Most quantum computers are sold on qubit counts. Quantinuum Helios is sold on accuracy. Its 98 barium-ion qubits run two-qubit gates at 99.921% fidelity, and in June 2026 the company behind it raised $1.68 billion in an upsized Nasdaq IPO under the ticker QNT.
That combination makes Helios the most scrutinized trapped-ion machine on the market. Investors want to know if the hardware justifies the valuation. Buyers want to know if it does anything a GPU cluster can't.
This full guide covers how Helios works, what its benchmarks really say, the logical-qubit and Helix error-correction results, the Sol and Apollo roadmap, the business behind QNT, how it compares with IBM, Google and IonQ, and the honest limits. For the head-to-head with its closest rival, see Inside Deep Tech's Quantinuum vs IonQ comparison.
Last updated: October 7, 2026.
Key Takeaways
- Helios has 98 barium qubits with 99.921% two-qubit and 99.9975% single-qubit fidelity.
- A ring-and-junction ion trap gives all-to-all connectivity, at about 55 ms per full circuit layer.
- Helix logical operations beat physical Clifford gates by 4.28x, no post-selection, in September 2026.
- Sol is due in 2027 and Apollo, the fault-tolerant machine, in 2029, per Quantinuum's Q2 update.
- Honest limits: slow gates, $8 million in Q2 revenue, and roadmap machines that don't exist yet.
What Quantinuum Helios is
Helios is Quantinuum's third-generation commercial quantum computer, after the H1 and H2 systems. It launched on November 5, 2025, and is sold both through Quantinuum's cloud and as an on-premises system, per the launch announcement.
The launch named four early customers: Amgen, BMW Group, JPMorganChase and SoftBank Corp. The same release announced a partnership with Singapore's National Quantum Office that puts a Helios in the country (Quantinuum).
The peer-reviewed description of the machine, "A 98-qubit trapped-ion quantum computer with all-to-all connectivity," was published by Nature on June 17, 2026. It's the reference to use when a vendor slide and a spec sheet disagree.
Helios specifications at a glance
| Parameter | Helios value | Source |
|---|---|---|
| Physical qubits | 98 ions of barium-137 | Nature |
| Connectivity | All-to-all, via ion transport | Data sheet |
| Two-qubit gate fidelity | 99.921% (typical error 8 x 10^-4, max 20 x 10^-4) | Launch, Data sheet |
| Single-qubit gate fidelity | 99.9975% | Launch |
| Operation zones | 8 zones; two-qubit gates in 4 of them | arXiv paper |
| Depth-1 layer time (98 qubits) | About 55 ms | arXiv paper |
| Power draw | Under 40 kW, base unit | Quantinuum |
| Classical co-processor | NVIDIA GPUs for real-time decoding | Quantinuum |
Source: Quantinuum Helios data sheet, Helios arXiv paper, Nature (2026), and Quantinuum launch release.
How Helios works: barium ions, a ring and a junction
Helios is a quantum charge-coupled device (QCCD). Instead of wiring every qubit to its neighbors, it physically moves ions around a microfabricated surface trap, so any qubit can be paired with any other. The trap is 15.3 mm long, with a 2.8 mm storage ring connected to the logic region by a four-way "X" junction (arXiv).
Think of it like a spinning hard drive. Ions rotate through the ring (memory), pass the junction into a cache, move into eight operation zones in batches of up to 16 qubits, then park in leg storage while the next batch is loaded. Sorting happens in parallel with laser cooling, which H1 and H2 couldn't do (Quantinuum).
The junction also cuts wiring. Helios uses 2.8 control signals per qubit, against 4.8 on H2 and 9.9 on H1, per the Helios paper. That ratio is the real scaling story, because control electronics, not qubits, tend to be the bottleneck.
Why Quantinuum switched from ytterbium to barium
H1 and H2 used ytterbium qubits, which rely mostly on ultraviolet lasers. Barium works with visible-wavelength lasers, where mature, cheaper industrial optics already exist, and it lets the machine detect "leakage" errors directly at the atomic level (Quantinuum).
