Three announcements landed at SC26 in Chicago on November 16, and none of them involved a physical quantum computer being built or shipped. What IBM, Nvidia, and Fermilab unveiled was something more strategically interesting: a plan to co-locate quantum and classical hardware at the Department of Energy's most advanced quantum research center, a new decoder that cuts quantum error-correction time by a factor of up to 13,000, and a Slurm update that lets any HPC site schedule quantum processors the way they already schedule GPUs. Taken together, they define where fault-tolerant quantum computing is actually headed, and who controls the classical infrastructure surrounding every qubit in every lab.
Three Announcements, One Week, Zero New Hardware
The core announcement, per the Fermilab primary announcement published November 16, is a planned modular quantum computing co-design platform. It would physically co-locate IBM quantum processors with Nvidia DGX B200 accelerated computing systems at Fermilab's Superconducting Quantum Materials and Systems Center, the DOE national lab that leads U.S. government research on superconducting cavity hardware. The initial phase focuses on emulating communication between quantum modules: how link errors, latency, and classical feedback loops affect performance when multiple quantum processors are networked together. A later phase would add Fermilab's own ultracoherent superconducting cavities as quantum memories and entanglement buffers between modules, the technology IBM needs to hit its 1,040-qubit modular system target by 2027.
Fermilab CTO Anna Grassellino called it a "proposed platform" in her own press quotes. IBM Fellow and Quantum CTO Jerry Chow said the work would "open new frontiers in high-energy and nuclear physics." Per the IBM and Fermilab joint press release, no site, system count, start date, or budget was disclosed. What was disclosed, separately but on the same day, were two software products already in production: Nvidia's CUDA-Q QEC 0.9.0 decoder and Slurm 26.11's native quantum processor scheduling. Those are the parts of the November 16 story that have actually shipped and are running in production today.
Fermilab is not a neutral party here. The Superconducting Quantum Materials and Systems Center is one of five DOE National Quantum Information Science Research Centers funded under the National Quantum Initiative Act, with a mandate to advance superconducting hardware specifically. Its director, Grassellino, has led the development of niobium cavities with coherence times far above the industry average, the technology that becomes quantum memory once a modular architecture is working. By partnering with both IBM and Nvidia, Fermilab positions itself as the interconnect layer between quantum processor vendors and the HPC centers that will ultimately run quantum workloads at scale.
What CUDA-Q QEC 0.9.0 Actually Delivers
The decoder at the center of the SC26 release is TCBP-JOBS: Trajectory-Calibrated Belief Propagation paired with GPU-parallel Jittered Ordered Statistics Decoding Beam Search. The job description is precise: when a quantum circuit runs, errors accumulate continuously, and a classical computer must identify the error pattern and feed corrections back before the quantum state collapses. Speed here is not a performance metric; it is a physical constraint. Too slow, and the quantum state you are trying to preserve has already decayed into noise. According to the Quantum Computing Report's SC26 coverage, TCBP-JOBS delivers latencies of 4.08 milliseconds for a 12-round gross-code memory and 13.5 milliseconds for logical lattice surgery at 0.2% physical error rate, using real IBM Nighthawk hardware. Those are feed-forward-compatible latencies, meaning they open the door to real-time error correction in mid-circuit operations rather than only post-processing.
The aggregate speed claim runs 10x to 13,000x faster than Tesseract, an exact search decoder that sets the accuracy bar at ultra-low noise. The wide range reflects the physics: Tesseract's precision advantage holds only at error rates below 0.05%, which no current quantum processor reliably achieves. At the 0.1% to 0.4% error rates where today's hardware actually operates, TCBP-JOBS wins by a large margin. Nvidia's demonstration decoded 24.2 million shots in 45 minutes; a comparable workload would have taken approximately 12 weeks on a leading single-threaded decoder, as reported by The Next Web. Relay-BP kernels, an earlier component of CUDA-Q QEC, improved 1.3x to 3.2x in the same release. The package is available open source on Nvidia's CUDA-Q GitHub repository.
Beyond speed, CUDA-Q QEC 0.9.0 ships a containerized Decoding Server that separates the decoder from the quantum hardware control stack, letting multiple quantum processors share decoding capacity across a network. That design matters for the co-design platform: when Fermilab eventually co-locates multiple IBM QPUs with Nvidia GPUs, a single decoder pool can serve all of them. Nvidia also released CUDA-Q QEC in a form that runs on AMD GPUs through its open integration, partly because IBM and AMD released the competing open-source Pelagos decoder at the same SC26 event, built on the Triton framework. The two decoders are compatible with the same QRMI interface, so a supercomputing center is not locked into a single decoder stack once it adopts Slurm 26.11.
