Why Trapped Ions, Why Now? Helios Comes to OCI
Quantinuum's Helios is being installed in an Oracle Cloud Infrastructure data center. It is the right machine for this moment, and the physics explains why.
 |
| The Helios chip, which generates tiny electromagnetic fields to trap single atomic ions hovering above the chip to be used for computation - From "https://www.quantinuum.com/blog/helios-delivers-quantum-advantage-with-real-world-impact" |
By Dr. Sanjay Basu
Oracle is installing a Quantinuum Helios quantum computer in one of its US-based OCI AI data centers. Quantinuum and Oracle announced the partnership on August 11, 2026. Helios will sit alongside OCI's GPU and HPC infrastructure and will be offered through an OCI quantum service. That is the news, and I will leave the corporate reasoning to the press releases.
What I want to write about is a more durable question. Why a trapped-ion machine? A decade of quantum computing has been a contest between physical modalities. Superconducting circuits, neutral atoms, trapped ions, photons, semiconductor spins, and the still-hypothetical topological qubit. Each has a credible story about the future. My argument here is narrower and, I think, more defensible. Right now, in 2026, trapped ions are the best modality for the work that matters most: turning noisy physical qubits into reliable logical ones. Helios is the clearest current example of that claim.
"Right now" matters in that sentence. Modalities rise and fall on engineering milestones, not on destiny. But when you ask which platform currently offers the best combination of fidelity, connectivity, coherence, and mid-circuit control, the answer keeps coming back to ions in traps.
The problem every modality is actually solving
The industry has mostly stopped arguing about qubit counts, and it was right to stop. A thousand noisy qubits that decohere before a meaningful circuit finishes are a physics experiment, not a computer. The real goal is fault tolerance. Encoding a logical qubit redundantly across many physical qubits, detecting errors as they happen, and correcting them faster than they accumulate, is the holy grail for now.
Fault tolerance has a threshold structure. If physical error rates stay below a code's threshold, adding more physical qubits suppresses logical errors exponentially. Above it, adding more hardware makes things worse. And the distance below threshold decides the cost. A system at 99.9% two-qubit fidelity may need several times fewer physical qubits per logical qubit than a system at 99.5% to reach the same logical error rate.
So the useful questions for any modality are:
How good are the gates, especially two-qubit gates, which dominate error budgets?
How long do qubits stay coherent, measured against how long operations take?
Which qubits can interact with which?
Can you measure some qubits mid-circuit, reset them, and branch on the result while the rest keep computing?
How does all of this hold up as the machine grows?
Trapped ions currently lead on the first four questions. The fifth is where the argument gets interesting.
What a trapped-ion qubit is
Helios holds individual atomic ions in electromagnetic traps. Its computational qubits are barium-137 ions (¹³⁷Ba⁺). Two states in the ion's ground-state hyperfine manifold represent |0⟩ and |1⟩. Helios also carries ytterbium-171 ions for sympathetic cooling: they remove unwanted motion from the ion crystal without taking part in the computation. Lasers prepare states, rotate individual qubits, entangle pairs, and read out results.
The deepest advantage here is simple. Nature manufactures the qubits. Every ¹³⁷Ba⁺ ion is identical to every other. There is no fabrication variance, no two-level-system defects in an amorphous oxide, no qubit-to-qubit frequency spread to calibrate around. The quantum energy structure is fixed by atomic physics. The engineering difficulty moves elsewhere, into trapping, optical delivery, motional control, transport, and readout. Those problems are hard, but they are organizable. They yield to better electrodes, better lasers, and better control software.
Contrast that with a superconducting transmon. It is an engineered circuit. It is exceptionally fast and it benefits from semiconductor fabrication, but every transmon is slightly different, and its coherence is limited by materials defects that the field has spent fifteen years fighting. A transmon qubit is made. A trapped-ion qubit is found.
Barium matters too. Quantinuum's documentation notes that barium's key optical transitions are at visible wavelengths or longer, which allows a more scalable laser architecture built from mature optical components, unlike the ultraviolet transitions of earlier ion species. In the published Helios work, gates use pairs of 515 nm laser beams separated by the qubit's roughly 8.04 GHz splitting. Laser phase and intensity noise, spontaneous emission, ion motion, and electrode voltage noise all still add error. High fidelity is an engineering accomplishment, not a birthright of the modality.
Modality by modality where ions win today
Let me be fair to the competition, because every alternative has real strengths.
