the entire GPU complement of an exascale machine, which is a reminder of why Google's chip matters: the physical device does this work natively. Neither source gives a simulation runtime, a node count or a speed-up, so the compute bill cannot be priced. GCN adds that JUPITER simulated 50 qubits under ideal, noiseless conditions, in 2025, which beat a 48-qubit record from 2019. That run assumed no errors at all. The 2026 run added realistic noise and used the machine's full 8,192-GPU configuration.

For a reader pricing quantum hardware against other silicon, the comparison point is conventional chips, where Tenstorrent raised $1.5 billion in a Series E on a $6.2 billion valuation. Quantum has no equivalent shipping product with a published price per useful operation. Google, whose quantum group produced the hardware result, is also pricing its model business in public, as the Gemini 4 Argon launch showed; the quantum side has no such price list.

Nvidia's 90 percent saving is the company's own claim. The Next Platform notes that the article does not name the methods it is measured against, and GCN says the comparison is to traditional cascade distillation factories.

Item	Google Quantum AI	Nvidia
Platform	Willow superconducting chip	JUPITER, 8,192 GH200 GPUs
Result	40x error reduction, 0.9999(1) fidelity	104-qubit noisy simulation
Retention / saving	8% of attempts kept	Up to 90% fewer physical qubits (Nvidia claim)
Kind of evidence	Hardware experiment	Classical simulation

What would change this read

The case that cultivation lowers the qubit bill rests on a small demonstration and a vendor simulation. Three things would move it. A larger code distance on hardware with the retention rate reported alongside it would show whether 8 percent holds as the code grows. A published Nvidia baseline would turn "up to 90 percent" into a comparable figure. And a Google statement of qubit count and cycle time for each accepted state would let a reader price cultivation against distillation in qubit-seconds, which is the unit a machine builder pays in. Until then, a single logical qubit can still require thousands of physical ones, according to The Next Platform, and neither result changes that count by itself.