Slurm 26.11 Turns QPUs Into Schedulable Compute Resources
The second shipped product is Slurm 26.11, which adds native quantum processor scheduling through the open-source Quantum Resource Management Interface (QRMI). Before this update, every supercomputing center running IBM, Quantinuum, or Pasqal quantum hardware had to build its own scheduling plugin. The situation was analogous to CPUs before POSIX: fast processors, but no common interface for jobs to request them. QRMI changes that. Researchers can now specify a QPU in a Slurm job script the way they would specify a GPU, a core count, or memory. Nvidia, which acquired SchedMD (the company behind Slurm), contributed the QRMI and QDMI interface work alongside IBM and Pasqal. The Fermilab HEPCloud infrastructure will manage researcher access and hybrid quantum-classical workflows, initially targeting lattice gauge theories, neutrino interactions, and nuclear simulations in high-energy and nuclear physics.
The scale of the existing deployment puts Slurm 26.11 in context. Per the SC26 disclosures, roughly 30 supercomputing centers are already deploying more than 50 quantum processors worldwide under the QRMI framework. Vendors already integrated include IBM, Quantinuum, Pasqal, IQM, Alice and Bob, and MQSC. Adding native Slurm scheduling converts those centers from specialized quantum labs into shared infrastructure any HPC user can request time on. The gatekeeping function shifts from hardware access to allocation policy, the same transition that happened when GPU clusters moved from dedicated AI labs into university HPC centers around 2019. For the quantum industry, that shift carries both opportunity and risk: more users means faster feedback on what hardware improvements matter most, but it also means quantum hardware reliability becomes visible to a large population of scientists who will not tolerate the error rates that quantum specialists currently accept.
IBM's Nighthawk Roadmap and the 1,040-Qubit Target
The quantum processor at the center of the announced platform is IBM's Nighthawk QPU: 120 qubits, capable of circuits with up to 7,500 two-qubit operations, running at 100,000 samples per second. Nighthawk r2, the generation shipping now, delivers 25x the throughput of IBM's Heron processor, the previous generation. The r3 generation on IBM's roadmap targets 10,000 two-qubit operations on 520 qubits, a step change in circuit depth that would put some fault-tolerant algorithms within practical reach for the first time. These are IBM's own figures, and they have not been independently audited.
Beyond r3, IBM has disclosed plans to link two quantum chips into a 1,040-qubit modular system by 2027, according to The Next Web's coverage of the SC26 announcements. The interconnect technology required is partly what Fermilab's SQMS Center is developing: ultracoherent superconducting cavities that serve as quantum memories and entanglement buffers between modules. The co-design platform announced at SC26 is specifically designed to test those interconnects in simulation before the physical hardware is built, letting IBM identify which link-error patterns matter most for the physics workloads Fermilab targets. According to StorageReview's SC26 roundup, the initial phase will emulate communication between quantum modules, studying how routing errors, latency, and classical feedback affect multi-QPU execution. The goal is to inform IBM's physical interconnect design before it commits to a manufacturing run.
The Fe4S4 Experiment: A 79x Accuracy Gain That Changes the Argument
Buried in the SC26 releases is a result that explains why IBM, Nvidia, and Fermilab felt ready to go public with a platform plan: a 79-fold accuracy improvement in simulating Fe4S4, an iron-sulfur cluster central to biological nitrogen fixation and electron transport in living cells. Classical computers struggle with Fe4S4 because its electrons interact strongly with each other, a regime where the approximations standard in quantum chemistry break down and even the best density-matrix renormalization group methods consume weeks of CPU time. The quantum run used IBM's sample-based quantum diagonalization method, seeded with a physically informed starting state, running up to 4,000 two-qubit operations on 72 qubits, with 5 million samples taken in total. The classical partner was RIKEN's ROQUO system, a supercomputer built on NVIDIA GB200 NVL4 systems, which handled the diagonalization step.
The combined run landed within 4.3 millihartree of the density-matrix renormalization group reference, the best classical benchmark for this molecule. Chemical accuracy is defined as 1 millihartree or better, so this result does not cross that line. But the 79x improvement over IBM's own previous attempt demonstrates that the quantum hardware is contributing genuine signal rather than amplifying noise. That distinction matters. For years, hybrid quantum-classical calculations were difficult to distinguish from pure-classical baselines, because quantum hardware errors swamped whatever quantum signal existed. The Fe4S4 result, funded by Japan's NEDO under its quantum-supercomputer hybrid platform project, makes a credible case that Nighthawk-generation hardware is now good enough to contribute something a classical supercomputer cannot replicate on its own. A second IBM-RIKEN project, modeling protein-ligand binding for molecules of up to 12,635 atoms versus a prior benchmark of 303, is a finalist for the 2026 ACM Gordon Bell Prize, the most recognized award in high-performance computing. That nomination reflects peer review, not vendor marketing.