Superconducting qubits (Google, IBM, Rigetti, and others) are the speed champions. Gate times run in the tens of nanoseconds, roughly a thousand times faster than trapped-ion gates. Fabrication borrows from the semiconductor industry. Google's Willow processor showed below-threshold surface-code scaling in 2024, a landmark result. But superconducting two-qubit fidelities at scale have generally sat around 99.5–99.8%. Coherence times are measured in hundreds of microseconds. And connectivity is fixed by wiring. A qubit talks to its lattice neighbors, and everything else costs SWAP gates. The surface code was chosen partly because it tolerates nearest-neighbor connectivity. It is a code designed around a hardware constraint. It also carries a heavy overhead of physical qubits per logical qubit.
Neutral atoms (QuEra, Atom Computing, Pasqal, Infleqtion) are the most serious challenger, and I say that without hedging. Like ions, the qubits are identical atoms. Optical tweezers can hold arrays of thousands of them and physically move atoms to reconfigure connectivity. The Harvard–QuEra work reached 99.5% parallel two-qubit gate fidelity, and logical-qubit demonstrations have come quickly. On raw scale, neutral atoms are ahead. But their two-qubit fidelities still trail the best ion results. Atom loss is a real error channel. Mid-circuit measurement without disturbing neighboring atoms remains harder than in ions, and the cycle time of a full QEC round is not obviously faster. In my view, neutral atoms are the modality most likely to challenge ions over the next few years. They have not overtaken them on the metrics that decide error correction.
Photonic systems (PsiQuantum, Xanadu) offer room-temperature operation for much of the stack and a natural fit with networking. But photons do not interact easily. Entanglement is probabilistic, optical loss is the dominant error, and the fusion-based architectures that fix these problems need very large machines before they become useful. The approach is a long-horizon bet on manufacturing scale.
Semiconductor spin qubits (Intel, Diraq, Quantum Motion) promise compatibility with CMOS fabs and tiny footprints. Current devices are small, and uniformity across a chip remains a materials problem, the same challenge superconducting circuits face.
Topological qubits (Microsoft) would, if realized, build error protection into the physics itself. The scientific community is still debating the evidence for the underlying states. It is a fascinating program and not yet a computing platform.
Against that field, trapped ions bring a set of advantages that no other modality currently matches all at once:
The dominant fact in that table is the ratio of coherence time to gate error. Ions are slow, but they are patient. A qubit that stays coherent for seconds can afford a gate that takes a hundred microseconds. What matters for correctness is not the absolute speed but how many high-fidelity operations fit inside a coherence window, and ions win that ratio comfortably.
Helios is a machine for moving qubits
The textbook picture of trapped-ion computing is a static chain of ions illuminated by lasers. Helios does not work that way. It uses a quantum charge-coupled device (QCCD) architecture. Ions are physically shuttled between storage and operation regions, much as data moves between memory and execution units in a classical processor. If you design computer architectures, this should feel familiar. It is a memory hierarchy for qubits.
Helios has 98 physical qubits. Its trap includes a storage ring, an X-shaped junction, cache and leg storage regions, and eight operation zones. Ions travel through the junction to zones where gates, cooling, preparation, and measurement occur, and up to 16 qubits can be processed in those zones at a time. The runtime schedules transport and operations dynamically, including in response to mid-circuit measurement outcomes.
This is what all-to-all connectivity means in Helios. There is no physical coupler between every pair of qubits. Instead, any two qubits can be brought together. Transport still takes time and adds error, and the Helios paper measures memory and leakage errors tied to transport. But the junction avoids the growing SWAP overhead of fixed-lattice routing. For a real algorithm, the question is not "can these two qubits interact?" It is "how much transport, waiting, cooling, and control does this interaction cost?"
That design choice links directly to the modality argument. Connectivity is not a convenience in error correction. It decides which codes you can run. A fixed 2D lattice pushes you toward the surface code. A machine with flexible connectivity can run high-rate codes, such as certain quantum LDPC and concatenated codes, that encode more logical qubits per physical qubit. The industry's roadmap toward practical fault tolerance increasingly runs through these more efficient codes, and they favor architectures in which any qubit can reach any other.
The numbers, read honestly
The peer-reviewed Helios paper reports average infidelities across operation zones of roughly 2.5×10⁻⁵ for single-qubit gates, 7.9×10⁻⁴ for two-qubit gates, and 4.8×10⁻⁴ for state preparation and measurement. That corresponds to an average two-qubit gate fidelity of 99.921%, at 98-qubit scale. The last part is the important one. Many modalities can post a record fidelity on a pair of qubits in a lab. Holding it across a full production system is the harder achievement.
