Zoom from the refrigerator to the nanowire, meet the Majorana zero modes that would store a qubit, and see how measurements stand in for braids.
Status as of October 2026. Every step says whether it is demonstrated, reported in a preprint, proposed, a company claim, or an ideal model computed in your browser. Numbers link to their sources.
Scroll to read, or press Play at the bottom. The arrow keys step back and forth; K plays and pauses. Each figure has its own controls.
Powers of ten: from metres to nanometres in six steps.
Most of a quantum computer is a refrigerator. Each plate is colder than the one above it. In one common model the plates sit at about 35 K, 2.9 K, 0.88 K and 0.082 K, and the bottom plate reaches 0.006 K [1]. The chip hangs below that coldest plate.
Temperatures are for an empty Bluefors XLD400 dilution refrigerator (Krinner et al., Table 1) [1]. Microsoft has not published its own cryostat layout, so the drawing is schematic.
At the bottom plate sits a small copper holder with microwave cables. Microsoft says its Majorana 1 chip can be held in the palm of one’s hand [2]. The holder here is drawn about 3 cm across.
The package and its connectors are illustrative; only the palm-sized description comes from the company [2].
On the chip, each qubit is shaped like the letter H. Microsoft calls it a tetron. In February 2025 it said Majorana 1 carries eight of them [2, 3]. The qubits are tiny: each wire is 3.5 µm long, about one twentieth of a hair’s width [12].
Drawn here as a 4 × 2 block, like the error-detection device of the roadmap [11]. We found no published layout of the Majorana 1 chip; the announcement gives only eight qubits [2, 3]. The spacing between tetrons is schematic.
Two parallel wires, each 3.5 µm long, joined in the middle by a superconducting bar 1 µm long [12]. The four wire ends are where Majorana zero modes are expected. Small quantum dots beside the wires read them out.
A wire should be many coherence lengths ξ long, ξ being how far a Majorana mode spreads. The tetron paper estimates L/ξ ≳ 1 for these wires, against about 10 needed for fault tolerance [12]. The cut marks the cross-section shown next.
Cut one wire across. An aluminium strip 60 nm wide and 3.5 nm thick lies on a semiconductor sandwich with a 9.1 nm layer of indium arsenide [4]. A gate above it pushes electrons away everywhere except under the aluminium. What is left is a wire of electrons a few tens of nanometres wide.
Stack in the published supplement of [4] (S1.2, S1.3): Al 60 × 3.5 nm; InAlAs top barrier 6 nm; InAs well 9.1 nm; InAlAs lower barrier 25 nm; graded buffer on semi-insulating InP. Above: dielectric 1 (about 2 nm AlO and 10 nm HfO), gate layer 1 (Ti/Au/Ti/Al, 2/15/2/10 nm), dielectric 2 (15 nm AlO). The field, about 2 T, points along the wire, into the page here [4].
Gate layer 1, held negative, empties the quantum well everywhere except under the Al strip. Gate layer 2 (Ti/Au, 10/60 nm) carries the dot plungers and the cutters at the junctions, so it lies outside this slice [4]. The arXiv version of [4] gave the Al as 6.5 nm thick; the published supplement says 3.5 nm.
Pulled apart, the wire is a stack where each layer has one job. The aluminium lends its electron pairing. The indium arsenide brings strong spin-orbit coupling and a large response to magnetic fields. The gate sets how many electrons sit in the wire. The magnet, pointing along the wire and so into the page, does the rest.
In June 2026 Microsoft reported tetrons with 10 nm of lead (Pb) in place of aluminium, on a stack grown on GaSb [25]. That work is a preprint.
Theory predicts that a wire tuned into a topological phase holds half an electron at each end.
A semiconductor wire, a superconductor touching it, and a magnetic field along the wire [8, 9]. Two knobs matter: the field strength and the gate voltage, which sets the energy μ of the electrons. The chart shows where the wire stays ordinary, called trivial, and where it should become topological.
The wire is topological when the Zeeman energy beats the pairing and the chemical potential together: , with [8, 9].
Watch the knob cross the boundary, or drag it. In the toy chain below, every electron is written as two halves called Majorana modes. In the trivial phase the halves pair up inside each box. In the topological phase they pair with their neighbours instead, and one half is left over at each end [7].
