The Moment Quantum Computers and GPUs Became One System

For years, quantum computing and classical supercomputing existed in separate worlds — connected by slow API calls and awkward middleware. At GTC 2026, NVIDIA changed that with NVQLink, the first universal interconnect that physically links quantum processors to GPU-accelerated supercomputers. With 400Gb/s bandwidth and sub-4-microsecond latency, NVQLink doesn't just improve the quantum-classical interface — it fundamentally redefines it.

"In the future, supercomputers will be quantum-GPU systems," declared Jensen Huang, NVIDIA's CEO. As of April 2026, that future is no longer theoretical. It's being deployed at national labs, integrated by 17 QPU builders, and already producing real scientific results.

Why NVQLink Matters Now

The quantum computing industry spent years in a qubit-count arms race. But the real bottleneck wasn't qubit quantity — it was the classical infrastructure required to make those qubits useful. Quantum error correction (QEC), the essential process of detecting and fixing errors in quantum computations, demands real-time feedback loops operating at microsecond timescales. Previous architectures, with their millisecond-scale REST API connections between quantum controllers and GPU servers, simply couldn't deliver.

NVQLink was designed to solve this exact problem. It builds on NVIDIA's existing CUDA-Q platform — an open-source quantum development framework supporting Python and C++ that already integrates with 75% of publicly available QPUs — by adding a hardware-level, low-latency interconnect between GPUs and quantum processors.

The timing is right because quantum hardware has finally reached the scale where error correction isn't just desirable — it's mandatory for useful computation. And error correction at scale requires the kind of real-time GPU acceleration that only a tight physical coupling can provide.

Inside the NVQLink Architecture

NVQLink introduces the Logical QPU machine model, consisting of three tightly coupled components:

Real-time Host: An NVIDIA GPU-accelerated node (Grace Hopper or Grace Blackwell), programmable via CUDA-Q in C++ or Python. This handles computationally heavy tasks like QEC decoding.

Quantum System Controller (QSC): Third-party FPGA/RFSoC hardware that directly controls qubits through pulse processing units. Supported controllers include those from Keysight Technologies, Qblox, QubiC, and Zurich Instruments.

Real-time Interconnect: A low-latency RDMA over Converged Ethernet (RoCE) network using standard NVIDIA ConnectX hardware with Precision Time Protocol (PTP) timestamping.

The measured performance is remarkable. Across 1,000 test samples, the system achieved a mean end-to-end latency of 3.84 microseconds with a standard deviation of just 0.035µs. On an RTX 6000 Blackwell Pro with ConnectX 7, the three-kernel dispatch mode hit 2.92µs — fast enough for real-time error correction on any current quantum processor.

The architecture supports 400Gbit/s Ethernet links with 256-port switch radix, meaning it scales with the same infrastructure powering today's AI data centers. A key design principle is IP preservation: the FPGA core is open-source, so quantum hardware builders can adopt the interface without disclosing proprietary firmware.

Real-World Proof: Quantinuum Helios and GPU-Accelerated Error Correction

The most compelling NVQLink demonstration came from Quantinuum's Helios processor — widely regarded as the world's most accurate commercial quantum computer. Using NVQLink, the Quantinuum team integrated an NVIDIA GPU-based decoder directly into the Helios control engine.

The results speak for themselves: decoding Bring's code (8 logical qubits encoded in 30 physical qubits) using a BP+OSD algorithm, they achieved a median decoding time of 67 microseconds — exceeding Helios's 2-millisecond requirement by 32x. This real-time error correction improved logical fidelity by over 3% and reduced the error rate by 5.4x (from 4.95% to 0.925%).

This wasn't a lab curiosity. It was a demonstration that GPU-accelerated quantum error correction works at production scale, on commercial hardware, with commercially relevant error rates. Quantinuum is now integrating NVIDIA GB200 with Helios via NVQLink to develop Generative Quantum AI (GenQAI) applications targeting power grid optimization, nuclear fuel arrangement, and molecular design for drug discovery.

Programming the Quantum-GPU Stack: CUDA-Q and cudaq-realtime

NVQLink's software interface is the new cudaq-realtime API, released as part of the CUDA-Q platform at GTC 2026. This API enables developers to write code that exchanges data between GPUs and QPUs at microsecond timescales.

