COMPUTE ATLAS Supercomputer supply-chain graph
1153 systems 482 sites Sourced data

Blog/2026-10-07

· Compute Atlas

Quantinuum Helios: what the claims support

Helios has 98 barium-ion qubits at 99.921% two-qubit fidelity and a reported 48 error-corrected logical qubits. The H-series reached QV 2^25. Here is what is measured, what is postselected, and who pays.

Quantinuum’s Helios, launched commercially on 5 November 2025, is a 98-qubit trapped-ion machine whose headline numbers are a two-qubit gate fidelity of 99.921 percent across all qubit pairs and logical-qubit counts of 94 and 48. Its predecessor, the 56-qubit H2, holds the highest quantum volume in our benchmark table, 2^25. The fidelity figures now have peer-reviewed backing, including a June 2026 Nature paper with Sandia National Laboratories. The logical-qubit claims are real but narrower than they sound: they rely on error-detecting codes with postselection, and the 48 corrected qubits are produced by a distance-4 concatenated code. What is not known is how this architecture scales beyond a single junction-and-ring trap, whether Helios’s quantum volume matches H2’s, or what any customer pays, because the public sources give no unit price.

From Honeywell to a listed company

The H-series began inside Honeywell. Our H1 page records a launch in October 2020 with 10 qubits, later raised to 20 fully connected qubits. Quantinuum was formed in 2021 from Honeywell Quantum Solutions and Cambridge Quantum, per our company record, and in 2026 became a public company: Renaissance Capital reported a 3 June 2026 pricing of 28 million shares at $60, about $1.7 billion, on Nasdaq under QNT. The company’s second-quarter 2026 results show $8 million of revenue, guidance of $28 million to $32 million for the year, and $2.1 billion in cash and short-term investments at 30 June. The gap between a $1.7 billion raise and $8 million of quarterly revenue is the clearest sign that Helios is being financed as a platform bet, with customers still mostly research and enterprise pilots.

Three generations in one table

SystemQubitsIonHeadline figures
H120not checkedQV 2^20, 99.914% two-qubit fidelity (Apr 2024)
H256YbQV 2^25 (Sep 2025), certified randomness (2025)
Helios98Ba99.921% two-qubit, 99.9975% single-qubit (Nov 2025)

The H1 and H2 rows are from our benchmark table: QV 1,048,576 and 99.914 percent on H1-1 were reported in April 2024, and QV 33,554,432 on H2 in September 2025, four times the roughly 8 million reported in May 2025. Quantinuum’s H2 product page lists 56 fully connected qubits, single-qubit fidelity above 99.99 percent and two-qubit fidelity above 99.9 percent, and states that subscription, Azure Quantum and the Oak Ridge National Laboratory quantum user program are access routes.

Two points of care. A QV of 2^25 by definition means 25-qubit circuits passed a heavy-output test, not that the machine behaves like 2^25 of anything; and the test is sensitive to connectivity, which all-to-all ion traps supply for free. We found no QV figure for Helios in the sources we read. Our benchmarks deep dive explains why QV stops being informative as machines grow past about 50 qubits.

The Helios hardware

The Helios paper, posted on 7 November 2025 with 186 contributors led by Anthony Ransford, describes a QCCD trap with a rotatable ion storage ring joining two operation regions through a junction. Qubits are barium-137 hyperfine states, a change from H1 and H2’s ytterbium, and connectivity is all-to-all. The paper reports average infidelities of 2.5 x 10^-5 for single-qubit gates, 7.9 x 10^-4 for two-qubit gates and 4.8 x 10^-4 for state preparation and measurement, averaged over operational zones. Those numbers match the 99.9975 and 99.921 percent in the launch release. Quantinuum’s launch blog calls the junction the first commercial implementation of its kind and describes ring storage, cache and logic zones.

Third-party review arrived in June 2026. Quantum Computing Report reports that Sandia and Quantinuum published peer-reviewed Helios data in Nature on 18 June 2026 (another report gives 17 June), including a state preparation and measurement fidelity of 99.967 percent, using random Clifford circuits with mid-circuit measurement and random circuit sampling. The Nature paper was a collaboration with the vendor, so it is peer-reviewed and Sandia-checked, not an audit. Sandia staff quoted by The Quantum Insider said the most important property of today’s machines is reliability, not speed.

Power and the 60 kW figure

The only power figure we found for Helios comes from Quantinuum’s August 2026 Oracle announcement, which says the system draws approximately 60 kW without an HVAC system, under one percent of leading supercomputers. That is a vendor statement. Because the release says “without an HVAC system”, facility cooling sits outside the figure, and we found nothing on how the number was measured. The launch blog also claims that a classical supercomputer would need more power than all stars in the visible universe to match Helios’s sampling in equal time, against about one data-center rack for Helios. We treat that comparison as rhetoric: it prices a task with no application against an estimated classical algorithm.

Sampling and certified randomness

Quantinuum’s sampling claims go back to the 56-qubit H2. In June 2024 it said H2-1 reached a cross-entropy score around 0.35, over 100 times Google’s 2019 result, and that the machine was impossible to simulate fully. The Helios paper’s abstract says random circuit sampling shows it operating well beyond classical simulation, without a cross-entropy value in the abstract.

