In September 2025, a team at Caltech assembled 6,100 qubits — the largest array anyone had ever built. Two months later, Quantinuum launched a commercial machine with 98.
The 98-qubit machine is the more capable computer, and it is not close.
That sounds like a contradiction. It is actually the single most useful thing to understand about quantum hardware right now: "number of qubits" is not a measure of computing power, and the five competing platforms are not even measuring the same quantity when they report it. A qubit built from a superconducting circuit and a qubit built from a single atom differ in how long they survive, how accurately they can be operated, which other qubits they can talk to, and whether they can be operated together at all.
This is a comparison of those five platforms — what each one physically is, what it is genuinely good at, and where it breaks. If you want the underlying physics first, start with quantum computing explained; if you want to know what these machines are for, see what quantum computers are good for.
Why Qubit Counts Don't Compare
Start with the two records, because the gap between them explains the whole field.
The Caltech result is real and impressive. Physicists led by Manuel Endres trapped 6,100 caesium atoms in a grid of optical tweezers — focused laser beams that each grip a single atom. They held those atoms in superposition for about 13 seconds, roughly ten times longer than previous arrays of this kind, and manipulated individual qubits with 99.98% accuracy. The work was published in Nature on 24 September 2025.
What that array had not yet done is entangle its qubits. Caltech's own announcement describes linking the qubits in entanglement as the next step. Without entanglement there are no two-qubit gates, and without two-qubit gates there is no quantum computation — only 6,100 very well-behaved atoms sitting in a grid.
Quantinuum's Helios, launched commercially on 5 November 2025, has 98 trapped barium ions. Every one of them can be entangled with any other. Its two-qubit gate fidelity is 99.921%, and the company reports 48 fully error-corrected logical qubits running on those 98 physical ones.
So one machine reports 6,100 qubits and produces zero logical qubits. The other reports 98 and produces 48. The count tells you about the apparatus; the fidelity and connectivity tell you about the computer.
Five Ways to Hold a Qubit
Every platform has to solve the same contradiction. A qubit must be isolated enough from its environment to hold a fragile quantum state, and connected enough to the outside world to be written, manipulated and read. Isolation and control pull in opposite directions, and each technology picks a different compromise.
The Five Platforms, Side by Side
Figures are the best publicly reported results as of October 2026, not typical production performance.
| Platform | What the qubit physically is | Operating temperature | Main obstacle |
|---|---|---|---|
| PlatformSuperconducting | What the qubit physically isA circuit of metal on a chip | Operating temperature10–20 millikelvin | Main obstacleShort coherence, limited connectivity |
| PlatformTrapped ion | What the qubit physically isA single electrically charged atom | Operating temperatureRoom-temperature vacuum | Main obstacleSlow gates, hard to scale |
| PlatformNeutral atom | What the qubit physically isAn uncharged atom held by lasers | Operating temperatureRoom-temperature vacuum | Main obstacleGate fidelity behind ions |
| PlatformPhotonic | What the qubit physically isA single particle of light | Operating temperature2–4 kelvin for detectors | Main obstaclePhotons get lost |
| PlatformTopological | What the qubit physically isA state spread across a device | Operating temperatureMillikelvin | Main obstacleNot yet demonstrated |
No platform leads on every axis
Speed, memory, connectivity, scale and manufacturability trade against each other. That is why serious money is still spread across all five.
Superconducting Qubits: Fast, Cold, and Industrially Mature
A superconducting qubit is a tiny loop of metal — usually aluminium — patterned onto a chip and cooled until it loses all electrical resistance. That is the same phenomenon, and the same temperature constraint, described in why we still lack room-temperature superconductors. Electrical current flowing in that loop behaves quantum-mechanically, and two of its energy levels are used as the 0 and 1 of the qubit.
The great advantage is speed and manufacturability. These are chips, made with adapted semiconductor fabrication, and their gates run in nanoseconds — thousands of times faster than a trapped-ion gate. This is the platform behind Google and IBM.
Google's Willow, announced on 9 December 2024, carries 105 of these qubits. Its significance was never the count. Willow was the first chip to show error correction working below threshold: as Google scaled its arrays from 3×3 to 5×5 to 7×7 qubits, the error rate halved at each step rather than growing. That result is covered in detail in our piece on Google's Willow chip and the 2026 roadmap.
The weakness is memory. Willow's T1 coherence times approach 100 microseconds — a genuine improvement over its predecessor, and still vastly shorter than a trapped ion's. Superconducting qubits also connect only to their immediate neighbours on the chip, so entangling two distant qubits means routing through a chain of intermediate operations, each adding error.
IBM's answer is scale and engineering discipline rather than exotic physics. Its Nighthawk processor carries 120 qubits, and its published roadmap targets Starling in 2029: roughly 200 error-corrected logical qubits built from on the order of 10,000 physical ones, running circuits of 100 million gates.