Ytterbium didn't disappear. Helios still co-traps ytterbium-171 ions as coolants, which absorb motional energy from the barium qubits between operations (arXiv). For how this design compares with fixed-qubit chips, see the team's superconducting vs trapped-ion qubits comparison.
What the Helios benchmarks actually show
Component numbers are strong. Averaged across all operation zones, the Nature paper reports single-qubit gate errors of 2.5 x 10^-5, two-qubit gate errors of 7.9 x 10^-4 and state preparation and measurement errors of 3.3 x 10^-4 (Nature).
System numbers are where to look harder. In full 98-qubit random circuits, the effective two-qubit error, which folds in memory and single-qubit errors, came out at 2.0 x 10^-3 (arXiv). That's roughly 2.5 times the headline gate error, by Inside Deep Tech's arithmetic from the same paper, and it's the number that limits real programs.
Quantinuum also ran random circuit sampling, the benchmark Google used for its 2019 supremacy claim. It says matching Helios classically in the same time would take more power than every star in the visible universe, while Helios used roughly the power of one data center rack (Quantinuum). That claim depends on today's best tensor-network methods, which keep improving.
The most useful science result so far is a superconductivity simulation. Using up to 90 qubits (72 system qubits plus 18 ancillas), Helios simulated a 6x6 Fermi-Hubbard lattice and measured light-induced pairing correlations (Quantinuum). It's a real physics experiment, not a commercial workload yet.
Logical qubits on Helios: 94, 50, 48, and what each means
Quantinuum's launch figures need decoding. On 98 physical qubits, it reported 94 error-detected logical qubits in a GHZ entangled state, 50 error-detected logical qubits in a magnetism simulation, and 48 error-corrected logical qubits, all beating their physical equivalents (Quantinuum).
Error detection and error correction aren't the same thing. Detection flags a corrupted run so it can be thrown away. Correction fixes errors while the program keeps running, which is what long algorithms need. The 48 error-corrected qubits used a 2:1 physical-to-logical encoding built with code concatenation (Quantinuum).
For the background on codes, thresholds and why logical qubits matter, read Inside Deep Tech's guide to quantum error correction in 2026.
Helix: the September 2026 error-correction result
The bigger 2026 result is Helix, an error-correction architecture built on a [[20,2,6]] code and tested on Helios. The September 2 preprint reports 4.6 x 10^-5 error per logical qubit per correction cycle and 2.8 x 10^-4 error per two-qubit logical Clifford operation, without post-selection (arXiv).
Quantinuum says that logical error rate is a 4.28x improvement over the physical Clifford error, and that adaptive syndrome extraction cut physical gates by 33% and run time by 23% (Quantinuum). The preprint also entangled Helix qubits with a distance-5 surface code block, a step toward injecting the non-Clifford "magic" states universal computing needs (arXiv).
Software: Guppy, Nexus, and NVIDIA in the loop
Helios shipped with a new stack. Guppy is a Python-based language for dynamic programs that mix quantum and classical logic, with loops, early exits and qubits allocated on the fly (Quantinuum). Programs written for QIR-compatible tools, including Qiskit, Cirq and CUDA-Q, can also target it (arXiv).
NVIDIA sits inside the control loop. Helios integrates NVIDIA GPUs for real-time decoding, and the launch added NVIDIA GB200 integration through NVQLink (Quantinuum). Nexus, the cloud platform, had more than 200 organizations and more than 1,000 active developers by September 2026, per Quantinuum.
Distribution is widening. In August 2026, Oracle and Quantinuum agreed to deploy Helios inside a U.S. Oracle Cloud Infrastructure AI data center, sold as an OCI service next to GPU and HPC capacity (Oracle).
What customers actually run on Helios
The launch customer list reads like a map of where quantum chemistry and simulation might pay first. Per Quantinuum's launch release, each one is exploring a different problem:
- Amgen: hybrid quantum machine learning for biologics discovery, as both investor and research collaborator.
- BMW Group: materials research on fuel cell catalysts.
- JPMorganChase: research into advanced financial analytics.
- SoftBank Corp.: organic materials for next-generation batteries, optical switches and solar cells.
