Cool a printed-circuit chip colder than deep space and it starts impersonating an atom — here is how that fake atom became quantum computing's loudest, fastest, most milestone-heavy front-runner, and what it still cannot do.
Part of the How to Build a Quantum Computer series.
Most qubit platforms spend enormous effort trying to catch a real atom — trapping it with lasers, levitating it in a vacuum, holding their breath so it does not drift away. Superconducting qubits do the opposite. They build the atom. You pattern a little circuit onto a wafer with the same lithography that prints classical processors, cool it to a whisker above absolute zero, and the circuit politely starts behaving like an atom: discrete energy levels, quantized transitions, the works. This is the route Google and IBM bet on, and it is the best-funded, most closely watched lane in the field. It is also, at its core, faintly absurd — one of humanity’s most advanced computers runs on a fake atom that is far bigger than a real one. Let us take the joke seriously, because the physics underneath it is genuinely beautiful.
Start with the question every qubit platform has to answer: what makes a good two-level system? A real atom is a good qubit because its electrons can only occupy a few discrete energy levels. Pick the lowest two, call them $\vert 0\rangle$ and $\vert 1\rangle$, and you have somewhere to store a bit of quantum information. The thing a superconducting qubit must reproduce is precisely that discreteness — and, less obviously, an uneven spacing between the levels.
Here is why evenness is fatal. The simplest superconducting circuit that oscillates is an LC resonator: an inductor $L$ and a capacitor $C$ trading energy back and forth, like a mass on a spring. Quantize it and its Hamiltonian (its energy written as a quantum operator) is the textbook harmonic oscillator,
\[H_{LC} = \frac{\hat Q^2}{2C} + \frac{\hat\Phi^2}{2L}, \qquad E_n = \hbar\omega\left(n+\tfrac{1}{2}\right), \quad \omega = \frac{1}{\sqrt{LC}} .\]The energy levels form a perfectly even ladder: every rung is separated by the same $\hbar\omega$. That sounds tidy, and it is exactly the problem. If you send in a microwave pulse tuned to drive $\vert 0\rangle\to\vert 1\rangle$, that same pulse is equally resonant with $\vert 1\rangle\to\vert 2\rangle$, $\vert 2\rangle\to\vert 3\rangle$, and on up the ladder — so trying to flip your bit just keeps driving the excitation up the ladder, straight out of the computational subspace. A harmonic oscillator, for all its elegance, makes a hopeless qubit. You need a ladder whose rungs are unequally spaced, so that a pulse resonant with $0\to1$ is detuned from $1\to2$ and leaves the higher levels alone.
The other half of the picture is what is actually doing the oscillating. It is tempting to imagine one lonely electron sloshing around, but the truth is stranger and more wonderful. Below its critical temperature the metal becomes a superconductor, and its conduction electrons pair up into Cooper pairs that condense into a single coherent quantum state — billions of charge carriers described by one macroscopic wavefunction, marching in lockstep. The “atom” you have built is enormous: a collective quantum object the size of a circuit, behaving as if it were a single particle with discrete energy levels. The 2025 Nobel Prize in Physics went to John Clarke, Michel Devoret, and John Martinis for the 1980s experiments that nailed down exactly this — that a macroscopic electrical circuit can tunnel and occupy quantized energy levels like a single quantum object
The component that turns a boring LC resonator into a usable artificial atom is the Josephson junction: two superconductors separated by an insulating barrier a couple of nanometers thick. Cooper pairs tunnel coherently across that gap, and Brian Josephson’s 1962 prediction
where $\phi$ is the quantum phase difference across the junction and $I_c$ is its critical current. Integrate the voltage relation and you find the junction stores energy as a function of phase:
\[U(\phi) = -E_J\cos\phi, \qquad E_J = \frac{\hbar I_c}{2e} = \frac{I_c\,\Phi_0}{2\pi},\]with $\Phi_0 = h/2e$ the flux quantum. That cosine is the whole trick. A linear inductor stores energy as $\tfrac12\Phi^2/L$ — a parabola, which gives the even ladder. The junction behaves like a nonlinear inductor, and the curvature of $-E_J\cos\phi$ changes as you climb it. Expand it about the bottom of the well:
\[-E_J\cos\phi \;=\; -E_J + \tfrac{1}{2}E_J\,\phi^2 - \tfrac{1}{24}E_J\,\phi^4 + \cdots\]The quadratic term is just a harmonic oscillator (even rungs); the quartic term is the anharmonicity that squeezes the rungs together as the energy rises — exactly the unevenness Figure 1 demanded.
