The DiVincenzo Criteria and What They Screen Out

In 1996, David DiVincenzo, then at IBM Research, wrote down what a machine would need in order to count as a quantum computer. He formalized the list in a 2000 paper, “The Physical Implementation of Quantum Computation”: five requirements for computation, plus two more for anything that has to send quantum information somewhere else. A quarter-century later, those seven items are still the first filter applied to a new hardware proposal, and they’re still the quickest way to tell whether an announcement describes a computer or a component of one.

The list does one job well and refuses to do another. It screens architectures, telling you whether a physical system has the ingredients to become a general-purpose quantum computer at all. It does not rank machines, and it never claimed to. Two platforms can satisfy all five computing criteria and be a decade apart in usefulness, because the criteria say nothing about how many qubits, how good the gates are beyond “good enough to be called gates,” or what it costs to run the thing.

This article works through each criterion, explains what it demands of real hardware, and shows how superconducting circuits, trapped ions, and photonics each meet it or fail to. It closes with the part that gets skipped most often: the engineering problems the list was never designed to catch.

Why DiVincenzo Wrote the List

The idea of a quantum computer came out of physics rather than computing. Yuri Manin in 1980 and Richard Feynman in the early 1980s both observed that classical machines struggle to simulate quantum systems, and that a machine built from quantum parts might not. By the mid-1990s, laboratories were demonstrating quantum logic operations on two or three particles. There was no agreement on what separated those demonstrations from a machine.

DiVincenzo wrote the list partly as a corrective. He had watched genuinely impressive experiments get described as steps toward a full-scale computer when they addressed only one narrow piece of the problem. In a later retrospective he said the criteria were meant to remind experimenters “not to pretend that your impressive little experiment is actually a big step towards the physical realization of a quantum computer.”

The specific target was liquid-state nuclear magnetic resonance. NMR quantum computing used nuclear spins in molecules dissolved in a liquid, addressed with radio-frequency pulses, and in the late 1990s it was the most capable platform in existence. In 2001 a group at IBM and Stanford, led by Lieven Vandersypen, ran Shor’s algorithm on seven nuclear spins and factored the number 15. Nothing else came close at the time.

It also went nowhere, and the criteria explain why. NMR worked on an ensemble of roughly 10^18 molecules at room temperature, never on a single pure quantum state. The signal from a pseudopure state falls off exponentially as spins are added, so every qubit past the seventh costs exponentially more sample or exponentially more averaging. Two of DiVincenzo’s words were aimed directly at this. “Scalable” ruled out systems whose overhead explodes with size. “Fiducial,” meaning a known reference state, ruled out systems that compute on top of thermal randomness. He has said as much: those were warning phrases, and the NMR community was the intended audience.

That episode is the reason to keep using the list. Every generation of quantum hardware produces a result that is real, publishable, and easy to mistake for progress toward a machine. The criteria are how you check.

The Five Computing Criteria

1. A Scalable Physical System With Well-Characterized Qubits

A qubit is a two-level quantum system whose state can be 0, 1, or a superposition of both, meaning a weighted combination of the two that only resolves into one answer when measured. The first criterion asks for a physical system that can host many of them, where each one’s behavior is known in detail.

Two words do the work. Scalable means adding qubits without a step change in noise, control complexity, or manufacturing difficulty. Well-characterized means you know each qubit’s energy levels, its coupling to its neighbors, and its error behavior, and that qubits of the same type behave the same way. A qubit that occasionally leaks into a third energy level is not a qubit, in the same way that a bit which sometimes reads as 2 is not a bit.

Superconducting circuits are strong on scalability and weaker on uniformity. They’re built with the same lithographic processes as conventional chips, and IBM’s 127-qubit Eagle processor, announced by the company in November 2021, showed that qubit counts can grow through packaging and wiring innovation rather than new physics. The cost is that every transmon is a fabricated object with its own resonance frequency, and fabrication spread means qubits on the same die are not identical. Calibration absorbs the difference, and calibration time grows with the device.

Trapped ions invert the tradeoff. Every ytterbium ion in a trap is identical to every other one, because atoms of the same isotope have the same energy levels by nature. That gives near-perfect characterization for free. The difficulty is that entangling gates in a linear trap use the shared vibrational motion of the whole ion chain, and as the chain grows, those vibrational modes crowd together in frequency until they can no longer be addressed cleanly. The trapped-ion review by Bruzewicz and colleagues in 2019 puts the practical ceiling for a single chain in the low hundreds of ions. The route past it is modular, with small traps connected either by physically shuttling ions between zones or by photonic links, and that architecture is still being built.

