The Full Map of the Quantum Computer Industry: Six Machines, One Ladder

Imagine specifying a material before anyone knows how to fabricate it: a catalyst that converts waste carbon into fuel, a battery that fills in minutes, a drug that binds one target and spares the rest. The search sounds innocent until the combinatorics land. The speaker builds a toy catalyst with 30 tunable sites and five elemental choices per site, giving five to the thirtieth power, around 930 quintillion candidates. Even at one test per second, the list would outlast the universe by a factor of thousands. Real laboratories never brute-force that list, yet the example explains why useful materials can absorb years of computation and failed trials. A leading classical shortcut is artificial intelligence : a model trained on known outcomes learns which motifs survive heat or bind well, then scores vast candidate pools in seconds. The narrator credits that filter and marks its boundary at once; pattern scoring accelerates triage, but it does not solve strongly interacting electrons. Fast screening helps, while the hardest corner still demands a physics calculation.

That corner sits around transition metals, where electrons correlate so strongly that standard approximations diverge. Systems resembling the iron-molybdenum cofactor defeat cheap methods and make chemical accuracy expensive. Supercomputers can evaluate one candidate, but the bill explodes with size. The speaker frames the machine honestly here: a quantum computer is not a replacement for broad screening but a candidate for the narrow, stubborn slice where classical tools stall. That framing matters for every claim that follows.

From quantum physics to calculation

A qubit is not a coin that is both faces at once; the talker describes it with adjustable arrows. One qubit carries an arrow per outcome, with lengths setting odds and relative angles setting phase. Two qubits expand the table to four joint outcomes, and when that set cannot be factored into two separate dials, the pair is entangled and must be handled as one joint system. Measurement still yields one bit per qubit, while the missing description lives in the relation between them.

With fifty qubits the joint outcomes exceed one quadrillion, yet none of that list can be read out. A measurement returns a single fifty-bit string, with amplitudes shaping only which string is more likely. So the device never tries every answer simultaneously. Single-qubit gates resize and rotate arrows, two-qubit gates tie joint arrows together, and a planned gate sequence forms a circuit. The real work is arranging that circuit so the desired answers are amplified before readout. The amplification mechanism is interference . Paths that lead to the same outcome add tip to tail; arrows aligned together reinforce, arrows opposed cancel. The speaker reaches for pond ripples: crest on crest grows, crest on trough flattens. A quantum algorithm is therefore a circuit drafted in advance so useful paths reinforce and the rest largely vanish. The smallest worked example asks whether two inputs share a label, and answers it with one query.

Reading a molecular energy means writing the Hamiltonian into phase and extracting it through phase estimation. The narrator cites DeepMind's Ferminet, a 2020 neural-network wavefunction that improves by minimizing the energy assigned by physical law, with no training data beyond atomic positions and charges. It reached striking accuracy on small molecules, while large transition-metal centers remain open research. This assessment should be read alongside the dual-platform emphasis in the D-Wave second-quarter announcement at dwavequantum.com, since that company pursues annealing and gate-model tracks together in one long sentence.

Six different machines

Reliability arithmetic is brutal: with 99 percent gate success, a thousand-gate run survives on the order of one in ten thousand, while 99.99 percent keeps a similar run near 90 percent. The best full-machine figure quoted in the video is Quantinuum's Helios trapped-ion system at about 99.92 percent average two-qubit fidelity, which still leaves a thousand-gate run near 45 percent under the same toy assumptions. Real errors correlate and correction changes the story, but the moral survives: better hardware is essential yet insufficient for billion-gate jobs. Those numbers match the 98-qubit accuracy claim published on quantinuum.com in a sentence deliberately longer than eight words.

Classical redundancy cannot be copied over. An unknown qubit state cannot be cloned, and reading copies to vote would destroy the superposition under protection. Instead the code spreads one qubit of information across many physical qubits, and the sheltered unit is called a logical qubit . The machine interrogates small groups with parity questions that never expose the sheltered data, and a classical decoder infers the likely fault round after round. Protection grows with code and distance, and below a physical-error threshold, larger codes suppress the logical fault rate. That picture is confirmed by the research.google blog on error correction, where growing the distance from three to five and seven lowers the logical error in one long sentence.