Quantinuum's own 2026 application work points the same way. Its Q2 update lists a new parallel quantum phase-estimation algorithm for molecular properties, an AI-driven quantum simulation with NVIDIA and a Fortune 100 pharma company, and magnetic-materials simulations it says go beyond practical classical methods (Quantinuum).
The Helios paper adds one more category: cryptographic protocols that generate certified randomness (arXiv). That's a narrow but real use, because the output can be verified as genuinely random.
Notice what's missing. None of these is a production workload with a published cost saving. They're research programs, which is exactly what you'd expect from a 98-qubit machine in 2026.
The roadmap: Helios, Sol, Apollo, Lumos
| System | Target | Planned capability | Status, Oct 2026 | Source |
|---|---|---|---|---|
| H2 | Predecessor | 56 qubits, racetrack trap | Still offered | Quantinuum |
| Helios | 2025 | 98 qubits, all-to-all, barium | Launched Nov 5, 2025 | Quantinuum |
| Sol | 2027 | Hundreds of qubits, 2D grid, two-qubit errors below 2 x 10^-4, about 2x faster | Trap chip back from fab, in validation | Quantinuum, Q2 |
| Apollo | 2029 | Universal, fully fault tolerant; hundreds of logical qubits | On schedule, per Q2 update | Quantinuum, Q2 |
| Lumos | 2033 goal | Utility-scale concept for DARPA QBI | DARPA Stage B | Quantinuum |
Source: Quantinuum 2024 roadmap, technical perspective, Q2 2026 results, DARPA QBI release.
Sol is the important hinge. It moves from Helios's ring to a full 2D grid of junctions, the layout Quantinuum says scales to Apollo. Sol is planned with about 100 junctions and circuits of up to 10,000 quantum operations (Quantinuum). Until Sol ships, the grid architecture is a plan, not a product.
Manufacturing is the next constraint. In September 2026, the Commerce Department finalized a CHIPS award of up to $100 million for trapped-ion bottlenecks such as low-loss integrated photonics and reliable lasers (NIST). Quantinuum named GlobalFoundries, on 300 mm wafers, and Monarch Quantum as onshore partners (Quantinuum).
The business behind Helios: QNT by the numbers
| Metric | Value | Source |
|---|---|---|
| IPO | 28 million shares at $60; $1.68 billion gross, June 2026 | Quantinuum |
| Debut market value | $15.7 billion at the first-day close | CNBC |
| Q2 2026 revenue | $8 million, up 279% year over year | Quantinuum |
| Q2 2026 GAAP net loss | $597 million (includes non-cash IPO charges) | Quantinuum |
| Q2 2026 adjusted EBITDA loss | $68 million | Quantinuum |
| Cash and short-term investments | $2.1 billion at June 30, 2026 | Quantinuum |
| 2026 revenue guidance | $28 million to $32 million | Quantinuum |
Source: Quantinuum IPO release, CNBC, and Quantinuum Q2 2026 results.
Honeywell, which formed Quantinuum by merging its quantum unit with Cambridge Quantum in 2021, remains the majority shareholder, per CNBC. The prospectus showed first-quarter 2026 revenue of $5.24 million, down 73% from a year earlier, a reminder of how lumpy quantum contracts are (CNBC).
By Inside Deep Tech's arithmetic, the $15.7 billion debut value is more than 500x the midpoint of 2026 revenue guidance (CNBC, Quantinuum). That multiple prices in Apollo, not Helios. For how QNT stacks up against listed peers, see the best quantum computing stocks to watch in 2026.
How Helios compares with IBM, Google, and IonQ
| Program | Qubit type | Latest public milestone | Fault-tolerance target | Source |
|---|---|---|---|---|
| Quantinuum Helios | Trapped ion (barium), QCCD | 98 qubits, 99.921% two-qubit fidelity; Helix logical Clifford beats physical | Apollo, 2029 | Quantinuum, arXiv |
| IBM | Superconducting | Roadmap via Loon, Kookaburra, Cockatoo | Starling, 2029: 200 logical qubits, 100 million operations | IBM |
| Google Willow | Superconducting | 105 qubits; below-threshold surface code (Dec 2024) | Not dated publicly | Google, Nature |
| IonQ | Trapped ion (electronic control) | Above 99.99% two-qubit fidelity on R&D prototypes (Oct 2025) | 256-qubit systems planned for 2026; 2 million qubits by 2030 | IonQ |
Source: Quantinuum, Helix preprint, IBM, Google, Nature, IonQ.
