Marin Ivezic
A diligence deck crosses the desk with three hardware numbers on it. One company reports more than 5,000 qubits. Another reports 105. A third reports an array of 6,100 atoms held in place by lasers. Ranked by that column, the order is obvious, and it’s also close to meaningless. The three numbers describe different physical objects, doing different kinds of work, under three different definitions of the word qubit.
A qubit is the quantum version of a bit. A classical bit is either 0 or 1; a qubit holds a weighted combination of both until it’s measured, and qubits can be correlated with each other, or entangled, so that a single operation acts on the whole set. That much is common to every machine in the field. Almost nothing else is.
We teach this material to people who have to make decisions about it rather than build it, and the framework below is the one we use. A modality is two things bolted together: a computational model, meaning how the machine is instructed, and a physical carrier, meaning what the qubit is physically made of. Confuse the two and you’ll compare a photon to an annealer and conclude that one of them is behind. Separate them and the field resolves into a small number of genuinely different bets.
Three Questions That Sort the Field
Every serious quantum computing company can be placed with three questions, asked in this order.
What computational model does the machine implement? This determines what problems it can address at all. A machine that cannot run arbitrary quantum circuits cannot run Shor’s algorithm, no matter how many qubits it has.
What is the qubit physically made of? This determines gate speed, error rates, connectivity, operating temperature, and the manufacturing path. It’s where most of the capital is going.
What is the error correction strategy, and what does it cost? This is the question that separates a demonstration from a business. Every current machine is noisy. The ratio of physical qubits consumed per usable logical qubit is the single number that most strongly determines when a company reaches a useful product, and it varies across platforms by a factor of several hundred.
The first question has four answers. The second has about seven that matter commercially. The third is where the field is currently being decided.
The Computational Models
The gate model is the mainstream. An algorithm is a sequence of discrete operations, or gates, applied to qubits, in the same way a classical program is a sequence of instructions. Peter Shor’s factoring algorithm and Lov Grover’s search algorithm are both written for this model, and so is essentially every algorithm with a proven speedup. The gate model is universal, meaning that with enough good qubits it can run anything a quantum computer can run. Superconducting circuits, trapped ions, neutral atoms, silicon spins, and most photonic designs all target it.
Measurement-based quantum computing, or MBQC, gets to the same place by a different route. Instead of applying gates one after another, the machine first prepares a large entangled state across many qubits, then computes by measuring them one at a time, choosing each measurement based on the results of the previous ones. Robert Raussendorf and Hans Briegel showed in 2001 that this is exactly as powerful as the gate model. It matters commercially because photons are hard to hold still and easy to measure, which makes MBQC and its variants the natural fit for light-based hardware. When you read that a photonic company is pursuing a fusion-based architecture, that’s a member of this family.
Annealing and adiabatic evolution work by physics rather than by instruction. The machine encodes an optimization problem as an energy function, prepares a system in an easy starting configuration, and then slowly deforms it toward the hard problem, letting the system settle into a low-energy state that represents a good answer. D-Wave has been shipping machines of this kind for over a decade, currently with more than 5,000 qubits. Those qubits are not comparable to gate-model qubits and were never meant to be. An annealer cannot run Shor’s algorithm. It’s a specialized optimizer, and it should be evaluated against classical optimization software rather than against IBM.
Topological computing is often listed as a fourth model, and that placement causes confusion. It’s better understood as an error correction strategy implemented in hardware. The proposal, originating with Alexei Kitaev in 1997, is to store information in exotic collective states of matter whose properties depend on global topology rather than local detail, so that local noise simply cannot corrupt the stored value. The computation on top of it is still gate-model. Topology is the protection, not the programming.
Error correction deserves its own definition, because it’s where the money goes. Physical qubits are the ones you can point at in the hardware. A logical qubit is a reliable qubit assembled out of many noisy physical ones, with the redundancy used to detect and fix errors continuously. The surface code is the best-understood scheme and also the most expensive, requiring on the order of 1,000 physical qubits per logical qubit at realistic error rates. Below threshold means the machine has crossed the point where adding more physical qubits makes the logical qubit better rather than worse; below threshold, error correction works, and above it, error correction adds noise faster than it removes it. Google’s Willow processor, 105 physical qubits, demonstrated this crossing in December 2024, with the logical error rate falling each time the code was enlarged. That result moved below-threshold error correction from a theoretical requirement to a demonstrated one.