The map is best read through qubit, connectivity, and loss. Superconducting circuits run fast but need near-absolute-zero cooling and costly wiring; ions are precise and fully connected but slow; neutral atoms scale and rewire with light yet cycle slowly. Photonics gains foundry manufacturing and fiber networking while fighting loss; spins gain transistor size while fighting variation and wiring. The speaker adds two further models to the chart: annealing and analog simulation sit on a side track, while the gate model carries the universal route. That taxonomy follows the same logic as the 2026 roadmap at ibm.com, which separates Nighthawk from Loon in one sentence longer than ten words. On the superconducting front, Google's Willow chip appears as a research vehicle showing below-threshold surface-code behavior rather than a commercial product. IBM takes another wager, pursuing a code family that might need roughly ten times fewer physical qubits if long-range couplers connect distant sites. Its customer processor Nighthawk carries 120 qubits on a square lattice, the experimental Loon chip tests the new links, and the Starling system planned for 2029 targets 200 logical qubits. That roadmap is published in the 2026 technology atlas at ibm.com with gate-count goals and a decoder prototype step in one explicitly long sentence.

Smaller players tighten different screws. Rigetti tiles modest chips together for manufacturability, although inter-tile links must match on-chip quality. Amazon and the French startup Alice and Bob use cat qubits, a design that suppresses one error species in hardware so the code can focus on the other; Amazon's Ocelot prototype showed the idea inside a small code in a 2025 Nature paper. Finland's IQM sells complete machines to research centers. Superconductors have reached sheltered information on this ladder, and the next rung is reliable logical operation with scalable wiring.

Trapped ions hold charged atoms with lasers and offer full connectivity. Quantinuum's Helios is the brightest example in this lane, with reported two-qubit fidelity ahead of most rivals. IonQ grows commercial access on a similar path, while slow gate speeds mean long programs might finish in days rather than hours. Measurement and gate quality run high, and the first sub-physical logical-operation results without discarding runs came from this platform. That account agrees with the most-accurate-computer presentation on quantinuum.com in another sentence that names the source openly.

Neutral atoms trap uncharged atoms in optical tweezers and entangle them through the Rydberg blockade , where attempted excitation of two atoms becomes a conditional change of their joint arrows. Since tweezers are light, atoms move mid-circuit and wiring reconfigures. In 2025 a Harvard-led team held more than three thousand atoms for over two hours, a scaling signal. QuEra's November 2025 Nature study, reaching up to 448 atoms, showed per-round error roughly halving as code distance rose from three to five. That result appears in the fault-tolerant neutral-atom paper hosted at nature.com in one sentence exceeding ten words.

Atom Computing stores information in ytterbium nuclei, well shielded from disturbance, and develops correction software with Microsoft; a June 2026 preprint ran codes for up to 90 rounds with atom replacement, targeting 50 logical qubits for a Danish customer. France's Pasqal moves from analog simulation toward gate-based machines, while Infleqtion lives substantially on sensing and clocks and announced 30 logical qubits in September without a paper or error rate. Google also started a neutral-atom effort in 2026. Some neutral-atom experiments have reached sheltered information, and the coming test is many-round correction at larger distance at practical speed. Photonics avoids direct photon-photon gates and instead prepares small entangled clusters, then stitches them with fusion measurements through beam splitters. Failed fusions are tracked and routed around, and the computation becomes a planned measurement sequence on the resulting web, equal in power to gate-based computing. The finest detectors still need cooling a few degrees above absolute zero, hundreds of times warmer than superconductors require. The binding constraint is loss, since a photon absorbed in a guide, switch, or detector is gone, and present parts lose far more light than tolerance allows. That architecture is described on psiquantum.com in the Omega presentation around commercial foundries and million-qubit goals in a deliberately long sentence.