Putting the junction in parallel with a capacitor gives the transmon, the design Jens Koch and colleagues introduced at Yale in 2007
where $\hat n$ counts Cooper pairs that have crossed the junction, $\hat\phi$ is the conjugate phase, $E_J$ is the Josephson energy above, and $E_C = e^2/2C_\Sigma$ is the charging energy set by the total capacitance. Treating the cosine to leading order, the qubit frequency and the anharmonicity come out as
\[\hbar\omega_{01} \approx \sqrt{8E_J E_C} - E_C, \qquad \alpha \equiv \hbar(\omega_{12}-\omega_{01}) \approx -E_C .\]To put numbers on it: a typical transmon runs with $E_J/h \sim 15$–$20$ GHz and $E_C/h \sim 0.2$–$0.3$ GHz, so $\omega_{01}/2\pi \approx 5$ GHz while the anharmonicity $\lvert\alpha\rvert/h$ is only about 200–300 MHz — a few percent of the transition frequency. That sliver of unevenness is the entire margin separating a working qubit from a leaky oscillator, which is why “small anharmonicity” is a phrase the field says nervously.
Why “transmon” rather than the earlier Cooper pair box, the very first solid-state qubit, demonstrated by Nakamura and colleagues in 1999
Two engineering facts about how the thing is made and cooled seed every later virtue and vice:
To operate the qubit you send microwave pulses — typically near 5 GHz, resonant with $\omega_{01}$ — down a control line. A calibrated pulse a few tens of nanoseconds long rotates $\vert 0\rangle\leftrightarrow\vert 1\rangle$ (a single-qubit gate). The anharmonicity sets a hard speed limit: a gate cannot run much faster than about $h/\lvert\alpha\rvert$ — a few nanoseconds — without driving population up into $\vert 2\rangle$, so pulses are deliberately shaped (the standard trick, DRAG, adds an out-of-phase component that cancels the leakage) to stay inside the qubit subspace.
Two-qubit gates split the field into two camps, and the split explains why an IBM chip and a Google chip look so different despite using the same transmon. Google and many others make their qubits frequency-tunable: they replace the single junction with a two-junction SQUID loop (a superconducting quantum interference device) whose effective $E_J$ — and hence $\omega_{01}$ — is set by a magnetic flux threaded through it, then entangle a pair by briefly tuning the two into resonance (or through a tunable coupler) to run a fast CZ or iSWAP gate. IBM takes the other road — fixed-frequency transmons driven purely by microwaves, entangled through the cross-resonance effect: drive one qubit at its neighbor’s frequency and the pair picks up a conditional rotation, no tuning required. Tunability buys speed; fixed frequency buys immunity to flux noise. You pay for whichever knob you keep.
One more operation, and it is the one the build-it story usually forgets: you have to read the answer out. You do not measure the transmon directly — you couple it to a separate microwave readout resonator and work in the dispersive regime, where qubit and resonator are far detuned. There the qubit’s state pulls the resonator’s frequency one way for $\vert 0\rangle$ and the other for $\vert 1\rangle$ (a shift $\pm\chi$), so you bounce a probe tone off the resonator, read the phase it returns with, and learn the qubit state without ever landing a real photon on the qubit. This whole framework — artificial atoms exchanging microwave photons with a cavity — is circuit quantum electrodynamics (circuit QED), the architecture the field is built on. Readout is the slowest, noisiest step in the stack (hundreds of nanoseconds, around 99% fidelity), which is why it looms large in every error budget and why each chip carries Purcell filters (which keep the qubit from radiating away down its readout line) and near-quantum-limited amplifiers to pull the signal out faster and cleaner.
Superconducting qubits lead the field for three reasons that all trace back to that fabrication-and-control story: they are fast, manufacturable, and crowded.
Fast. Gate times are tens of nanoseconds. Google’s processors, for instance, run single-qubit gates in about 25 ns and two-qubit (CZ) gates in about 34 ns
Manufacturable. Because qubits are printed, the whole apparatus of semiconductor manufacturing — lithography, integration, on-chip wiring, foundry processes — is available. Scaling up means making a bigger chip, not building a bigger laser-and-vacuum cathedral. This is why the raw qubit-count records belong to this platform, and why IBM, Google, and a crowd of startups can iterate on hardware every year.
Crowded. Nearly every quantum-computing headline of the past decade debuted on superconducting hardware, which means a deep, fast-moving ecosystem of tools, talent, and money. Fidelities reflect the maturity: Google’s Willow processor reports single-qubit gate fidelity around 99.97% and a best-case two-qubit (CZ) fidelity around 99.88%, each benchmarked one gate at a time
Now settle the bill. Superconducting’s weaknesses are as real as its strengths, and they are not bugs to be patched so much as the flip side of the same design choices.