The Physical Carriers
Superconducting circuits. Tiny electrical circuits built on chips, cooled to around 10 millikelvin, where certain metals lose all electrical resistance and the circuit behaves like a single artificial atom. Gates take nanoseconds, the fastest of any platform. IBM and Google lead, with IBM shipping Heron processors at 133 and 156 qubits and publishing a roadmap toward a fault-tolerant machine before the end of the decade. The weakness is coherence time, meaning how long a qubit holds its state before ambient noise destroys it, which sits in the range of hundreds of microseconds. The platform also carries a heavy infrastructure burden: dilution refrigerators, helium-3 supply, and thousands of control wires running from room temperature down into the cold stage. That wiring is a real scaling constraint and one that gets discussed less than qubit counts do.
Trapped ions. Individual charged atoms held in place by electromagnetic fields and manipulated with lasers or microwaves. Ions are identical to each other by the laws of physics, which removes the fabrication variability that plagues solid-state platforms, and they hold coherence for seconds. Gate fidelities are the highest in the field. Critically, any ion in a chain can interact with any other, which is called all-to-all connectivity, and it allows the use of quantum low-density parity-check codes, or qLDPC codes, a family of error correction schemes that pack more logical qubits into fewer physical ones than the surface code can. Quantinuum reports logical qubits produced at a ratio close to two physical qubits per logical qubit, against roughly 1,000 to 1 for a surface code on nearest-neighbour hardware. If that ratio holds as systems grow, it reorders the timeline. The offsetting weakness is speed: ion gates run in microseconds to milliseconds, roughly a thousand times slower than superconducting gates. IonQ paid roughly a billion dollars for Oxford Ionics in 2025 largely to replace laser control with on-chip electronic control, which addresses the scaling problem rather than the speed one.
Neutral atoms. Uncharged atoms, usually rubidium or strontium, held in arrays of focused laser beams called optical tweezers. Two-qubit gates work by exciting atoms into Rydberg states, in which the outer electron orbits so far from the nucleus that neighbouring atoms interact strongly. Two properties make this platform commercially interesting. Arrays are large and cheap to enlarge, with Caltech reporting an array of 6,100 atoms in 2025, and the tweezers can physically move atoms mid-computation, so connectivity is reconfigurable rather than fixed in the wiring. QuEra and its Harvard and MIT collaborators have reported dozens of logical qubits running on a few hundred atoms using high-rate codes. Pasqal has installed systems at European supercomputing centres, and the hardware runs at room temperature in a standard rack. The open questions are two-qubit gate fidelity, still short of trapped ions, and atom loss during long computations.
Photons. Qubits encoded in particles of light, manipulated by beam splitters and phase shifters etched into silicon photonic chips. The main body of the machine runs at room temperature, and only the single-photon detectors need cooling. PsiQuantum fabricates its chips at GlobalFoundries on standard 300 mm semiconductor lines and published results on its Omega chipset in Nature in 2025. Xanadu and Quandela pursue related routes. The structural difficulty is that entangling operations between photons are probabilistic, succeeding some fraction of the time rather than on demand, which forces an architecture that generates enormous numbers of small entangled states and stitches together the ones that succeed. The photonic bet is a manufacturing bet: that a semiconductor fab can produce the required volume of components cheaply enough to absorb an overhead that no other platform has to pay. It’s the highest-variance position in the field, and the one where fab access is worth more than qubit count.
Silicon spin qubits. Information stored in the spin of a single electron confined in a quantum dot on isotopically purified silicon, fabricated with the same complementary metal-oxide-semiconductor, or CMOS, processes used for conventional chips. Intel’s Tunnel Falls device carries 12 qubits and comes off Intel’s own 300 mm lines. Diraq, working with imec, has reported above 99% two-qubit gate fidelity on devices selected at random from industrially fabricated wafers, which is the result that matters, because it speaks to yield rather than to a single hero device. Silicon spin qubits are physically minute, and some designs operate above 1 kelvin rather than at 10 millikelvin, which removes the helium-3 dependency. The gap is qubit count. A dozen qubits against a thousand is a real distance, and the thesis rests on the claim that a mature manufacturing base closes it faster than hand-assembly can.
