July 21, 2026 – Quantinuum (NASDAQ: QNT) and SoftBank Corp. (TOKYO: 9434) published Quantum Computing Frontiers, a 64-page joint white paper subtitled “A Use-Case Timeline for Quantum Chemistry and Topological Data Analysis.” SoftBank put up the download page the following day and ran a second news item on July 29. Twenty contributors are credited: sixteen from Quantinuum, four from SoftBank.
The paper maps two application domains against four generations of Quantinuum hardware: Helios (current), Sol (2027), Apollo (2029), and Lumos, the utility-scale design Quantinuum submitted to DARPA’s Quantum Benchmarking Initiative and now targets for the 2030s. The table that matters is in Chapter 5, and it places chemically accurate simulation with more than 100 logical spin orbitals in the Lumos column. Helios and Sol get roughly ten.
Chapter 3 covers the chemistry experiment. The team ran quantum phase difference estimation on ethylene at its conical intersection, using a minimal two-electron, two-orbital active space and the STO-3G basis, tapered by particle-number and spin symmetry down to a two-qubit effective Hamiltonian. Both logical qubits were encoded in the Steane [[7,1,3]] color code. Arbitrary-angle logical rotations were implemented through recursive gate teleportation with error detection rather than magic-state distillation. Circuits ran on H2-2, Helios-1, and Helios-1E between January 12 and March 27, 2026, with an unencoded logical baseline on the H2-2E emulator. From 1,400 circuit samples the estimated energy gap came out at −0.0025561(24) Hartree against a reference value of −0.002559 Ha.
The comparison across the three conditions produced a mixed result. On H2, the encoded circuit with Steane error-correction gadgets showed less decoherence than the encoded circuit without them at depths of three or more, and more at depths of one and two. Against the unencoded physical implementation, the encoded circuit was worse at every depth tested. The authors state the outcome without softening it: “break-even with respect to the physical baseline was not reached.” They attribute the shortfall to memory errors accumulating during the extended idle periods that encoding, decoding, and correction impose on trapped ions.
Chapter 4 turns to topological data analysis. The algorithm estimates normalized Laplacian moments of Hodge Laplacians on clique complexes, following Akhalwaya and colleagues in Quantum, where classical trace estimation requires a sample count scaling as $$O(2^d)$$ in the moment order and the quantum version requires a count independent of it. Validation used a complete k-partite graph on 20 vertices, 1,000 shots, run on a noisy emulator of Helios rather than on the machine.
The application study used the BUPT telecommunications fraud dataset. Betti numbers vanished above dimension one because the graph proved too sparse to contain higher-order cliques, so both the Betti numbers and the Laplacian moments were computed exactly, on classical hardware. Adding Laplacian-moment features to a graph neural network raised macro recall from 0.8317 ± 0.0219 to 0.8449 ± 0.0237 across ten random seeds, with a two-sample t-test returning p = 0.026. At a fixed precision of 0.9, the authors convert that recall gain into roughly USD 0.220 billion of avoided annual fraud losses, calculated against an assumed USD 10 billion pool, about a quarter of the Communications Fraud Control Association’s $38.95 billion estimate for 2023 global telecom fraud.
Both roadmap tables are framed by the authors as an assumption-driven envelope rather than a deterministic forecast. Duncan Jones, General Manager of Quantinuum’s Applications Group, said in the announcement that organizations “do not need to wait for large-scale, fault-tolerant systems” before looking for value. SoftBank CTO Ryuji Wakikawa framed the open question as which problem classes become executable at which stage of hardware maturity.
My Analysis
The most useful sentence in 64 pages is an admission
Most vendor white papers are engineered so that no sentence in them can ever turn out to be wrong. This one contains a sentence that could easily have been left out and wasn’t, and it is the reason to read the document: on a two-qubit problem, with a distance-3 code, on the most accurate commercial hardware in the world, error correction made the answer worse.
Take that in both directions. Break-even at the component level is settled science by now, including on this exact platform: logical CNOT and SPAM, Bell and cat state preparation, logical qubit teleportation, Google’s below-threshold surface-code memory, the Harvard/QuEra 48-logical-qubit sampling circuit. Break-even for a complete algorithm, including the non-Clifford rotations that phase estimation actually requires, has not been shown by anyone. The distance between those two facts is where my CRQC Quantum Capability Framework parks most of its uncertainty, and this paper puts a measured number on it instead of a roadmap arrow.