The cold, and the wiring wall. A dilution refrigerator is expensive and finicky, but the deeper problem is what scaling does inside it. Every qubit needs its own control and readout lines threaded from room temperature down to 10 mK, and every one of those lines carries heat and noise. IBM’s 1121-qubit Condor processor reportedly packs over a mile of high-density cryogenic wiring into a single fridge
Coherence is short. Two numbers matter here: $T_1$, how long the excited state survives before it relaxes (energy loss), and $T_2$, how long the qubit keeps its phase before it dephases — bounded by $T_2 \le 2T_1$, and usually $T_2$ is both the shorter and the harder-won. In typical transmons both sit around 0.1–0.4 ms
Snowflakes and nearest neighbors. The fabrication spread comes due here. Because every junction is slightly different, every qubit must be individually tuned, and some pairs land at colliding frequencies that have to be designed around. Connectivity is mostly nearest-neighbor on a 2D grid, so entangling a distant pair means relaying the information through a chain of SWAP gates — the opposite of the all-to-all connectivity ions enjoy. Tunable couplers help switch interactions on and off and suppress crosstalk, but crosstalk remains the platform’s perennial headache.
Put those together and the in-machine numbers are soberer than the lab records. On real multi-qubit processors the median two-qubit gate fidelity is around 99.5%: Rigetti’s 84-qubit Ankaa-3 reports a median two-qubit fidelity of 99.5%
And the gap that frames the whole endeavor: running large-scale algorithms — Shor-scale factoring, fault-tolerant chemistry — calls for logical error rates of roughly $10^{-10}$ or lower, and commonly $10^{-12}$–$10^{-15}$ for cracking real cryptographic keys, because total error grows with the astronomical gate count
The superconducting scoreboard is a parade of hard-won milestones, each of which means less than the headline and more than the cynics allow. The trick is to read each one precisely.
2019 — Sycamore and “quantum supremacy.” Google’s 53-qubit Sycamore sampled a random quantum circuit in about 200 seconds and claimed a classical supercomputer would need on the order of 10,000 years to match it
2024 — Willow goes below threshold. Google’s 105-qubit Willow chip ran surface codes at code distances 3, 5, and 7 and showed the logical error rate dropping by a factor $\Lambda \approx 2.14$ each time the code grew
2025 — “Quantum Echoes” and verifiable advantage. In October 2025 Google reported, again on Willow, a measurement of an out-of-time-order correlator (an “echo” probing how quantum information scrambles) running roughly 13,000 times faster than the best known classical algorithm
The qubit-count flex. IBM’s Condor reached 1121 physical qubits in 2023
The players, and their bets. Here the milestones resolve into a single argument: different error-correction philosophies, one shared transmon. IBM (Condor / Heron / Nighthawk) is betting on quantum LDPC codes — its “gross code” packs 12 logical qubits into 144 data qubits, far cheaper than the surface code’s overhead
Superconducting qubits are a fast but high-maintenance front-runner. They win on speed, manufacturability, and ecosystem; they pay for it in cryogenics, short coherence, fabrication spread, and one weakness that does not show up on any gate-fidelity spec: they are terrible at networking.
That last point deserves a sentence, because this series keeps one eye on whether a qubit can network as well as compute. A superconducting qubit speaks in ~5 GHz microwave photons, and microwave photons cannot travel down an optical fiber the way telecom-band light does — glass is transparent only near 200 THz, some four orders of magnitude higher in frequency, and a lone 5 GHz photon would in any case drown in the thermal noise of any room-temperature link. Linking two superconducting processors over fiber requires microwave-to-optical quantum transduction, a translation step whose efficiency — especially in the low-added-noise regime a true quantum link demands — is still only a few percent in practice, a few tens of percent at the most optimistic edge
| Superconducting | Trapped ion | Neutral atom | Photonic | |
|---|---|---|---|---|
| Qubit | printed circuit (transmon) | a real atom (ion) | a real atom | a photon |
| Gate speed | very fast (~10–70 ns) | slow (~µs) | slow (~µs) | n/a (measurement) |
| Coherence | short (~0.1–1.7 ms) | very long (s–min) | long (s) | loss-limited |
| Connectivity | mostly nearest-neighbor | all-to-all | reconfigurable | hard (no interaction) |
| Operating temp | ~10 mK (dilution fridge) | room-temp vacuum | room-temp vacuum | room temperature |
| As a network node | needs transduction | natural (emits photons) | promising | it is the photon |
No platform wins every row, which is the whole reason five of them are still in the race. If you want a cloud-accessible, general-purpose, engineering-leading quantum computer today, superconducting is the most realistic pick — a math genius who happens to live in a refrigerator and cannot use a phone. But if your wish list includes long memory and a photon you can fire at a distant node, the next routes in this series are about to look very attractive. Up next: what happens when you stop faking the atom and trap a real one — trapped ions.
More in this series — How to Build a Quantum Computer: Superconducting · Trapped ion · Neutral atom · Photonic · Other platforms