Cat qubits. A superconducting variant worth separating out, because it attacks the overhead problem directly rather than by scaling. Information is encoded in a superposition of two oscillating microwave states, engineered so that one type of error, the bit flip, is suppressed exponentially by the physics, leaving only the other type, the phase flip, to be corrected by a simple and cheap code. Amazon’s Ocelot chip encoded a logical qubit in nine physical components where a comparable surface code would need roughly 49. Alice & Bob claims bit-flip stability exceeding an hour. The unsolved problem is performing two-qubit gates without destroying the noise bias that the whole design depends on, and until that’s demonstrated at high fidelity, the hardware saving remains projected rather than realized.
Defect spins in diamond and other materials. Nitrogen-vacancy centres in diamond and similar defects hold spin states that can be controlled optically at room temperature. As a route to a general-purpose processor they lag badly. As nodes in a quantum network, as sensors, and as the interface between a stationary qubit and a flying photon, they’re already useful. Evaluate them as networking and sensing companies, not as computing companies.
The Axes That Separate the Platforms
Once the carriers are laid out, the comparison collapses onto five axes, and only five.
Gate speed ranges over three orders of magnitude, from nanoseconds for superconducting circuits to milliseconds for some ion operations. Speed sets how many operations fit inside a useful runtime, and a fault-tolerant algorithm may need hundreds of millions of them.
Connectivity determines which error correction codes are available. Nearest-neighbour hardware is stuck with the surface code and its overhead. All-to-all or reconfigurable connectivity opens the qLDPC family and cuts the physical-to-logical ratio dramatically. This axis is currently doing more to change timelines than raw qubit count is.
Physical-to-logical ratio is the compounding of the two above. A platform at 2 to 1 needs a few hundred physical qubits for a hundred logical ones. A platform at 1,000 to 1 needs a hundred thousand.
Manufacturing path asks whether the company is assembling machines or printing them. Photonics and silicon spins are printing. Ions and atoms are assembling, with the compensation that their qubits arrive perfect from nature rather than from a fab.
Operating environment covers cryogenics, laser systems, vacuum, power draw, and floor space. It’s the axis that determines whether a customer can host a machine or has to rent time on someone else’s.
There’s a useful precedent for what happens next. Through the 1990s, processors were sold on clock speed, and buyers compared megahertz across chips that differed enormously in how much work they completed per cycle. The comparison held until Intel’s Pentium 4 pushed clock rates to 3.8 GHz while delivering less real throughput than slower competitors, at which point the industry abandoned the number and moved to benchmarks of delivered work. Quantum computing is at the megahertz stage. Qubit count is the number on the box, and it will be replaced within a few years by some measure of logical operations completed at a given error rate. A diligence process that already tracks logical operations at a given error rate, alongside qubit counts, will not have to re-rank its portfolio when the benchmark changes.
Reading a Modality Claim
Four distinctions separate a real result from a press release, and we drill all four in class.
Physical or logical. A claim of 1,000 qubits almost always means physical qubits. A claim of 50 logical qubits is a far stronger statement and a far rarer one. If a deck doesn’t say which, assume physical.
Announced or demonstrated. Roadmap targets are commitments, not results. IBM’s Starling target of 200 logical qubits by 2029 is a published intention with engineering behind it, and it is not a machine. Both belong in a model, in different columns.
Peer-reviewed, preprint, or press release. The gap between the three is where most valuation error lives. A Nature paper has been through referees who were paid nothing to be generous.
What the referees actually said. This is the one that gets skipped, and Microsoft’s Majorana 1 chip in February 2025 is the worked example. The announcement described an eight-qubit-capacity device on a new class of material, and it was accompanied by a paper in Nature. A Majorana zero mode is a predicted quasiparticle that would appear in pairs at the ends of a suitable superconducting wire, with the qubit’s value stored jointly across the pair, so that no local disturbance can read or corrupt it. That’s the prize: hardware error rates orders of magnitude below anything achievable today. The Nature reviewers, however, stated in the published assessment that the measurements did not by themselves establish whether the states observed were topological. Non-topological explanations for the same signatures exist, and the field has been here before. A prominent 2018 Nature paper claiming quantized Majorana conductance was retracted in 2021 after its data handling was examined. No braiding operation has been performed, no topological gate has been executed, and no error correction has been demonstrated on this platform.
That makes the programme a research position rather than a product line. Topological qubits belong on a strategic watch list with a defined trigger, which is the first experiment whose outcome cannot be explained by ordinary bound states or disorder. They don’t belong in a revenue model or in a threat timeline, and a fund that has them in either is mispricing the rest of its book.
The Exotic Tail
Beyond the seven carriers sits a long tail of approaches that appear in surveys and occasionally in pitch decks. Most of them are research instruments rather than products, and a short taxonomy prevents them from being either dismissed or overvalued.