Trapped Ions: The Accuracy Leader
A trapped-ion qubit is a single atom with an electron stripped away, suspended in a vacuum by oscillating electric fields and manipulated with precisely tuned lasers. The qubit is encoded in the atom's internal electronic states.
Two properties make this the accuracy leader. First, the qubits are identical — every barium ion in the universe is exactly the same, so there is no fabrication variation to calibrate around, unlike superconducting circuits which must each be tuned. Second, ions can be physically shuttled around the trap, which gives all-to-all connectivity: any qubit can be entangled directly with any other, with no routing overhead.
That is why Helios gets 48 error-corrected logical qubits out of 98 physical ones. In October 2025, IonQ went further still, reporting 99.99% two-qubit gate fidelity — the first time anyone has crossed "four nines" — on R&D prototypes using barium ions and chip-based electronic control. That figure is from prototypes, not a shipped system, and the distinction matters.
The cost is speed. Ion gates are driven by lasers and run in microseconds, far slower than superconducting gates. Scaling is the other problem: more ions in one trap means more vibrational modes to control, so the path forward involves connecting many small traps together rather than building one enormous one.
Neutral Atoms: The Scaling Bet
Neutral atoms are the same idea as trapped ions with the charge removed. Because the atoms are uncharged they do not repel one another, so they can be packed into large arrays — which is exactly how Caltech reached 6,100. Each atom is held by an optical tweezer, and gates are performed by exciting atoms into large, puffed-up Rydberg states that let neighbours interact.
The scaling advantage is real and the coherence is excellent. The honest caveat is the one the Caltech result illustrates: assembling a large array and operating it as a computer are different achievements, and published two-qubit gate fidelities on this platform still trail trapped ions. The bet is that the scaling advantage eventually outweighs that gap. It has not been settled.
Photonic Qubits: Examining the "No Cryogenics" Claim
Photonic quantum computing encodes qubits in single particles of light moving through waveguides on a chip. Its headline selling point is that photons barely interact with their environment, so the qubits do not need to be chilled into stillness.
That claim needs a qualifier. The chip avoids the dilution refrigerator, but the single-photon detectors that read the result are superconducting devices and need cooling of their own. PsiQuantum's Omega system runs at roughly 2–4 kelvin — substantially warmer than the 10–20 millikelvin a superconducting qubit demands, and a genuinely easier engineering problem, but a long way from room temperature. The company has contracted an industrial cryogenic plant for its Brisbane machine.
The central difficulty is loss. A photon that scatters out of a waveguide is simply gone, and loss accumulates with every component the light passes through. Competing approaches handle this differently — building in enough redundancy to tolerate loss, or encoding qubits in a way that resists it — but no photonic machine has yet matched the gate fidelities demonstrated on ions.
Topological Qubits: The One That Isn't Proven
The remaining platform is different in kind: it does not yet clearly exist.
The idea is elegant. Store quantum information not in one particle but in a global property of the system — something that local noise cannot disturb, because disturbing it would require changing the system's overall shape. Such a qubit would be error-resistant by construction rather than by correction.
In February 2025 Microsoft announced Majorana 1, a chip it said was built on a topological core and designed to accommodate eight qubits. The accompanying paper in Nature was considerably more measured than the press release. The authors wrote that "these measurements do not, by themselves, determine whether the low-energy states detected by interferometry are topological."
Nature's editors attached a note stating that "the results in this manuscript do not represent evidence for the presence of Majorana zero modes in the reported devices." Winfried Hensinger of the University of Sussex put the gap plainly in APS Physics: "The peer-reviewed publication is quite clear [that it contains] no proof for topological qubits. But the press release speaks differently."
Microsoft has been here before — it retracted an earlier Majorana claim after problems emerged in the data analysis. If topological qubits work, they change the economics of the entire field. That is precisely why the evidence bar should be high.
Why Fidelity Beats Count
The reason a 98-qubit machine outruns a 6,100-qubit array comes down to arithmetic that compounds brutally.
Errors multiply. If each gate succeeds with probability f, a circuit of n gates finishes cleanly with probability f raised to the power n. Run a 1,000-gate circuit and the difference between fidelities that look almost identical becomes total.
At 99% fidelity, a 1,000-gate circuit succeeds about 0.004% of the time. At 99.9%, about 37%. At 99.99%, about 90%. Each extra nine buys roughly ten times more circuit — which is why "four nines" was treated as a milestone rather than an incremental gain.
This arithmetic is idealised; real machines also contend with crosstalk, readout error and qubits that decay while waiting. It understates the problem rather than overstating it.
