Everyone counts qubits. This week IBM, AMD and Oak Ridge put their money on a different bottleneck: the conventional silicon that must read every error, decide what went wrong, and answer before the next one lands.
A Formula 1 pit crew can change four tyres in about two seconds, yet nobody says the car is the slow part. Now imagine the race is run in microseconds and the crew has to read the mind of the car before it finishes the lap. That is the situation in quantum computing, and this week it got a new address.
On 5 November, IBM, AMD and Oak Ridge National Laboratory announced a platform that puts a 120-qubit IBM Nighthawk next to ORNL’s Umbra supercomputer, built on AMD Instinct MI355X GPUs. The stated purpose is not more qubits. It is to train and test decoders: the classical software that interprets error signals fast enough to keep a calculation alive.
Our lens for this issue: when a new technology looks like a hardware race, ask which conventional system it secretly depends on. Carry one question through the pages that follow. In your own roadmap, what is the unglamorous component whose latency quietly sets the speed of everything else?
A quantum machine makes a mistake every few hundred operations. Useful programs need billions of clean ones. The bridge between the two is a piece of ordinary electronics racing a clock measured in millionths of a second.
Think of a quantum bit as a spinning coin balanced on its edge. It is wonderfully powerful and it falls over if anyone breathes near it. Today’s physical qubits, SiliconANGLE reports, “make errors roughly once every few hundred operations,” while a code-breaking algorithm like Shor’s needs billions of error-free steps. No amount of better engineering closes a gap of that size in one move. So the field uses a trick: spread one logical bit of information across many physical qubits, and keep checking them against each other.
You cannot simply read a qubit to see if it has flipped; reading destroys what you are protecting. Error correction sidesteps this with helper qubits that measure only the relationships between neighbours, like a referee who reports “those two disagree” without saying who is right. Each round of checks produces a stream of tiny clues called a syndrome.
Someone has to turn those clues into a verdict, and the verdict has to arrive before the next round of errors. That someone is the decoder. It is conventional hardware, running conventional algorithms, under a deadline set by quantum physics.
Google’s Willow experiment gives the scale. Its surface-code cycle took 1.1 microseconds, and its real-time decoder at code distance 5 had an average latency of 63 microseconds (arXiv 2408.13687). By our arithmetic, that is roughly 57 cycles of the quantum chip elapsing for each decision. It works because decoding is pipelined, but it shows the decoder is not idle slack. It is the critical path.
Scale it up and the problem steepens. A 2026 preprint on surface-code decoding says the data rates can reach “gigabits or even terabits per second,” and warns of a “backlog problem”: if the decoder falls behind, the slowdown compounds exponentially.
The 5 November announcement pairs IBM Nighthawk processors with ORNL’s Umbra, which uses AMD Instinct MI355X GPUs and 5th Gen EPYC CPUs. IBM’s Quantum System API gives researchers pulse-level control, mid-circuit measurement and the raw measurement signals, so decoders can be trained on real hardware noise rather than idealised models. IBM frames it as a step toward Starling, its planned fault-tolerant machine for 2029.
IBM’s Relay-BP decoder is a refinement of belief propagation, a message-passing method in which each check tells its neighbours its best guess. Its authors call it “inherently parallel,” which is why it suits chips built for parallel work. A follow-up FPGA implementation for the 144-qubit “gross” code reports a belief-propagation iteration time of 24 nanoseconds and an average decoding time under 1 microsecond per cycle, assuming circuit-level error below 3×10⁻³ (arXiv 2510.21600).
Google’s Willow result showed the logic holds: each time the code distance grew by 2, logical errors fell by a factor Λ of 2.14 ± 0.02, on a 101-qubit distance-7 surface code. A factor above 1 is the sign you are below threshold, where adding qubits helps rather than hurts.
The U.S. Department of Energy’s Q Competition targets 100 logical qubits running 100,000 hard operations by late 2028. IBM’s own current mark, per the platform materials, is up to 70 logical qubits and 438 hard operations. The gap in operations is about 230×.