Quantinuum has also demonstrated 48 logical qubits on the 98-physical-qubit system, an encoding ratio of about two physical qubits per logical qubit. Surface-code approaches on other platforms often assume hundreds of physical qubits per logical qubit for comparable protection. That ratio is the modality argument in one number: high fidelity plus flexible connectivity lets you choose efficient codes.
"48 logical qubits" still needs a careful reading. It does not mean Helios can run an arbitrarily long fault-tolerant program on 48 fully protected qubits. The protection achieved depends on the code, the circuit, and the experiment. The right engineering questions are how logical error scales with circuit depth, which logical operations are protected, how much decoding is required, and how much time and hardware each logical operation consumes.
Fidelity is the prerequisite; useful throughput is the goal.
The honest weaknesses
A thesis is only as good as its treatment of the counterarguments, and trapped ions have two real ones.
Speed. Trapped-ion gates are slower than superconducting gates by roughly three orders of magnitude, and transport adds more time. For algorithms that need billions of logical operations, wall-clock time matters, and a slow clock can turn a feasible computation into an impractical one. The ion community's answer is parallelism across operation zones, faster gate schemes, and, most importantly, fewer total operations through more efficient codes. The bet is that doing fewer, better operations beats doing many fast, noisy ones. For today's workloads, where the binding constraint is reliability rather than runtime, that bet is sound.
There is also a less obvious upside to the slower clock. A superconducting QEC cycle runs in about a microsecond, so the classical decoder must keep up with a relentless stream of syndrome data, which pushes teams toward custom FPGA and ASIC decoders. A trapped-ion QEC cycle runs on a much longer timescale, which gives classical decoders room to breathe. That is one reason a trapped-ion machine sits comfortably next to general-purpose GPU and CPU infrastructure: the classical side of the loop does not have to be exotic.
Scale. The biggest trapped-ion systems hold around a hundred qubits. Neutral-atom arrays hold thousands. Scaling QCCD traps means more electrodes, more zones, more optical channels, and eventually linking modules through photonic interconnects. Quantinuum's roadmap, with Sol and then Apollo targeting fault tolerance by the end of the decade, depends on solving these problems. The barium switch to visible wavelengths and the junction-based trap design are exactly the kinds of choices that make scaling tractable. But tractable is not solved, and I would not pretend otherwise.
My position is not that ions win forever. It is that they win now, on the metric that decides whether fault tolerance arrives: error rates per operation across a full, connected, controllable system.
Why the machine belongs next to classical compute
One architectural point follows directly from the physics, so I will make it without speculating about anyone's business strategy.
A useful quantum computation is almost never quantum alone. A variational chemistry workflow prepares a molecular Hamiltonian classically, compiles circuits, submits them to the QPU, collects measurements, estimates an objective, and updates parameters, often hundreds of times. Error-correction research adds decoding and feedback. Classical simulation on GPUs validates results and benchmarks against the best classical methods. The QPU is one stage in a heterogeneous pipeline.
A high-fidelity trapped-ion machine increases the value of that pipeline, because its results are worth more classical processing. When your quantum output is dominated by noise, careful classical post-processing buys little. When it is reliable, the surrounding compute can do real work. Putting Helios in the same data center as GPUs, storage, and enterprise data services turns what would be a remote experiment into a component of a system.
To be precise, this does not mean a cloud network controls individual ion gates. The time-critical control stays inside the quantum system. Colocation affects the surrounding workflow: data movement, orchestration, governance, and iteration.
What I would measure
Deployment is the beginning of the evaluation, not the end. For representative workloads, I would want end-to-end results on:
These measurements will also test the modality thesis. If trapped ions really are ahead, it should show up as better answer quality per unit of cost and time, not only as prettier fidelity numbers.
The bet on patience
Every quantum modality is a philosophical stance on where difficulty should live. Superconducting qubits put it in materials and fabrication, then answer with speed. Neutral atoms put it in optical control at scale, then answer with numbers. Photonics puts it in manufacturing, then answers with the promise of volume. Trapped ions put it in trapping, transport, and optical engineering, and answer with quality: identical qubits, long coherence, and gates accurate enough that error correction becomes efficient rather than merely possible.
In 2026, quality is the constraint that binds. That is why a trapped-ion machine is the right quantum computer to put in a hyperscale data center today. Helios will not replace a GPU cluster, and it does not need to. The interesting future is heterogeneous: each processor doing the part of the problem it suits, under one operational model. The quantum part of that future will be decided by who reaches reliable logical qubits first. For now, the atoms in the trap are ahead.
I included IonQ's 99.99% result as independent evidence that the fidelity lead belongs to trapped ions as a whole, not just to Quantinuum. Sources:
Comments
Post a Comment