When the commercial, scalable, fault-tolerant quantum computing era really begins, when it becomes widely available, it will – least at the start – be a cloud service that is integrated with classical, powerful supercomputers, accelerator-like nodes that will run alongside with CPUs and GPUs and take on the workloads that are too powerful for their classical kin.

Increasingly, major players in high-end computing as well as smaller vendor and startups are beginning to put pieces in place that will allow such integrated systems to operate smoothly. As we noted last year, Nvidia, which supplies much of the fuel for the still-expanding AI market, has begun equipping its offerings with capabilities to link HPC with quantum. For example, NVQLink is a high-speed interconnect for linking classical supercomputers to quantum systems, while CUDA-Q is Nvidia’s quantum-classical platform.

More recently, startup Quantum Elements is using a combination of AI and digital twins to speed up the arrival of commercial, fault-tolerant quantum, and Quantum Machines this week launched The Open Acceleration Stack, a framework aimed at users who want to integrate any classical process into their quantum control stack, which co-founder and chief technology officer Yonatan Cohen said “reflects the industry’s shift from quantum computing demonstration to scaling and integration. It meets the needs of two critical areas of quantum development: real-time error correction and advanced qubit calibration, and provides the framework to scale both hardware and software with user experience and performance in mind.”

The Need For Quantum-Classical Systems

The idea of integrated quantum and HPC systems has been chewed on for several years. Startup superconducting quantum processor maker QuantWare wrote that “as the boundaries between classical and quantum computing move closer together, the industry is converging on one vision: the future of high-performance computing will be heterogeneous where quantum computing will emerge as another ‘tool in the toolbox’ of available compute architectures.”

It’s even become a national security issue, with the Center for Strategic and International Studies (CSIS) wrote this month that “integrating quantum computers into U.S. world-class supercomputers is now a strategic imperative for U.S. technological leadership in the next era of computing. While the United States leads in supercomputing and quantum computing, it lags behind Europe and Japan in developing hybrid quantum-supercomputing systems.”

The Big Blue Blueprint

IBM this month unveiled a reference architecture that executives say gives the industry a roadmap for bringing quantum and classical computing together to run workloads in what they call quantum-centric supercomputing (QCSC). IBM sees it as a “blueprint for the future of computing, but one that is meant to show compatibility and complementarity with what exists,” Jerry Chow, IBM Fellow and chief technology officer for QSCS at Big Blue, tells The Next Platform.

“Quantum and HPC need to come together, and there are all kinds of places that are putting these into datacenters,” Chow says. “What we wanted to do was really put a stick into the ground to show a blueprint for technically how this can look and how a heterogeneous compute of having quantum alongside GPUs and CPUs in a high-performance computing platform can really interoperate and communicate, can be orchestrated, and can be programmed for end applications.”

The reference architecture, detailed in a research paper, includes multiple layers, with the hardware infrastructure as the foundation that itself is divvied up into three tiers with their own computational capabilities, interconnects, and proximity to each other, according to IBM scientists. The base is the quantum system, which includes classical runtime and one or more interconnected QPUs, with the runtime comprising specialized classical accelerators – FPGAs and ASICs – and CPUs whose job is to enable QPU operations from error correction coding to qubit calibration to active qubit reset.

Making up the second tier are programmable CPU and GPU systems that are co-located with the quantum system and connected via a low-latency, near-time interconnect, like RDMA over Converged Internet (ROCE), Ultra Ethernet, and NVQLink, among others. The last are partner scale-out systems, either in the cloud or on premises.

Atop the infrastructure is the orchestration layer that includes the Quantum Resource Management Interface (QRMI), an open source library that abstracts away hardware-specific details and delivers APIs for quantum resource acquisition, task running, and systems monitoring. There also is application middleware that provides as communication tool between the independent quantum and classical programming models, and application software.

“Whereas CPUs represent information using binary code and GPUs use tensors, QPUs rely on circuits for their programming model,” the scientists wrote. “Evolving existing solvers into QCSC solvers requires an application layer where computational libraries can decompose a problem into components that launch in different environments. This layer facilitates an interplay between classical libraries and quantum libraries that prepare, optimize, and post-process quantum workloads into pre-defined circuits relative to the application domain, often using classical resources to do so.”

Chow says IBM already has been exploring the integration of quantum and classical, adding that quantum is reaching comparability with classical when it comes to physics and chemistry problems, something found through work done with Cleveland Clinic using a QCSC workflow. Big Blue has also worked with early deployments of the reference architecture with the RIKEN supercomputing environments and its Fugaku supercomputer.

“Overall, the architecture is that to show a number of different uses cases that have this either tight temporal or spatial co-location as a guiding direction,” he says. “That's why we are trying to inform this primarily with an evolution of the architecture. It’s not to say this is the one architecture to rule them all. It's meant to really show progressively more tightly coupled resources. In the long run, really drive a lot of co-design of the systems to scale with the application and the algorithms and libraries as they expand in the key application verticals.”

IBM has a timeline that outlines what the vendor sees as the evolution of the quantum-classical computing integration over the next few years.

Another timeline is playing a role here, Chows says. Key enablers of the reference architecture and the work IBM has done in this area are the capabilities in, first, its Heron 133- to 156-qubit superconducting quantum chip released in 2023 and, now, its Nighthawk 120-qubit chip, shown below, rolled out in November 2025.

Those Nighthawk chips brought IBM “to the point where it is beyond what you can simulate exactly for certain circuits,” he says. “So it becomes, then, a proving ground for exploration for many of our users. It becomes not research on the device, but it becomes really exploring research and exploration with the quantum processor. A big part of that is, how do you leverage it alongside what people are typically doing from a classical perspective?”

Quantum isn’t going to replace every part of the classical infrastructure lineup, he says. Like CPUs and GPUs work together not, QPUs will become an important part of the stack.

“From an algorithm perspective, you make sure that you use your accelerators in the area that they're best,” Chows says. “You're doing your static batch with CPUs, you're doing your matrix and tensors with GPUs – you’re always going to do it there – and you're going to do quantum circuits, which is really the language that we have available to us for quantum computing on something that uses entanglements or superposition support on a quantum computer. The art is going to come down to, from an algorithm perspective, how do I best wield these different pieces? What's exciting about where we are is that by having these hybrid models and having this reference architectures, that people can start to think about how best to leverage these for their capabilities.”