Craig Gidney’s 2025 estimate says RSA-2048 could fall to a quantum computer with fewer than a million noisy qubits in under a week, against about 20 million qubits in 2019’s estimate. Better decoding and better algorithms, not just better chips, drive that curve.
The claimed prize in all of these sources is fault-tolerant quantum computing by the end of the decade. None of it is demonstrated yet. Three readings follow, flagged as ours rather than the papers’.
First, latency becomes a spec. Buyers of cloud quantum time have asked about qubit counts and fidelities. Expect the question “what is your decode latency, and where does the decoder live?” to join them. The closer the classical hardware sits to the qubits, the less time the signal spends travelling.
Second, the GPU vendors are quantum vendors. AMD’s presence in an IBM announcement is the signal. If the decoder is the pacing item, accelerator makers hold a seat at a table they were never invited to.
Third, software gets a moat. Quantinuum’s Nexus platform, per The Qubit Report, compiles ordinary programs into fault-tolerant form. Whoever hides error correction from the programmer owns the developer.
For a technology leader: the near-term action is not buying qubits. It is mapping which of your workloads would tolerate a slow, expensive, error-corrected machine, and starting the post-quantum cryptography migration now.
1 · Check. Helper qubits compare neighbouring data qubits and flip a signal if they disagree. The data is never read directly.
2 · Decode. The pattern of disagreements, the syndrome, streams to a classical chip. It guesses the most likely error that caused it.
3 · Correct. The fix is applied, or tracked in software, before the next round of errors can pile on top.
The consensus reading of this week: IBM keeps pushing toward Starling, so watch qubit counts and logical-qubit records. Our Move 37 is less comfortable. The announcement is not about the quantum chip at all. A bank of AMD GPUs next to a 120-qubit processor is a statement that the scarce resource is real-time classical compute colocated with the cryostat, and that is a resource you can rent, bundle and meter.
Our inference, and it could be wrong: quantum cloud pricing may drift from “per shot” toward something that includes decoding capacity, the way cloud storage learned to charge for throughput as well as terabytes. Qubit makers are easier to copy than a tuned decoder stack trained on a specific device’s noise.
The announcement has no latency number. Neither IBM’s blog nor the press release states a decoder latency, a logical error rate or a logical-qubit count for the new platform. It is a testbed, announced as one.
The big data-rate figure is a vendor projection. Riverlane’s “100 TB/s” for large machines comes from a company that sells decoders. An academic paper puts rates at gigabits to terabits per second, orders of magnitude lower.
Qubit quality may dissolve the problem. Alice & Bob estimate that cat qubits could cut an RSA-2048 attack from about a million qubits and 500 MW to about 23,000 and 12 MW. If hardware errors fall far enough, the decoder’s job shrinks. That physics is not yet built.
The 2028 gap is the real test. Going from 438 to 100,000 hard operations is about 230×. If it closes, the decoder story gains weight; if it slips, the bottleneck may be elsewhere.
In the Phase 3 INTerpath-001 trial of Merck and Moderna’s mRNA vaccine in resected stage 2b–4 melanoma (1,137 patients), the combination reduced risk of recurrence or death by 42%, a hazard ratio of 0.58. Just as striking for operations: 99% of patients in the vaccine arm received a manufactured vaccine, each encoding up to 34 neoantigens unique to the patient. The efficacy data were first presented at ESMO last month; BioSpace’s report is dated 6 November, and no approval date is given.
A Power Peak Digest roundup of 5 November reports that India’s Union Cabinet approved Rs 14,500 crore of viability-gap funding for about 92 GWh of storage over FY27 to FY31: 60 GWh of pumped hydro, 30 GWh of batteries and 2 GWh of emerging technologies. The same roundup cites an IEA report putting India’s renewable capacity past 300 GW. This is a secondary news roundup, so treat the figures as reported, not independently confirmed.