The calibration data carries a second detail that deserves more attention than it will get. Helios-1, the newer machine with better physical error rates, fitted a decay rate of 0.1086(201) against H2-2’s 0.0741(67). The newer hardware performed worse. The authors explain why, and the explanation has nothing to do with qubits: Helios has a real-time engine that schedules ion transport dynamically, which suits a circuit full of repeat-until-success operations, but automatic dynamical decoupling had not yet been compiled into the Helios stack when these runs happened, so coherent memory noise suppression had to be inserted by hand at the logical circuit level. Two effects competing, one of them a missing compiler pass.
That is the entire argument for treating quantum computing as a systems integration problem, compressed into one table. I have made the same point about Quantinuum’s iceberg-code result, and it holds here: the road to fault tolerance runs through the control stack at least as much as through the trap.
What the $220 million figure actually measures
I expect the fraud number to be the thing that travels, so let me be precise about what produced it.
No quantum computer was involved. The BUPT graph is too sparse to contain the higher-order cliques the algorithm needs, every Betti number above dimension one came out zero, and the team therefore computed the Laplacian moments exactly, classically, by diagonalization. What the experiment demonstrates is that classically computed Laplacian moments are a modestly useful extra feature for a GNN: macro recall of 0.8449 ± 0.0237 with them against 0.8317 ± 0.0219 without, error bars that overlap comfortably, ten seeds, p = 0.026. The paper says outright that this is not a demonstration of quantum advantage.
Then it converts a 1.3-point recall gain at one chosen operating point into $220 million a year by scaling across a quarter of a global fraud-loss estimate. Each half is defensible on its own. Placed next to each other, on the strength of ten seeds, they produce a figure that will outlive every caveat attached to it.
The resource estimates in the same chapter make the honest version of the point. Running fifty blocks of the Laplacian moments algorithm on graphs the size of the real two-hop ego-graphs in that dataset, which reach 20,000 vertices, needs T-gate counts in the billions. That is Apollo-and-later territory, not NISQ. Calling TDA the shorter commercialization path is fair only if you mean the shorter path to classical Laplacian features.
The dates run later than the pitch
Now the part I actually care about. Both companies want earlier dates. SoftBank is building an investment case for quantum AI data centers, and Quantinuum listed publicly this year, which makes its roadmap part of its equity story. Here is what they published anyway, for the logical layer:
- Helios, now: ~10 spin orbitals, logical error rate ~10⁻³, code distance 3 to 5
- Sol, 2027: still ~10 spin orbitals, ~10⁻⁴, distance 5 to 7
- Apollo, 2029: ~100 spin orbitals, ~10⁻⁷, distance 7 to 9, excited-state QPE
- Lumos, 2030s: 100+ spin orbitals at ≤10⁻¹⁰, chemical accuracy, scalable workflows
The physical-level row is more generous, reaching about 50 spin orbitals on Helios today with error detection, but only for shallow circuits, since success probability decays exponentially in circuit size. Deep circuits need the logical layer, and the logical layer reaches a hundred spin orbitals in 2029.
FeMoco, the nitrogenase cofactor that anchors half the industrial pitch for quantum chemistry, is standardly modeled with a 152-spin-orbital active space. On this table, the molecule everyone quotes does not fit until the generation after Apollo. I argued something close to that in Quantum Chemistry’s Honest Ledger, and it carries more weight coming from the people selling the machine. (The roadmap tables still date the Helios column to 2025 in a paper published in July 2026, which tells you roughly how long a joint document takes to clear two companies’ review.)
How this reads against the CRQC clock
The paper says nothing about cryptography, and I am not going to pretend otherwise. The logical-layer numbers are readable against the CRQC question anyway, because they are the same numbers.
Quantinuum describes Apollo publicly as thousands of physical qubits supporting hundreds of logical ones, arriving in 2029. Craig Gidney’s 2025 estimate for RSA-2048 puts the requirement under a million noisy physical qubits running for about a week. Two to three orders of magnitude of physical qubits separate the most aggressive generation on a vendor roadmap from the cheapest published attack, and Lumos in the 2030s is the entry to that regime rather than the end of it. Nothing here shifts my Q-Day view. What it adds is a data point from a self-interested source that happens to fall on the conservative side, which is the most useful kind of data point there is.
The milestone to watch is the one the authors set for themselves. They name a repeat of this experiment on next-generation hardware, with a higher-distance code, as the natural benchmark for algorithm-level fault tolerance. That is a dated and falsifiable target, and it is theirs rather than mine. If a two-qubit chemistry circuit still runs better without error correction than with it when Sol arrives in 2027, the logical-layer trajectory is slower than the roadmap assumes, and every downstream date in both tables moves with it.