The more concrete use of sampling is certified randomness, published in Nature in March 2025. A client sends random circuits to the untrusted quantum server, which must return results within a time limit of about 2.2 seconds per sample. Classical supercomputers then check the results through cross-entropy scores. On H2-1 the protocol certified 71,313 bits of entropy. Verification needed 1.1 x 10^18 floating-point operations per second sustained across Frontier, Summit, Perlmutter and Polaris, and the strictest security level yielded about one bit per second. The authors state that the adversary model is restricted and that assumptions limit production use. This is the clearest case where a quantum machine produced something a supercomputer had to verify at great cost, and also the clearest case of how small the output is: 71,313 bits is less than 9 kilobytes.

What the fidelity numbers do and do not show

A two-qubit infidelity of 7.9 x 10^-4 means roughly one gate in 1,270 goes wrong. As our own arithmetic, a circuit with 1,000 such gates and no other error source would run cleanly about 45 percent of the time, which is why Quantinuum pairs raw fidelity with error detection for larger programs. The figure is an average over operating zones, and the paper itself says the three error types are not fundamentally limited. It is not directly comparable with the best-coupler errors that IBM posts for Heron and Nighthawk, which are selected minima on a single pair, or with Rigetti’s whole-device median, which is another statistic again. Our benchmarks deep dive walks through why these should not be ranked against each other.

The speed side is less visible. None of the Helios sources we read gives a circuit-layer rate comparable to IBM’s CLOPS, so a reader cannot compare throughput. That matters for workloads needing millions of shots, such as error-mitigated chemistry, and the headline fidelity figures do not address it.

The logical-qubit claims, read closely

Quantinuum’s launch claims were 94 error-detected logical qubits, globally entangled; 50 error-detected logical qubits in a magnetism simulation; and 48 error-corrected logical qubits at 2:1 encoding. The supporting preprint, posted on 25 February 2026 by 45 authors, is titled “Computing with many encoded logical qubits beyond break-even” and covers 48 to 94 logical qubits. Quantum Computing Report says the codes are iceberg codes: a [[k+2, k, 2]] error-detecting family, and a two-level concatenated version with distance 4 for correction. Logical error rates came out 10 to 100 times lower than the physical ones, with logical gates near one error in 10,000.

The caveats are in the same sources. The Quantum Insider reports that runs where errors were detected were discarded, calls the work “partially fault-tolerant,” and says the techniques may not scale indefinitely to very large circuits. A distance-2 error-detecting code cannot correct errors on its own; it flags them, and the experiment throws away flagged shots. The count of 94 therefore describes a code that fits 94 logical qubits in 98 physical ones with detection, not a machine that corrects 94. The 48 corrected qubits use distance 4 but come from a concatenated, high-rate construction rather than a surface code. Quantinuum’s earnings release adds a claim of near five-nines logical fidelity, which we could not tie to a specific experiment. Compare Google’s Willow, which shows a below-threshold surface-code memory with exponential suppression but only one logical qubit. The two results measure different things, and neither has yet shown a universal set of logical gates running at fault-tolerant error rates.

Quantum-HPC partnerships and who pays

Quantinuum’s integration strategy has four threads.

RIKEN is the oldest. In January 2024 it selected the H1 for a quantum-HPC hybrid platform with Fugaku, in a NEDO-commissioned project involving SoftBank, the University of Tokyo and Osaka University. Our Reimei record notes a 20-qubit machine at the Wako campus announced in February 2025, with the Fugaku link described by RIKEN as a planned milestone. In April 2026 RIKEN procured an H2, delivered near Tokyo, to replace the H1 in the Reimei-Fugaku platform. A RIKEN researcher said that with 56 qubits the team expects the machine to help demonstrate quantum advantage.

NVIDIA is the second thread. The Helios launch included GB200 integration through NVQLink and the Guppy programming language, and the launch blog says Grace Hopper GPUs are used for real-time error decoding.

Singapore is third. A strategic partnership with the National Quantum Office, A*STAR and Singapore’s national quantum and supercomputing centres promises a Helios installation in 2026, the first outside the United States, immediate cloud access, and an R&D centre; the announcement gives no funding amounts or commercial terms. We could not confirm that the installation has been completed.

Oracle is the newest. Reported on 11 August 2026 (Quantum Computing Report gives 12 August), the multi-year deal places a Helios in a US Oracle Cloud Infrastructure AI data center, to be offered as a managed service with a preview in coming months. No financial terms were disclosed. Customers of the launch included Amgen, BMW Group, JPMorganChase and SoftBank. None of the pieces we read gives a Helios price, so who pays what is known only at the level of named partners.

What we could not confirm

We found no Helios quantum volume, no published unit price, and no completion date for the Singapore or Oracle installations. The 60 kW figure is a company statement and its scope beyond the quantum system is unclear. We could not tie the five-nines logical-fidelity claim to a named experiment. We did not read the Nature papers directly and relied on arXiv versions and trade reports for them. The two sources dating the Oracle announcement disagree by a day. We did not verify the company’s quarterly bookings, which appear only in a secondary summary. We did not fetch Quantinuum’s roadmap for later machines, so we say nothing on it.

What to watch

Watch for the first logical-level result on Helios that does not rely on postselection, a Helios quantum volume figure, and the first independent measurement from the Oracle or Singapore units. Watch also whether RIKEN’s H2 and Fugaku link produces a published hybrid workflow with real throughput numbers, since the Reimei-Fugaku coupling is the one place the “quantum-HPC” label can be tested against a named supercomputer. Finally, whether the DARPA evaluation described in our benchmarks coverage, where Quantinuum is one of 11 Stage B companies, produces any published scores.

Sources

deep-divequantumbenchmarks