Astronauts completed installation of the eighth new ISS roll-out solar array in a 6-hour, 59-minute spacewalk, the 286th for station assembly and maintenance. Each new array generates about 8 kilowatts, against 19 kilowatts for each original wing when new, but the eight together make up to 30 percent more power available to the station. The final array unrolled to 63 feet long and 20 feet wide.
Three stories, one pattern. A quantum computer cannot run without a conventional computer reading its mistakes in microseconds. A personalised cancer vaccine only matters if a factory can make one for 99% of patients. A solar array refit only counts when eight roll-outs and five years of spacewalks add up to usable watts. In each case the headline technology is carried by a less glamorous system, and that system sets the speed.
The discipline is plain: when you evaluate a breakthrough, price the dependency. Ask what must be true of the unglamorous layer, and whether anyone has measured it. This week the honest answer for quantum decoding is that nobody has published the number that matters most for the new platform.
1. IBM Quantum blog, ORNL–AMD error-correction co-design platform, 5 Nov 2026 — https://www.ibm.com/quantum/blog/ornl-amd-error-correction-co-design-platform
2. PR Newswire, “Oak Ridge National Laboratory, IBM and AMD launch new co-design platform for quantum error correction,” 5 Nov 2026 — https://www.prnewswire.co.uk/news-releases/oak-ridge-national-laboratory-ibm-and-amd-launch-new-co-design-platform-for-quantum-error-correction-302929996.html
3. SiliconANGLE, “IBM, AMD and Oak Ridge join forces to launch quantum computing error-correction platform,” 5 Nov 2026 — https://siliconangle.com/2026/11/05/ibm-amd-and-oak-ridge-join-forces-to-launch-quantum-computing-error-correction-platform/
4. The Qubit Report, weekly round-up, week ending 7 Nov 2026 — https://thequbitreport.com/quantum-computing-weekly-round-up/2026/11/06/quantum-computing-weekly-round-up-week-ending-november-7-2026
5. Google Quantum AI, “Quantum error correction below the surface code threshold,” arXiv 2408.13687 — https://arxiv.org/abs/2408.13687
6. Müller et al. (IBM), “Improved belief propagation is sufficient for real-time decoding of quantum memory,” arXiv 2506.01779 — https://arxiv.org/abs/2506.01779
7. “Real-time decoding of the gross code memory with FPGAs,” arXiv 2510.21600 — https://arxiv.org/abs/2510.21600
8. “Information-efficient decoding of surface codes,” arXiv 2512.14255 — https://www.arxiv.org/pdf/2512.14255
9. Riverlane press release, 8 Jan 2025 (vendor projection of 100 TB/s) — https://www.riverlane.com/press-release/riverlane-ushers-in-the-year-of-quantum-with-nature-electronics-publication
10. Gidney, “How to factor 2048 bit RSA integers with less than a million noisy qubits,” arXiv 2505.15917 — https://arxiv.org/abs/2505.15917
11. BioSpace, Merck/Moderna INTerpath-001, 6 Nov 2026 — https://www.biospace.com/business/merck-moderna-serve-99-of-trial-patients-with-mrna-vaccine-as-personalized-shots-near-market
12. Power Peak Digest, 5 Nov 2026 — https://powerpeakdigest.com/power-sector-news-roundup-for-november-5-2026
13. Spaceflight Now, 5 Nov 2026 — https://spaceflightnow.com/2026/11/05/spacewalkers-complete-five-year-effort-to-upgrade-space-station-power-supply/
Note: the “So What?”, The Contrarian and Long View are Inflection’s own inference. The decoder-latency ratio (≈57 cycles) is our arithmetic from sources 5. The Alice & Bob figures are a company estimate relayed by source 4; the 100 TB/s figure is a vendor projection. Sources were read via summarised fetches; check the originals before relying on exact wording.
What is the unglamorous system in your roadmap whose latency quietly sets the pace of everything else?
Researched, written & designed with Claude. Typeset in Poppins & Lora on the Anthropic palette.