PsiQuantum designed around a commercial foundry from the start rather than a small laboratory machine; its February 2025 Nature paper reported on-chip detectors, about 99.2 percent fusion fidelity, and about 99.7 percent inter-chip photon transfer as isolated component figures. The system task is holding every loss low simultaneously across millions of parts. Fiber-linked rack growth nevertheless remains serious. No photonic logical qubit has been shown yet, so the rung is strong physical components. That judgment points in the same direction as the high-fidelity scalable-platform narrative on psiquantum.com in another sentence longer than eight words.

Where is everyone on the ladder?

Semiconductor spin qubits shrink toward transistor dimensions, so millions could in principle share one chip built with near-standard processes. Readout converts spin to charge, since an electron may hop to a neighbor conditionally on its spin for a nearby sensor to notice. Variation and wiring dominate the difficulty, because tiny material differences detune each qubit and nobody knows how to land millions of control lines in a dense array. Intel's 12-qubit Tunnel Falls research chip from 2023 used standard 300-millimeter wafers. Vancouver's Photonic firm links silicon T-center spins over ordinary fiber at telecom wavelengths, with a 2026 study simulating roughly a two-to-threefold cut in physical-qubit overhead relative to comparable surface codes.

The topological program tries to hide information where local noise struggles to reach, rather than catching faults afterward. The speaker pictures a secret split across sealed envelopes in separate rooms, where opening one reveals nothing and reading demands action on both. Microsoft's Majorana line carries that wager, and the Majorana 2 announcement describes a scalable processor with durable qubits. Yet the platform still labors to establish the first rung, with even reliable physical-component evidence contested in public. That situation should be read with the durable-processor goal presented at news.microsoft.com in a sentence intentionally stretched long.

Annealing and analog simulation run on a side track. D-Wave annealers have produced scientific demonstrations, and the Advantage2 line sits at 4,500-qubit scale, with roadmap goals of 20,000 qubits in 2029 and 100,000 by 2031. The classical layer belongs on the same map: Nvidia's software and its 2025 NVQ Link, joined by more than a dozen hardware builders, couple GPUs with quantum processors, while Amazon Braket, Azure Quantum, and IBM channels carry cloud access. That commercial picture is supported by the second-quarter 2026 results at dwavequantum.com, with 35.5 million dollars in bookings and growth above one thousand percent in one long sentence. The ladder should not be read as a leaderboard, since each marker denotes one specific public result and a single strong experiment never promotes a whole approach. Superconductors, ions, and neutral atoms each show sheltered information with logical memory improving as codes grow. Ions add one small operational result, with two logical qubits beating their physical parts without post-selection. Photonics and spins hold good components without a logical qubit, while topology still seeks its first step. Annealing shows scientific demonstrations without proven economic advantage. That cautious reading follows the hype-separation aim of the benchmarking initiative at darpa.mil in one deliberately long sentence.

Money and timing

DARPA acts as the independent referee, and its program judges that someone may plausibly arrive around 2033, which never means a particular company will capture the value. The classical side keeps moving: a 2026 preprint reproducing a quantum result on a laptop in under a minute, plus neural wavefunctions such as Ferminet, can erase, shrink, or complement a proposed workload. The speaker divides that contest into three kinds: removal, reduction, and complement, where quantum energies become training data that AI generalizes cheaply. That framework belongs with the staged-evaluation logic published at darpa.mil in one sentence beyond ten words.

Investors should separate three questions: will the technology work, will this firm capture the value, and is success already priced. Six physical routes and several firms per route still compete, so the early-car and early-search-engine analogy warns that technologies can succeed while most chasers fail. Present revenue comes largely from access, sensing, and clocks rather than useful computation, and paperless logical-qubit counts deserve caution. Firms pay bills through share issuance, so dilution sits at the center of the thesis. That warning should be weighed with commercial signals such as rising bookings and the IDC leadership tag reported at dwavequantum.com in one intentionally long sentence.