The chain is Kitaev’s model [7]: . Bond width and brightness show which coupling is stronger, inside a box or between neighbours; glows are the weights of the lowest mode, computed exactly here.
The map from the wire’s to the chain’s is a cartoon that keeps the phase boundary in place. It is not a derivation of the wire model.
The two leftover halves form one ordinary electron level near zero energy. It can be empty or filled, at almost the same energy; the tiny difference between the two falls exponentially as the wire gets longer [7]. This pair of states is the raw material of a topological qubit.
The splitting behaves as [7, 4, 12]. In the chain with it is , which the unit tests check.
Join two topological wires with a superconducting bar and you get a tetron: four Majorana zero modes, γ1 to γ4, at the four ends [10, 11]. The top wire of the H is the chain you just tuned.
Numbering follows Microsoft’s roadmap: γ1 and γ2 on one wire, γ3 and γ4 on the other, γ1 and γ3 on the left [11].
The H is a small island. Adding one more electron to it costs energy, so the island keeps its total even or odd. With the total fixed, the four ends share exactly two states. That is one qubit [10, 11]. Whether real devices host these modes is still argued about: see chapter 8.
Units: the field is given as its Zeeman energy VZ = gμBB/2, and VZ and μ in units of the pairing energy Δ; the chain energies in units of t, the energy for an electron to hop between boxes.
Not in any one place: in the shared parity of pairs of ends.
Is the top wire’s level empty or filled? That is its parity, even or odd, and it is the qubit’s Z value [11]. The answer belongs to the two ends together. Neither end alone knows it. On a device, a tetron qubit is prepared by measuring its Z (or X) parity and keeping the outcome [11].
Roadmap convention: on the island with [11].
Shake one end with a stray voltage. The qubit lives in the pair, so a kick at one end can only act through the coupling between the two ends. In theory that coupling shrinks exponentially once the wire is many coherence lengths ξ long (ξ is how far a Majorana mode spreads). Try the slider, then kick. Microsoft estimates L/ξ ≳ 1 for its current wires; about 10 is needed [12].
The kick is modelled as a phase with , : a hypothetical noise strength chosen to make the scaling visible.
In Microsoft’s aluminium tetron, both parities switched on their own [12]. The X loop switched every 14.5 µs. The authors put this down to parity switches inside a wire, from a leftover coupling between modes at its ends: the coupling the last step assumed was tiny. The Z parity switched every 12.4 ms, put down to quasiparticles, stray unpaired electrons, landing on the wire. A quasiparticle flips the island parity [10]. One that lands on end 1 flips both Z and X, because both pairs contain γ1. Press Send a quasiparticle to watch one.
A quasiparticle landing on end 1 acts as : it anticommutes with , and the island parity. Landing on end 2 it would flip only , on end 3 only , and on end 4 neither, while the island parity flips every time.
The authors also write that trivial models with additional hidden MZMs can be fine-tuned to give similar results [12].
A quantum dot listens to two ends at once.
To read Z, quantum dots are tunnel-coupled to both ends of one wire. An electron can then go around a loop: dot, one end, through the wire, the other end, back to the dot. A magnetic flux threads the loop [4].
In the 2025 device three dots and the wire form the interferometer loop; the flux is set by a small out-of-plane field [4]. That device had one wire; the second wire is drawn to match the tetron of [12].
Interference around the loop depends on the parity. It shifts the dot’s quantum capacitance, a measure of how easily the dot takes on charge. In the 2025 device the two values were about 0.32 and 1.28 femtofarads (fF) [4]. A microwave tone reflected from the dot’s circuit reads the difference. Which parity gives the high value depends on the flux: at another flux setting the two levels swap, so the labels here hold for one flux setting of this model. The authors read the two levels as the two parities; whether that reading holds is disputed (chapter 8) [4, 19].
Measured centres 322 aF and 1280 aF, with widths 99 aF and 92 aF at 91 µs integration (Supplementary S3.3) [4].