Photonic qubits are encoded in properties of individual particles of light, usually polarization or which of two paths the photon takes. Photons of the same frequency are identical, so characterization is not the issue. Generation is. Producing a single photon on demand, routing it through the right components, and detecting it at the end each carry a loss probability, and those probabilities compound. PsiQuantum’s argument for photonics is manufacturing: optical circuits can be fabricated in existing semiconductor foundries at volume, which is the only known way to reach a million components. Whether that argument holds is a question about yield and loss budgets, not about physics.

Neutral atoms deserve a mention here because they have the ion advantage of identical qubits with a different scaling path. Atoms held in optical tweezer arrays reached 256 qubits in a single system in work from Ebadi and colleagues published in Nature in 2021, arranged and rearranged by moving the tweezers.

2. Initialization to a Fiducial State

The second criterion requires the ability to reset the register to a known starting state, normally all qubits in 0. A fiducial state is just that: a reference state you can prepare reliably and trust. Classical machines clear memory at boot for the same reason. Computation on unknown initial data produces unknown results, and in the quantum case the unknown initial data becomes entangled with everything the algorithm does afterward.

Superconducting qubits get initialization close to free. At the millikelvin temperatures inside a dilution refrigerator, the excited state is thermally unpopulated and a transmon relaxes into 0 on its own within a few multiples of its relaxation time. Waiting works, though waiting is slow, so most systems now use active reset: measure the qubit, and apply a bit flip if it reads 1. The same coupling to a cold environment that resets the qubit for free also destroys its coherence.

Trapped ions use optical pumping. A laser tuned so that it excites the ion only from one of the two qubit states drives the population repeatedly until it falls into the other state and stops interacting. Combined with laser cooling of the ion’s motion, this prepares the register in microseconds, with fidelities above 99 percent in the figures collected by Bruzewicz and colleagues.

Photonics reframes the question. Photons are created rather than reset, so initialization means producing a photon in a known state at a known time. Single-photon sources based on parametric processes are probabilistic, and a source that fires half the time leaves you with a missing qubit half the time. Deterministic emitters such as quantum dots have improved considerably, and this is now an efficiency problem rather than a conceptual one.

Initialization is the criterion that fails quietly. Error correction consumes a continuous supply of freshly reset qubits, so a platform where reset is slow relative to gate operations pays for it in every correction cycle, not just at the start of the program.

3. Coherence Times Far Longer Than Gate Times

Decoherence is the process by which a qubit’s superposition leaks into its environment and degrades into an ordinary classical mixture of 0 and 1. Two timescales describe it. T1 is the energy relaxation time, how long an excited qubit stays excited. T2 is the phase coherence time, how long the relative phase between the two states survives, and it’s usually the shorter and more demanding of the two.

The criterion does not ask for long coherence in absolute terms. It asks for coherence that’s long compared to the time a single gate operation takes, because that ratio sets how many operations fit inside one quantum state’s lifetime. DiVincenzo’s own formulation asks only that coherence be long compared to operation times; a commonly cited heuristic puts the target ratio on the order of 10,000 to 100,000, and that number comes from later fault-tolerance threshold discussions rather than from the 2000 paper. A qubit with a 1-microsecond coherence time and a 0.5-microsecond gate can execute two operations, which is a physics experiment rather than a computer.

Superconducting qubits are fast and leaky. In the survey by Kjaergaard and colleagues in 2020, gate times sit in the range of tens of nanoseconds for single-qubit operations and around 100 nanoseconds for two-qubit gates, which is the fastest of any leading platform. Coherence has improved by several orders of magnitude since the early 2000s, and Place and colleagues in 2021, using tantalum rather than niobium films, reported transmon T1 near 0.3 milliseconds. Typical production devices run well below that. The resulting ratio of coherence to gate time is in the thousands, which satisfies the criterion with less headroom than the number suggests once accumulated gate error is counted.

Trapped ions run the same tradeoff in reverse. Their qubits are stored in hyperfine states of the atom, the same kind of transition used in atomic clocks, and they’re almost indifferent to their surroundings. Wang and colleagues in 2021 measured a single ytterbium ion with a coherence time estimated above one hour. Gate times, though, are tens to hundreds of microseconds, several hundred times slower than superconducting gates. Ions win on the ratio and lose on wall-clock throughput, which is a real operational difference when a circuit has millions of operations in it.

