A different future for quantum computing
Nvidia's NVQLink is an open platform architecture for connecting GPU-accelerated computing directly to a quantum processor’s control system (image: Nvidia)
In research published in October 2025, a quantum processor in Japan received help from 152,064 conventional computing nodes. The machine was not defeating one of the world’s largest supercomputers. It was working with it.
Now a newer experiment in Germany suggests that this unusual collaboration could become part of normal supercomputing.
In a preprint published on 14 August 2026, researchers demonstrated hybrid workflows connecting the petascale SuperMUC-NG supercomputer at the Leibniz Supercomputing Centre with a 20-qubit superconducting quantum processor supplied by IQM Quantum Computers.
The breakthrough was not a faster quantum calculation. It was that the supercomputer could treat the quantum processor as a resource to be allocated when needed.
The experiment points towards a different future for quantum computing. Instead of replacing conventional supercomputers, quantum processors may become specialised components within them.
Much as a graphics processing unit, or GPU, handles selected calculations for artificial intelligence and computer graphics, a quantum processing unit, or QPU, could be called upon for the particular parts of a scientific problem that might benefit from quantum mechanics.
Researchers are building the architecture before they know whether quantum hardware will become good enough to justify it. Without that architecture, however, even a powerful QPU would remain difficult to use.
The quantum computer was never going to work alone
The popular image of quantum computing is a succession story. Classical computers reach their limits; quantum computers take over.
The engineering reality is less dramatic and more interesting.
A QPU is designed for particular kinds of operation. Even if future quantum machines become far larger and more reliable, they will still depend on conventional processors to prepare data, control the hardware, correct errors and interpret the results.
Imagine researchers trying to understand how electrons behave inside a molecule; a calculation central to chemistry and materials science that becomes dramatically more difficult as the system grows.
A conventional supercomputer can prepare the molecular data and perform most of the calculation. The QPU might be asked to handle one especially difficult operation. Its output would return to the classical system for further analysis.
The quantum processor would not design a medicine or battery by itself. It would act as a co-processor for one carefully selected part of a much larger job.
This approach is known as hybrid quantum-classical computing. CPUs and GPUs perform most of the work, while the QPU is brought in when a suitable quantum method might offer something the classical processors cannot.
SuperMUC-NG turns the QPU into a bookable resource
Anyone using a large supercomputer does not normally choose processors one by one. They submit a job to a workload manager, which finds the necessary resources and decides when the calculation can run.
At SuperMUC-NG, that workload manager is Slurm, software widely used by supercomputing centres.
Using Munich Quantum Valley’s Munich Quantum Software Stack, researchers exposed the IQM quantum processor to Slurm as what the system calls a “generic resource”. Another part of the software compiled the quantum circuits and sent them to the QPU.
The effect was deceptively simple: the quantum processor became something a computing job could request.
Users did not have to leave the existing supercomputing environment and submit the quantum part through a separate system. Nor did the Leibniz Supercomputing Centre have to rewrite the core of its scheduler.
The researchers reported that the hybrid workflows ran without significant additional latency from the integration. That does not mean the QPU accelerated the scientific calculation. It means adding the quantum processor did not create a large new orchestration delay.
Putting two machines together is not enough
Quantum computers are already available through cloud platforms. A user prepares a quantum circuit, sends it over a network and waits for the answer.
That can work when the QPU is contacted only occasionally. It becomes restrictive when a calculation requires data to move repeatedly between quantum and classical hardware.
Placing a QPU in the same computing centre reduces the physical distance, but co-location does not automatically make the machines work as one.
Hybrid systems also need fast connections, compatible control electronics and software capable of coordinating several types of processor while a calculation is running. Some operations, particularly those involved in controlling and correcting quantum hardware, may require responses within microseconds.
Nvidia is addressing this problem through NVQLink, an open platform architecture for connecting GPU-accelerated computing directly to a quantum processor’s control system.
A 2025 technical preprint reported a maximum round-trip latency of 3.96 microseconds for Nvidia’s reference network. The figure measures communication through the control infrastructure, not the running time of a complete quantum application.
Connections at that speed could allow GPUs to assist with QPU calibration and quantum error-correction decoding while the quantum processor is operating. Nvidia lists latency below four microseconds and throughput of up to 400 gigabits per second for the NVQLink reference platform.
IBM has mapped out a broader version of the same future. Its reference architecture for 'quantum-centric supercomputing', published in March 2026, describes a progression from QPUs used as offload engines to systems in which quantum hardware, classical processors and software are designed together.
IBM and Japanese research institute RIKEN have already explored the middle ground. Their closed-loop workflow connected an on-site IBM Heron quantum processor with all 152,064 computing nodes of Japan’s Fugaku supercomputer. The QPU generated samples while Fugaku performed the extensive classical work around them.
The results were comparable with some classical approximation methods, but they did not show the QPU beating Fugaku. The experiment demonstrated that quantum and classical processors could repeatedly exchange data within one scientific calculation.
Who will build the CUDA of quantum computing?
GPUs did not transform computing simply because their chips could perform certain calculations quickly. They became transformative when developers gained practical tools for using them.
Nvidia’s CUDA software platform gave programmers a way to send selected work to GPUs without controlling every part of the hardware directly. Quantum computing is now approaching its own version of that software contest.
The surrounding system must know when to use the QPU, how to translate a problem into quantum instructions and what to do with the answer.
Compilers must prepare circuits for different quantum machines. Workload managers must allocate scarce QPU time. Classical processors must perform the preparation, optimisation and error mitigation around each quantum operation.
Several organisations are trying to establish this connecting layer. Nvidia is building around NVQLink and CUDA-Q. IBM is using its Qiskit ecosystem. Munich Quantum Valley has developed a stack compatible with established supercomputing tools. Hardware manufacturers such as IQM must ensure their processors fit into these wider environments.
The most influential company in quantum computing may therefore not be the one announcing the largest number of qubits. It could be the one whose software becomes the standard route through which researchers use QPUs from several manufacturers.
Quantum computing has spent years trying to prove that it can outperform the supercomputer. Its more immediate future may depend on learning how to join one.
Further reading on MoveTheNeedle.news:
Can a second dimension solve trapped-ion quantum computing’s scaling problem?
PsiQuantum’s $1bn bet on fault tolerance: why the race now feels real
Quobly wants to bring semiconductor-scale manufacturing to quantum computing