Press Measure. The simulator draws an outcome by the Born rule, which gives the odds of each outcome from the state. Measure again: the same answer, until the parity flips. Longer integration averages away noise. In the 2025 device the signal equalled the noise, a signal-to-noise ratio (SNR) of 1, after 3.6 µs [4].
SNR grows as : 5.01 at 91 µs, so 0.52 in 1 µs and 1 in 3.6 µs [4].
Record a long trace and the capacitance jumps between two levels like a telegraph. The typical (mean) time between jumps is the parity lifetime: more than 1 ms in the 2025 device [4]. The best integration time is a trade-off. From the measured noise and an assumed 2 ms lifetime, the authors estimate that 32.5 µs would give an error near 1%. That is a projection, not a measured error rate [4].
Assignment error, Supplementary Eq. S31 [4]: with ms.
Choosing another lifetime keeps the 2025 SNR, a what-if. Reported values: X loop 14.5 µs and Z loop 12.4 ms in an aluminium tetron [12]; Z parity about 20 s in an InAs–Pb device [25].
Exchanging Majoranas without moving anything.
Swapping two Majorana modes changes the qubit in a way that depends only on the swap, not on the details of the path. That is the hoped-for protection. Moving modes around each other is hard, so in 2008 Bonderson, Freedman and Nayak showed that a sequence of measurements can do the same job [13].
An exchange of and is . For it is the phase gate up to a global phase.
Two tetrons: a helper above, the qubit below. Measure X on the helper, then the joint Z parity of both tetrons, then Y on the helper, then X on the helper again [11]. The helper starts in Z = +1, so each of the four outcomes here is a fair coin toss; the last three always are. The joint ZZ measurement entangles helper and qubit: the record shades that stretch, and the qubit’s purity, on the right, drops from 1 (a state of its own) to 0.5 (fully entangled). The Y measurement on the helper ends it, and what is left on the qubit is the braid. Press Run to watch one sequence.
Read right to left: , roadmap Eq. 13 [11]. Only one step involves both tetrons. Purity of the qubit’s reduced state, from the exact state after each measurement: 1, 0.5, 1, 1 in all 16 branches, checked in the unit tests.
The result equals swapping γ1 and γ2 of the qubit, the gate S, followed by a Pauli Z flip when the first three outcomes multiply to +1 [11]. Software tracks the flip. Nothing has to move.
, Eq. 11 [11]. All 16 branches are verified in the unit tests.
Why a four-MZM ZZ step and not two-MZM steps across tetrons: the charging energy fixes each island’s parity, so “all measurements must involve an even number of Majorana operators on each island” [11, 10]. With six MZMs on one island (a hexon), the textbook ancilla-pair sequence (1,5), (2,5), (5,6) also gives up to a Majorana string; the tests check all 8 branches [13, 10].
On the right, the four measurements in time turn into the braid they equal: two worldlines that cross once. The braiding lab shows the same idea for a different kind of anyon.
Each branch is the braid times a Pauli, so a sequence of such steps composes like braids [11, 13].
Open the braiding labMajorana braids give only Clifford gates.
Start the qubit at the top of its sphere and apply any braids you like. It only ever lands on six points: the two poles and four points on the equator [27]. Press the braid buttons, or try every braid word.
The braid images of generate the 24-element single-qubit Clifford group (modulo phase), found here by exhaustive search.
Each braid turns the questions X, Y and Z into X, Y or Z again, up to a sign. Gates that do this are called Clifford gates. A computer that only has them can be simulated efficiently on an ordinary computer [28].
For example : , , . The table under the sphere is computed live.
To reach the rest of the sphere you need something braids cannot give, such as a T gate: an eighth of a turn, half the smallest turn braids can make. The usual plan is magic states: make many noisy copies of a special state and purify them [14]. Microsoft’s roadmap treats these non-Clifford operations separately [11].
The T state lies on the equator between two braid destinations.
A different anyon, the Fibonacci anyon, would be universal by braiding alone [15]. Its braids fill the sphere on the right. No material has been shown to host it. Quantum processors have simulated it [16, 17].
Generators here: , , checked for .
Open the braiding labFrom one tetron to arrays that catch errors: the published plan.