Photons barely have a coherence time in the usual sense. Polarization does drift in ordinary fiber, because birefringence and environmental disturbance rotate the state, and that drift is managed with polarization-maintaining fiber or active stabilization rather than with better isolation. The dominant practical failure is simpler. Absorption and scattering destroy the photon outright. For photonics, loss plays the role that decoherence plays elsewhere, and the same criterion translates into a question about how many optical components a photon can pass through before it’s likely to be gone.

This is the criterion where a vendor’s numbers are most worth reading together rather than separately. Coherence time alone tells you nothing. Coherence time divided by two-qubit gate time tells you the depth of circuit the hardware can attempt before error correction has to carry the load.

4. A Universal Set of Quantum Gates

A universal gate set is a finite collection of operations from which any quantum computation can be built. Classically, the NAND gate alone is universal. In the quantum case, arbitrary single-qubit rotations plus one entangling two-qubit gate will do, and the choice of entangling gate is fairly free: CNOT, controlled-Z, and iSWAP all work.

The criterion asks only whether the set exists, not whether the gates are good. That’s the most common misreading. A platform with a universal gate set at 90 percent fidelity satisfies criterion 4 and cannot compute anything.

Superconducting systems implement single-qubit rotations with shaped microwave pulses and two-qubit gates through tunable couplers or fixed capacitive coupling. Google’s Sycamore processor used a tunable coupler for an iSWAP-family gate and reported a median two-qubit gate error of 0.62 percent in isolated operation, roughly 99.4 percent fidelity, in the 2019 sampling experiment in Nature. IBM’s devices reach comparable two-qubit fidelities using the echoed cross-resonance interaction introduced by Sheldon and colleagues in Physical Review A in 2016. Both approaches are universal, and both are constrained by nearest-neighbor connectivity: entangling two distant qubits requires a chain of swap operations, and each swap costs fidelity.

Trapped ions use resonant laser pulses for single-qubit rotations and the Mølmer-Sørensen interaction for entanglement, which couples two ions through their shared motion with a bichromatic laser field. The architectural advantage is all-to-all connectivity, since any two ions in a chain share the same motional modes and can be entangled directly. For algorithms with long-range structure, that removes a large amount of swap overhead, which is why ion systems have historically posted high scores on depth-sensitive benchmarks with modest qubit counts.

Photonics is the awkward case. Photons don’t interact with each other, so a deterministic two-photon gate needs either a strong optical nonlinearity or a matter intermediary, and neither is practical at scale. The Knill-Laflamme-Milburn scheme showed in 2001 that linear optics plus measurement and feedforward is universal, with the gates succeeding only probabilistically and announcing their success through a detector click. Current photonic architectures build on the successor idea, measurement-based quantum computing, in which a large entangled resource called a cluster state is prepared in advance and the computation proceeds by measuring photons one at a time in bases chosen as you go. Universality holds. The overhead is enormous, and reducing it is the whole photonic engineering program.

Silicon spin qubits belong in this section too, since three teams reported two-qubit gate fidelity in silicon above 99 percent in early 2022, Xue and colleagues and Noiri and colleagues in Nature and Mills and colleagues in Science Advances, which is the point at which a modality stops being a curiosity and starts being a candidate.

5. Qubit-Specific Measurement

The last computing criterion requires reading out an individual qubit, accurately, without scrambling the others. Ideally the measurement is quantum non-demolition, meaning that a qubit measured as 1 is left in the state 1 rather than kicked somewhere unpredictable, so the result can be trusted and the qubit reused.

This is not only an output requirement. Error correction works by repeatedly measuring auxiliary qubits mid-circuit to detect errors on the data qubits without measuring the data itself. A platform that can only read out the whole register at the end can run algorithms but cannot correct errors, which caps it at whatever depth its raw fidelity allows.

Superconducting readout is dispersive. Each qubit is coupled to a microwave resonator whose frequency shifts slightly depending on the qubit state, and probing the resonator reveals the state without directly disturbing the qubit. Because each resonator has its own frequency, many qubits can be read simultaneously on one line through frequency multiplexing. Readout takes on the order of tens to hundreds of nanoseconds and is now the fastest of the leading platforms, though readout error typically remains larger than gate error.

Ion readout uses fluorescence. A detection laser is chosen so that one qubit state scatters photons and the other stays dark, and a camera or photomultiplier array records which ions light up. Spatial resolution provides the qubit specificity, since ions in a chain sit micrometers apart. Fidelities above 99 percent are routine in the figures Bruzewicz and colleagues collect. The cost is time, because collecting enough photons to be confident takes hundreds of microseconds to milliseconds, which is slow for a mid-circuit measurement inside a feedback loop. Some systems use two ion species, one for data and one for readout, so that the detection laser doesn’t touch the data qubits at all.