The Bear Case: Why Critics Are Not Wrong to Push Back
The bear case, however, is straightforward: the co-design platform does not exist yet, and the hardware that would make it useful is years away. Grassellino herself called it a "proposed platform." No site, start date, or budget has been named. The 13,000x decoder speedup, while real, compares against a single-threaded exact decoder rather than the strongest competing GPU-based decoder. When IBM and AMD's open-source Pelagos decoder is the comparison point instead, the advantage narrows: Pelagos achieves 97x faster throughput than a single CPU core and scales 7.5x across eight AMD Instinct GPUs on ORNL's Lux system. That suggests GPU-accelerated quantum decoding is a broadly solved problem across multiple vendors, not a unique Nvidia capability.
The risk is also that fault-tolerant quantum computing remains far from practical scale. The most demanding algorithms for chemistry and cryptography require millions of error-corrected logical qubits. IBM's 2027 target of 1,040 physical qubits falls orders of magnitude short of that threshold. Skeptics point out that the intermediate milestone of "quantum-centric supercomputing," where quantum processors contribute signal to hybrid calculations, depends on that signal being distinguishable from hardware noise. The Fe4S4 result is 79x better than IBM's own prior attempt, but still does not reach chemical accuracy. Each improvement narrows the gap, but the distance remaining is large enough that the timeline to practical fault-tolerant advantage is genuinely uncertain, and no independent analysis has validated IBM's specific roadmap targets.
The competitive landscape adds pressure from multiple directions. Google's Willow chip, demonstrated in late 2024, was the first superconducting processor to perform below the fault-tolerance threshold rather than above it, a critical milestone IBM has not yet claimed. Microsoft is pursuing topological qubits, which promise lower error rates but have not produced a programmable device at scale. IonQ and Quantinuum's trapped-ion systems achieve higher fidelity per qubit than superconducting approaches but run more slowly. IBM's bet is that raw scale (more qubits, faster sampling, better decoders) beats fidelity at the current margin, and Nighthawk supports that bet for some workloads. Whether it holds across the full range of algorithms that matter commercially is an open question. The Quantum Insider described the partnership as "a landmark in quantum-HPC integration," which is accurate if the hardware follows the plan, and premature if it does not.
What the November 16 announcements do establish clearly is Nvidia's strategic position. By acquiring SchedMD and shipping CUDA-Q QEC as open source, Nvidia is positioning its GPU stack as the classical backbone of every quantum computing center, regardless of which QPU vendor wins the hardware race. Even if IBM's quantum hardware takes another five years to demonstrate fault-tolerant advantage, Nvidia's decoders and schedulers are already in production at 30 supercomputing centers today. That is not a quantum bet; it is a compute infrastructure bet that happens to pay off under multiple quantum scenarios.
Nvidia's real quantum bet is not that it wins the qubit race: it is that every qubit, from every vendor, needs a GPU to decode its errors.
Key Takeaways
- 13,000x faster decoding: Nvidia's CUDA-Q QEC 0.9.0 compresses a 12-week quantum error-correction study to 45 minutes using the TCBP-JOBS algorithm on IBM Nighthawk hardware, with 4.08ms feed-forward latency.
- IBM Nighthawk r2 is 25x faster than Heron: At 120 qubits, 7,500 two-qubit operations, and 100,000 samples per second, with r3 targeting 520 qubits and the 1,040-qubit modular system coming in 2027.
- 30 supercomputing centers, 50+ QPUs live: Slurm 26.11's QRMI integration makes quantum processors schedulable alongside CPUs and GPUs across IBM, Quantinuum, Pasqal, IQM, Alice and Bob, and MQSC hardware.
- Fe4S4 accuracy improved 79x: IBM and RIKEN's joint chemistry run reached within 4.3 millihartree of the DMRG reference on 72 qubits; the protein-ligand follow-up modeling 12,635 atoms is a finalist for the 2026 ACM Gordon Bell Prize.
- The co-design platform is planned, not shipped: No site, start date, or budget; the hardware phase requires Fermilab cavity interconnects that do not yet exist, and IBM's 1,040-qubit modular target remains a 2027 commitment.
Questions Worth Asking
- If Nvidia's decoder works with any QPU vendor's hardware through QRMI, does IBM's specific quantum roadmap matter to Nvidia's business at all, or does Nvidia win regardless of which qubit technology dominates?
- The Fe4S4 result is 79x better than IBM's own prior attempt but still short of chemical accuracy. At what accuracy level does hybrid quantum-classical chemistry become commercially useful, and who sets that threshold?
- With 30 supercomputing centers already running quantum hardware under QRMI, is the barrier to quantum access now mostly bureaucratic rather than technical, and if so, who controls the queue?