Microsoft’s roadmap lists four devices: one tetron to benchmark a qubit by measurements; two tetrons for Clifford gates by measurement; a 4 × 2 array of eight for error detection; and a large array for lattice surgery between two logical qubits [11]. As of October 2026, reported results are parity measurements on single tetrons, one of them inside a multi-tetron device [12, 25]. No generation has been completed: even the first one’s measurement-based benchmarking has not been reported.
Eight tetrons in two columns, as two patches of four. The planned code is the ladder code, which for this size is the distance-2 Bacon-Shor code [11]. Each round measures XX across every row, then ZZ down each column within a patch.
After each step the products of the last two rounds give the checks XXXX and ZZZZ; their product is YYYY, as in roadmap Fig. 6 [11].
Press Inject to flip one qubit. The next round’s checks change, which flags the error. This code can notice an error but not tell which qubit to fix: detection, not yet correction [11].
An X error anticommutes with the ZZ step, so ZZZZ flips; a Z error flips XXXX. The simulator injects one of the two at random.
For bigger arrays the roadmap proposes Hastings-Haah Floquet codes or a surface code built from pairwise measurements. It draws a 27 × 13 array holding two logical qubits of distance 7 [11]. These are designs, not devices.
On 19 February 2025 Microsoft announced Majorana 1 as “the world’s first Quantum Processing Unit (QPU) powered by a Topological Core, designed to scale to a million qubits on a single chip”, with eight topological qubits on it [3]. On 2 June 2026 it announced Majorana 2, built with lead, with what it calls “a mean qubit lifetime of 20 seconds”, and named 2029 as its target for a scalable quantum computer [26]. These are company statements.
In the matching preprint, the 20 s is the Z-parity lifetime of one wire of a tetron in a multi-tetron device; X measurements are not reported there [25].
What has been shown, what is claimed, and what is disputed.
Peer reviewed: a single-shot, flux-dependent signal with two capacitance levels in an InAs–Al device, which the authors interpret as the fermion parity of the wire [4]. Dwell times exceeded 1 ms, and the authors project a 1% assignment error for an assumed 2 ms lifetime [4]. They write that this measurement does not, by itself, determine whether the states are topological [4]. Nature’s editors added a note.
“The editorial team wishes to point out that the results in this manuscript do not represent evidence for the presence of Majorana zero modes in the reported devices.”
Two loops on one tetron gave X and Z measurements, with lifetimes of 14.5 µs and 12.4 ms and assignment errors of 16% and 0.5% [12]. Physical Review A accepted that paper in October 2026. A 2026 preprint with lead reached Z-parity lifetimes of about 20 s, and a topological gap near 70 µeV in a test wire, against about 30 µeV before [25]. That preprint reports parity of one wire, not X measurements.
The gap values come from Microsoft’s topological gap protocol [29], which is the method under dispute.
Henry Legg argued that the topological gap protocol gives answers that depend on analysis choices, and that the 2025 device regions show no clear gap [18, 19]. Without a gap, he argues, the signal may not be the parity of a superconducting wire at all [19]. Microsoft rebutted [22]. In June 2026 Nature published Legg’s critique with Microsoft’s reply [20, 21]. At the APS Global Physics Summit in March 2025, Microsoft showed X-measurement data whose two-level signal, by its own account, could not be seen by eye; several physicists stayed unconvinced [23]. Doubt has a history here: in 2021 Nature retracted a 2018 paper that had reported quantized Majorana conductance [30].
We found no published report, as of October 2026, of a braid by measurement, a protected gate, or error detection on a tetron array. In February 2025 DARPA selected Microsoft and PsiQuantum for the final phase of its US2QC program [24]. DARPA gives that phase the same goals as Stage C of its Quantum Benchmarking Initiative [31], and in March 2026 it wrote that the two US2QC performers had advanced to Stage C [32]. Stage C checks whether a design can be built as intended. It is an evaluation, not a demonstration.
Primary papers where possible; company statements are marked as such in the text. Checked October 2026.
The models on this page are ideal and exact: Majorana operators from the Jordan-Wigner map, braids , parity projectors, the Kitaev chain and the readout statistics. They run in your browser and are checked by unit tests. Device drawings are schematic unless a dimension is cited.