Photonic measurement is the easy part, and in measurement-based architectures it’s the computation. Marsili and colleagues reported superconducting nanowire single-photon detectors with 93 percent system detection efficiency in Nature Photonics in 2013, and directing a photon through a polarizing beam splitter into two detectors performs a basis measurement outright. The complication isn’t isolation, since photons in different modes don’t share a detector, but dark counts and the loss that precedes detection.

The Two Communication Criteria

The first five criteria describe a machine working alone. The last two describe what it takes to send quantum information out of one and into another. A stationary qubit is one that stays put, such as an ion in a trap or a circuit on a chip. A flying qubit is one that travels, which in practice always means a photon.

These criteria used to look optional. They don’t anymore, because most credible paths to a large machine now involve connecting modules rather than building one enormous device.

6. Interconversion of Stationary and Flying Qubits

This criterion asks for a transducer: a way to write the state of a stationary qubit onto a photon, and to read a photon’s state back into a stationary qubit.

Trapped ions do this natively. An excited ion decays and emits a photon whose polarization is entangled with the ion’s internal state, and interfering two such photons from two separate traps entangles the two ions without either ever meeting the other. The Innsbruck group, in work by Krutyanskiy and colleagues published in 2023, reported ion-ion entanglement across a 230-meter fiber link between separate buildings. Nitrogen-vacancy centers in diamond work the same way, and the Delft group used exactly this mechanism for the 2015 loophole-free Bell test over 1.3 kilometers.

Superconducting circuits do not. Their qubits live at microwave frequencies of a few gigahertz, and microwave photons are absorbed within meters and cannot enter an optical fiber. Connecting two superconducting modules over any real distance requires converting a microwave excitation into a telecom-band optical photon and back, which is an open research problem. Prototype microwave-to-optical transducers exist, and their end-to-end efficiency is low. This is the single largest architectural gap in the most heavily funded modality, and it’s worth asking any superconducting vendor with a modular roadmap how they intend to close it. An answer that amounts to sending classical measurement results between modules is not a quantum interconnect.

Photonics satisfies the criterion by construction, since its qubits are already flying. The mirror-image problem appears instead: storing a photonic qubit requires a quantum memory, whether an optical delay line, an atomic ensemble, or a solid-state emitter, and memory is what photonic repeater schemes are short of.

7. Faithful Transmission of Flying Qubits

The final criterion asks whether a flying qubit can actually get from one place to another with its state intact. Quantum signals cannot be amplified the way classical ones are, because the no-cloning theorem forbids making a copy of an unknown quantum state. Every repeater in a classical network is doing something a quantum network is not allowed to do.

In fiber, the limit is attenuation. Telecom fiber at 1550 nanometers loses about 0.2 decibels per kilometer, so roughly 1 percent of photons survive 100 kilometers. Point-to-point quantum links over metropolitan distances work today and are the basis of deployed key distribution systems. Beyond that, you need quantum repeaters, which store entanglement at intermediate nodes and extend it by entanglement swapping. A repeater node is itself a small quantum computer with a memory, which means criteria 1 through 5 come back around at every hop.

Free space avoids attenuation but adds beam divergence and atmospheric turbulence. The Micius satellite, reported by Jian-Wei Pan’s group in Science in 2017, distributed entangled photon pairs to ground stations separated by 1,200 kilometers. Most pairs were lost. The ones that arrived were faithful, which is the distinction the criterion draws: loss is an erasure you can detect and discard, while an unpredictable rotation of the photon’s state is corruption you cannot.

Ions face a wavelength mismatch on this criterion. Ytterbium ions emit in the ultraviolet at 369 nanometers, which fiber absorbs almost immediately, so networking them over distance requires quantum frequency conversion into the telecom band. That works, and it costs efficiency at every conversion.

What the Criteria Do Not Measure

By the 2010s, several platforms could claim all five computing criteria in a single device. DiVincenzo’s own assessment, offered in 2018, was that they should be read as promise criteria: satisfying them means you have components that could in principle be assembled into a system, and assembling them is a separate and much larger job.

That job is where most engineering effort now goes, and none of it appears on the list.