Error correction is what eventually breaks this ceiling, by spending many physical qubits to build one reliable logical qubit. But the exchange rate depends on the physical error rate: the better your gates, the fewer physical qubits each logical qubit costs. Fidelity is not an alternative to scale. It is what determines how much your scale is worth.
How to Read the Next Qubit Announcement
Hardware announcements arrive constantly and are written to impress. Four questions separate substance from theatre:
- Physical or logical qubits? Logical qubits are error-corrected and worth vastly more. A claim that does not say which is usually quoting physical.
- What is the two-qubit gate fidelity? Single-qubit fidelity is far easier and is often the number quoted when the two-qubit figure is unflattering.
- Has anything been entangled? An array that has not demonstrated entanglement is not yet a computer.
- Prototype or shipped system? Record fidelities frequently come from lab prototypes with a handful of qubits.
A fifth question is worth asking of any announcement touching cryptography: does it change the timeline for breaking encryption? Usually it does not — see will quantum computers break encryption for why the gap remains large.
The Bottom Line
There is no winning qubit type yet, and anyone who tells you otherwise is selling something.
Superconducting circuits have speed, manufacturing maturity and the strongest error-correction results. Trapped ions have the best fidelity and the only all-to-all connectivity, at the cost of slow gates. Neutral atoms have the clearest path to very large arrays, with gate quality still catching up. Photonics sidesteps the dilution refrigerator but has not solved loss. Topological qubits would be the most elegant answer if anyone can demonstrate one.
What has genuinely changed is that the field now has a shared scoreboard. Below-threshold error correction, logical-qubit counts and two-qubit gate fidelity are measurable and comparable in a way that raw qubit counts never were. Follow those three numbers across platforms — and treat the headline count as what it is, a description of the apparatus rather than of the computer. More on the field's direction at our quantum computing hub.
Frequently Asked Questions
Which type of qubit is best?
None of them, yet. Trapped ions lead on gate fidelity and connectivity, superconducting circuits lead on speed and manufacturing maturity, and neutral atoms lead on array size. Each leads on a different axis, and the axes trade against one another. The question that will eventually be settled is which platform reaches large numbers of error-corrected logical qubits first — and as of late 2026 that is genuinely open, which is why major investment continues across all of them.
What is the difference between a physical and a logical qubit?
A physical qubit is one piece of hardware — a circuit, an atom, a photon. It makes errors constantly. A logical qubit is a single reliable qubit built out of many physical ones, with error correction running continuously to detect and fix faults. The exchange rate depends on hardware quality: Quantinuum's Helios produces 48 error-corrected logical qubits from 98 physical ions, while IBM's 2029 target expects roughly 200 logical qubits from on the order of 10,000 physical ones.
Why do quantum computers need to be so cold?
Superconducting qubits need temperatures of 10–20 millikelvin because heat is random energy, and random energy destroys the fragile superposition the qubit depends on. At those temperatures the circuit superconducts and stray thermal excitations are suppressed. Not every platform needs this: trapped ions and neutral atoms operate in room-temperature vacuum chambers, cooled by lasers rather than refrigerators, and photonic chips need only enough cooling for their detectors.
Did Microsoft actually build a topological qubit?
That claim is disputed. Microsoft announced Majorana 1 in February 2025 as a chip with a topological core. The accompanying Nature paper was more cautious, stating that its measurements do not by themselves determine whether the states detected are topological, and Nature's editors attached a note saying the results do not represent evidence for Majorana zero modes in the devices. Several physicists have said publicly that the peer-reviewed evidence does not support the press release.
How many qubits do you need to do something useful?
It depends entirely on the task and on error rates, not on a single threshold. Useful quantum simulation of small molecules may need a few hundred good logical qubits; breaking widely used public-key encryption needs far more. Because error correction consumes physical qubits to produce logical ones, improving gate fidelity reduces the physical qubit count needed for the same job — which is why fidelity records matter as much as size records.
Why do companies report such different qubit numbers?
Because they are counting different things and optimising for different goals. A neutral-atom array can hold thousands of atoms because uncharged atoms pack easily, while a trapped-ion machine holds dozens because ions repel each other and each added ion complicates control. Neither number says how accurately the qubits can be operated or whether they can be entangled, which is why cross-platform comparisons based on count alone are misleading.
Sources
- Meet Willow, our state-of-the-art quantum chip — Google
- Caltech Team Sets Record with 6,100-Qubit Array — California Institute of Technology
- Introducing Helios: The Most Accurate Quantum Computer in the World — Quantinuum
- Interferometric single-shot parity measurement in InAs–Al hybrid devices — Aghaee et al., Nature 638, 651–655 (2025)
- Experts Weigh in on Microsoft's Topological Qubit Claim — American Physical Society, Physics
- Experimental comparison of two quantum computing architectures — Linke et al., PNAS