The trade-off is speed against connectivity. In a QUOPS benchmark built with Sandia National Laboratories, Quantinuum says Willow and IBM's Boston system run much faster gates, while Helios reaches larger successful circuits at lower operation rates (Quantinuum). QUOPS was developed with Quantinuum's input, so treat it as useful but not neutral.
IonQ's fidelity claim is also worth reading closely. Its 99.99% two-qubit result came from R&D prototypes, not a commercial system (IonQ). For the full four-way picture, see Inside Deep Tech's IBM vs Google vs IonQ vs Quantinuum hardware race.
Honest limits of Quantinuum Helios
The real limits as of October 2026:
- Speed. A full 98-qubit layer takes about 55 ms, and each two-qubit gate about 70 microseconds. Superconducting rivals run much faster gates, which matters for error-correction cycles and wall-clock time.
- System error vs gate error. The effective two-qubit error in full circuits was 2.0 x 10^-3, against 7.9 x 10^-4 for the gate alone. Transport and memory errors eat much of the headline fidelity.
- Spec spread. The data sheet lists a typical two-qubit error of 8 x 10^-4 but a maximum of 20 x 10^-4, and notes performance varies by operation zone.
- Detection is not correction. The 94-logical-qubit result is error detection, which discards bad runs. Fully corrected, non-Clifford computation at scale is still Apollo's job.
- Unproven roadmap. Sol's trap chip is still in validation, and Apollo, Lumos and the 2D grid exist on paper.
- Revenue. Quantinuum guides to $28 million to $32 million of 2026 revenue against a $15.7 billion debut valuation. Customers are still mostly researching, not deploying.
- Self-reported benchmarks. Most Helios numbers come from Quantinuum's own papers. The Nature paper is peer reviewed; the Helix preprint isn't yet.
Frequently asked questions
What is Quantinuum Helios?
Quantinuum Helios is a 98-qubit trapped-ion quantum computer that launched on November 5, 2025. It uses barium ions moved around a chip-based trap to give all-to-all connectivity at 99.921% two-qubit fidelity.
How many qubits does Quantinuum Helios have?
Helios has 98 physical qubits. Quantinuum has encoded up to 94 error-detected and 48 error-corrected logical qubits on them in specific experiments.
Is Quantinuum Helios the most accurate quantum computer?
Among commercially available systems, Quantinuum says yes, based on physical two-qubit gate fidelity as of December 31, 2025 (Quantinuum). IonQ has reported above 99.99% on R&D prototypes, which aren't commercial systems.
How do you get access to Quantinuum Helios?
You can use Helios through Quantinuum's cloud platform, Nexus, or buy an on-premises system (Quantinuum). Oracle also plans an OCI service running Helios in a U.S. data center.
What comes after Helios?
Sol is planned for 2027 with hundreds of qubits on a 2D grid, followed by Apollo, a fully fault-tolerant machine targeted for 2029 (Quantinuum).
Is Quantinuum a public company?
Yes. Quantinuum listed on Nasdaq under the ticker QNT in June 2026 after an upsized IPO at $60 per share. Honeywell remains the majority shareholder.
Can Helios break encryption?
No. Breaking RSA needs large-scale, fault-tolerant machines with far more logical qubits than Helios has. Quantinuum's own fault-tolerant target, Apollo, is planned for 2029.
What to watch next
Three signals will tell you whether Helios is a stepping stone or a peak: a peer-reviewed Helix paper with non-Clifford logic, Sol shipping in 2027 as the Q2 update promises, and the Oracle OCI service moving from preview to paying customers (Oracle).
If you're evaluating quantum access this quarter, benchmark Helios against a superconducting system on your own circuits, and price time-to-solution, not fidelity alone. Start with the rival that matters most.