Boson sampling sends indistinguishable photons through an optical network and samples the pattern of where they come out. Proposed by Scott Aaronson and Alex Arkhipov in 2011, it’s not programmable and cannot run general algorithms, but the output distribution is believed to be intractable to simulate classically, which made it the vehicle for two headline advantage demonstrations: the Jiuzhang experiment with 76 photons in 2020 and Xanadu’s Borealis with 216 squeezed modes in 2022. Treat these as physics results with narrow commercial application, mainly certified randomness.
Continuous-variable photonics encodes information in the amplitude and phase of light fields rather than in single photons, and reaches fault tolerance through the Gottesman-Kitaev-Preskill state, a way of encoding a qubit in a light field so that small drifts in amplitude or phase can be detected and corrected. It’s a genuine competing route within photonics rather than an exotic curiosity.
Quantum walks, holonomic gates, dissipative computing, and quantum cellular automata are control techniques and theoretical models rather than hardware platforms. Holonomic gates in particular may end up improving several platforms, since driving a system around a closed loop in parameter space produces an operation that’s insensitive to certain timing errors. When one of these appears in a company description, the right question is which physical carrier it runs on.
Quantum acoustics uses quantized sound waves in nanomechanical resonators as memory and as a bridge between otherwise incompatible systems. Its near-term role is as a component inside other machines.
Hybrid architectures are not exotic at all and will probably be the norm. Ion traps linked by photonic channels, superconducting processors with specialized memory, annealers used to warm-start gate-model routines, and quantum processors attached to classical high-performance computing clusters are all in development now. A portfolio built on the assumption that one carrier wins outright is betting against the history of computing hardware, which has consistently produced specialization rather than a single winner.
What This Means for Cryptography Timelines
The reason modality analysis reaches beyond hardware investing is that only some of these machines threaten current encryption, and knowing which ones changes the shape of the risk.
A cryptographically relevant quantum computer, or CRQC, is a machine capable of running Shor’s algorithm at the scale needed to break the public-key cryptography protecting internet traffic, code signing, and financial messaging today. That requires a universal, error-corrected gate-model machine. Annealers cannot do it, boson samplers cannot do it, and analog simulators cannot do it, regardless of qubit count. Claims to the contrary surface periodically and have not survived examination.
The resource estimates have moved, and they’ve moved in one direction. Craig Gidney’s 2025 analysis puts the requirement for factoring a 2048-bit RSA key at fewer than one million noisy physical qubits, down from around 20 million in the equivalent 2019 estimate. That reduction came from better algorithms and better error correction rather than from better hardware. The qLDPC ratios now being reported on ion and atom platforms pull in the same direction, since they reduce the physical qubits needed per logical qubit. Neither development produces a machine tomorrow. Both compress the distance.
The response doesn’t depend on picking the winning modality, which is the practical point for anyone doing this analysis. The United States has standardized post-quantum cryptography, or PQC, in ML-KEM for key establishment and ML-DSA and SLH-DSA for signatures, published by NIST in 2024, with FN-DSA following. The NSA’s CNSA 2.0 requirements put national security systems on a transition path completing in 2033. The driver for acting sooner is harvest now, decrypt later, or HNDL, in which an adversary records encrypted traffic today and decrypts it once a capable machine exists. For data with a confidentiality requirement measured in decades, the deadline is set by the data’s lifetime rather than by anyone’s qubit roadmap. Migration methodology is covered in depth at pqcframework.org.
For an investor, the practical consequence is narrower. Portfolio companies handling long-lived confidential data have a cryptographic migration cost that belongs in diligence now, alongside the usual security review, and it’s independent of which hardware bet you like.
Where to Take This Further
A map is a starting position. Judging a specific company still means reading its results properly, knowing which of the five axes it’s actually competing on, and recognizing when a number in a deck has been chosen because it flatters.
Quantum Academy’s hardware and architecture training covers the material above in depth, with the modality comparisons worked through against published results rather than vendor summaries, and with the error correction arithmetic done step by step so that a physical-to-logical ratio stops being a slogan. It’s built for people who make decisions about quantum technology rather than for people who build it, and it assumes no physics background beyond what’s in this article. Programs and enrollment details are at quantumacademy.com/.
For deeper technical treatment of any individual modality, the full modality taxonomy on PostQuantum.com goes carrier by carrier in considerably more detail than a single article allows.