The core abstraction is cudaq::device_call, which lets quantum kernels invoke GPU functions and receive results within microseconds:

auto syndrome = mz(ancilla_qubits);
cudaq::device_call(/*gpu_id=*/1, surface_code_enqueue, syndrome);
auto correction = cudaq::device_call(/*gpu_id=*/1, surface_code_decode);

Python developers get an equally intuitive interface for QEC workflows:

@cudaq.kernel
def qec_circuit() -> int:
    qec.reset_decoder(0)
    syndromes = measure_stabilizers(logical)
    qec.enqueue_syndromes(0, syndromes, 0)  # Asynchronous
    corrections = qec.get_corrections(0, 1, False)

The asynchronous enqueue pattern is crucial — it allows the QPU to continue operating while the GPU decodes, maximizing overall system utilization.

The cudaq-realtime library offers four kernel execution modes: three-kernel dispatch (default, transport-agnostic), unified kernel (lowest latency), transport-only forwarding (for benchmarking), and cooperative kernel (for distributed workloads like multi-block belief-propagation decoders). Development without FPGA hardware is supported through emulation mode.

The Expanding Ecosystem

As of April 2026, the NVQLink ecosystem includes 17 QPU builders (Alice & Bob, Atom Computing, Diraq, Infleqtion, IonQ, IQM, ORCA Computing, Pasqal, Quantinuum, QuEra, Rigetti, and more), 5 controller builders, and 9 U.S. national laboratories. Supercomputing centers across Asia and Europe — including Japan's AIST G-QuAT and Singapore's National Quantum Computing Hub — have joined the platform.

The scientific demonstrations at GTC 2026 showcased impressive scale:

• Biomolecular simulation: A UCL consortium combined IQM's 54-qubit system with 120 NVIDIA H100 GPUs for hybrid simulation of a G-protein-coupled receptor (GPCR).
• Record-breaking simulation: CINECA and Kipu Quantum executed the largest-known statevector simulation — a 43-qubit quantum optimization routine using 2,048 Ampere GPUs.
• Cancer research: Infleqtion's Q4Bio project consumed 24,000 A100 GPU-node-hours on NERSC's Perlmutter supercomputer to train quantum neural networks for cancer biomarker discovery.
• QEC acceleration: University of Edinburgh researchers built a "vibe decoder" for color codes on GH200 GPUs, achieving 900x speedup over previous state-of-the-art.
• Autonomous algorithm discovery: Companies like Hiverge and Quantinuum are using LLM agents to translate natural-language problem descriptions into executable quantum circuits.

Getting Started Today

You don't need quantum hardware to begin working with this stack. Here's a practical path:

Step 1: Install CUDA-Q. A simple pip install cudaq gets you started. GPU acceleration is optional — the built-in simulator works on CPU.

Step 2: Learn the fundamentals. The CUDA-Q Academic repository on GitHub provides free Jupyter notebook modules covering hybrid quantum-classical algorithms from basics to optimization.

Step 3: Access real QPUs via cloud. TII's Quantum Computing Cloud Platform and Scaleway's Quantum-as-a-Service both offer CUDA-Q-compatible access to physical quantum hardware and simulators.

Step 4: Experiment with cudaq-realtime. The library ships with built-in latency benchmarking tools and supports emulation mode (./hololink_test.sh --emulate) for development without FPGA hardware.

PNNL (Pacific Northwest National Laboratory) is also developing an open-source GPU acceleration framework using NVQLink, specifically designed to lower barriers for scientists and engineers exploring quantum control and measurement.

What to Watch Next

The convergence of quantum computing and GPU-accelerated AI is accelerating faster than most predicted. Classiq has already demonstrated a 26x speedup in quantum circuit synthesis and execution using CUDA-Q on a single A100 GPU (from 67 minutes to 2.5 minutes for a 31-qubit circuit). Enterprise quantum-AI pilots are live in finance, pharma, and aerospace, with mainstream adoption expected to accelerate through 2030.

NVQLink represents quantum computing's "TCP/IP moment" — an open, vendor-neutral interconnect that unifies a fragmented hardware landscape into a coherent system. With 17 QPU builders, 9 national labs, and growing cloud availability, the infrastructure for practical quantum-GPU computing isn't coming. It's here. The question is no longer whether hybrid quantum-classical systems will become the standard — it's how quickly your organization will start building on them.