Quantum hardware: how qubits are actually built

The previous lesson treated qubits abstractly: two-level systems you can rotate, entangle, and measure. Building one means finding a physical system with two usable quantum states and enough isolation that superpositions survive long enough to compute.

Every candidate platform is judged on the same scorecard:

• Coherence time: how long a superposition survives before decoherence scrambles it.
• Gate fidelity: what fraction of operations do what they should. Two-qubit gates are always the weak point.
• Gate speed: how many operations fit inside one coherence time. What matters is the ratio, operations per coherence window, not either number alone.
• Connectivity: which qubit pairs can interact directly.
• Scalability: can you manufacture and control thousands, then millions?

No platform wins every column, and that is the story of this lesson: quantum hardware is an engineering trade-space, not a race with one obvious leader.

Decoherence is not one phenomenon but a family. The two canonical decay channels have standard names:

• T1 (energy relaxation): the excited state $|1\rangle$ decays toward $|0\rangle$, losing energy to the environment.
• T2 (dephasing): the relative phase between $|0\rangle$ and $|1\rangle$ drifts randomly, blurring exactly the information that interference needs. T2 is never better than twice T1 and is usually the tighter constraint.

The sources are mundane: thermal photons, magnetic field fluctuations, vibrations, material defects at chip surfaces, and imperfect control pulses that over- or under-rotate the state. This is why quantum hardware lives in extreme environments, dilution refrigerators near 10 millikelvin, ultra-high vacuum chambers, laser-stabilized optical tables.

A useful gotcha: adding more qubits usually worsens each qubit. More control lines mean more heat and crosstalk, so scaling and quality pull against each other. Progress means improving both at once.

The most industrialized platform builds qubits from superconducting circuits: loops of niobium or aluminum on a chip, cooled until electrical resistance vanishes. A Josephson junction (a nanometer-thin insulating barrier between superconductors) gives the circuit unevenly spaced energy levels, so the bottom two can serve as $|0\rangle$ and $|1\rangle$. The dominant design is the transmon.

Strengths:

• Fast gates: tens of nanoseconds, the fastest of any major platform.
• Chip fabrication: leverages semiconductor-style lithography; hundreds of qubits per chip is routine.
• Microwave control electronics are mature.

Costs:

• Short coherence: tens to hundreds of microseconds, so the speed advantage is spent racing decoherence.
• Fixed, sparse connectivity: qubits talk only to on-chip neighbors, typically a 2D grid.
• Every qubit is slightly different (fabrication variance), demanding per-qubit calibration.
• Dilution refrigerators and wiring per qubit make scaling to millions a serious open problem.

A trapped-ion qubit is a single charged atom (ytterbium and barium are popular) held in a vacuum by oscillating electric fields, with two internal electronic states as $|0\rangle$ and $|1\rangle$. Lasers drive the gates; a shared vibrational mode of the ion chain acts as a quantum bus, letting any ion entangle with any other.

Strengths:

• Best-in-class fidelity: two-qubit gates at 99.9%+ are demonstrated across all qubit pairs on leading systems.
• Long coherence: seconds to minutes; atoms are perfect and identical, no fabrication variance.
• All-to-all connectivity within a chain, which shortens circuits dramatically.

Costs:

• Slow gates: microseconds to milliseconds, thousands of times slower than superconducting gates.
• Chains beyond ~50 ions become unwieldy; scaling requires shuttling ions between zones or photonically linking modules, both of which add overhead.

The capsule summary: ions trade raw speed for quality and connectivity.

Neutral atoms are the platform whose scale grew fastest in recent years. Individual atoms (often rubidium) are held in optical tweezers, tightly focused laser beams arranged in reconfigurable 2D grids; arrays beyond a thousand atoms have been demonstrated. Exciting an atom to a giant Rydberg state makes neighbors interact, producing entangling gates. The draws: identical atoms, room-temperature vacuum systems (no dilution refrigerator), and the ability to physically move atoms mid-computation to rewire connectivity. Gate fidelities trail the best ions and superconductors but are climbing quickly.

Photonic platforms encode qubits in single particles of light. Photons barely decohere and travel at room temperature through fiber, but they do not naturally interact, so two-qubit gates are probabilistic, made workable through measurement tricks and massive multiplexing. Silicon-photonics manufacturing is the scaling bet. Photons are also the only practical carrier for networking quantum processors, whatever the processors are made of.

Read the columns as strategies: superconductors bet on speed and fabrication; ions on quality and connectivity; atoms on identical qubits at scale; photons on manufacturing and networking. Fidelity numbers shift year to year; the trade-off structure is far more durable.

Raw specs mislead; ratios decide. Compare two stylized machines:

Machine S (superconducting-like)
  gate time      : 50 ns
  coherence (T2) : 100 µs
  ops per window : 100,000 ns / 50 ns  = ~2,000 gates

Machine I (ion-like)
  gate time      : 100 µs
  coherence (T2) : 10 s
  ops per window : 10,000,000 µs / 100 µs = ~100,000 gates

The "slow" ion machine fits 50x more gates into its coherence window; per wall-clock hour, though, the superconducting machine executes far more circuits, useful when algorithms need millions of repeated runs (sampling, variational loops).

A second ratio matters just as much: circuit depth after routing. On a fixed 2D grid, interacting two distant qubits requires chains of SWAP gates, each one an extra error opportunity. All-to-all machines skip that tax entirely. The right question is never "which platform is best?" but "per useful circuit, how much error accumulates?", a preview of the error-correction arithmetic in the next lesson.

Where does this leave the field? Every leading platform now demonstrates qubits good enough to run small circuits, and none is close to the millions of excellent qubits that headline applications like breaking RSA would demand. The binding constraint has shifted from "can we make a qubit?" to two harder questions:

1. Can gate errors get low enough? Around the ~99.9% fidelity mark, something qualitative changes: error correction begins to help rather than hurt, because the redundancy it adds removes more error than it introduces. That tipping point is called the threshold, and crossing it cleanly is the defining experiment of the current era.
2. Can control scale? Lasers, microwave lines, and readout electronics per qubit must shrink in cost and complexity by orders of magnitude.

Every platform in this lesson is racing toward the same destination: enough good physical qubits, wired into the error-correcting codes that turn them into a few logical qubits that are effectively perfect. How that construction works is the next lesson.