Error correction overhead. The criteria distinguish nothing between a physical qubit, meaning one piece of hardware, and a logical qubit, meaning one error-corrected qubit encoded across many physical ones. The exchange rate between them is set by gate fidelity, and it’s brutal. The most recent public resource estimate for factoring RSA-2048, published by Craig Gidney in 2025, puts the requirement below a million noisy physical qubits, revised down from roughly 20 million in a 2019 estimate by Gidney and Martin Ekerå. Both numbers describe hardware that satisfies all five criteria. The six-year gap between them is a fact about algorithms and codes, not about criteria.

Control electronics and wiring. A superconducting qubit needs control and readout lines running from room temperature into a dilution refrigerator with finite cooling power and finite physical space. Multiplying qubits multiplies lines, and past a certain count the refrigerator becomes the constraint. Cryogenic control chips and frequency multiplexing are attempts to fix this, and the criteria have nothing to say about either.

Crosstalk and calibration at scale. Two qubits that behave perfectly in isolation may not when 200 neighbors are being driven. Frequency collisions between fabricated qubits, stray coupling, and drift all grow with device size, and recalibration time can grow faster than the qubit count.

Fabrication yield. A chip design that works is not a chip design that can be manufactured at acceptable yield, and the criteria treat the qubit as given.

Benchmarks that combine the criteria. Because no single criterion captures machine capability, the field built composite metrics. Quantum volume, introduced by IBM, measures the largest square random circuit a device can run successfully, which folds together qubit count, connectivity, and gate fidelity into one number. Honeywell’s ten-qubit H1 system reached a quantum volume of 128 on the company’s own reported measurement in September 2020, beating superconducting devices with far more qubits, and the underlying architecture was described by Pino and colleagues in Nature in 2021. That result is the clearest illustration available that qubit count on its own predicts nothing. Later benchmarks such as algorithmic qubit counts and application-level suites push further in the same direction.

We say this to learners plainly, and it’s a judgment about how the market communicates rather than about any particular company: the qubit count is the least informative number a quantum hardware vendor publishes, and it’s the one that reaches the headline.

Reading a Hardware Announcement

The practical use of DiVincenzo’s list is as a set of questions to hold against a claim. When a new result is announced, work down the seven and see which ones the announcement is silent about, because the silences are the information.

1. Scaling path. How many qubits does this design support, and what physically stops it there? A vendor who can name their own ceiling and describe the architecture that goes past it is telling you something. A vendor who says scaling is only a matter of funding is not.
2. Initialization and reset. What’s the preparation fidelity, and how long does a reset take relative to a gate? Slow reset becomes an error-correction tax.
3. Coherence against gate time. Ask for both numbers and divide. The ratio, not the coherence time, sets achievable circuit depth.
4. Two-qubit gate fidelity and connectivity. Single-qubit fidelity is almost always the better number and almost never the limiting one. Ask which qubit pairs the quoted fidelity applies to, and what connectivity the device offers, because all-to-all connectivity and nearest-neighbor connectivity are not comparable at equal fidelity.
5. Measurement. Is mid-circuit measurement supported, and at what fidelity and duration? Without it, error correction is off the table regardless of everything else.
6. Interconnect. If the roadmap is modular, what carries quantum information between modules, and does it exist today?
7. Transmission. For anything involving distance, what wavelength, what loss budget, and what stands in for a repeater?

Two habits go with the list. Separate demonstrated from announced, since a device that has been fabricated and characterized is a different object from a device on a slide. And separate physical from logical, because 1,000 physical qubits and 1,000 logical qubits are separated by something like three orders of magnitude of hardware.

The criteria were written to stop a specific over-claim, and the over-claim they were written against is gone. NMR quantum computing stalled at seven spins, exactly where DiVincenzo’s phrasing predicted. What replaced it is a field where every serious platform satisfies the list on a small device, and where the distance between satisfying it and building something useful is the entire remaining problem. That’s the reading to take away. The criteria tell you whether an approach is worth watching. Nothing on the list tells you when.

Where to Take This Next

Reading hardware claims well takes more than the checklist. It takes enough familiarity with each modality to know which numbers a given architecture finds easy to report and which ones it tends to leave out, and enough grounding in error correction to convert a physical qubit count into something meaningful.

That’s the ground our quantum technology programs cover, working through the modalities one at a time, with the benchmarks and the resource estimates alongside them. You can see the current program list and enrollment details at quantumacademy.com/.

For the migration and cryptographic side of quantum readiness, the methodology at pqcframework.org is the companion resource, and PostQuantum.com carries deeper technical analysis of individual hardware results as they’re published.