Dr. Kris Naudts, Zeynep Koruturk and Donald Harmitt @ Firgun Ventures.

Firgun Ventures is a VC firm investing in Series A/B quantum scale-ups globally.

Few questions in the quantum sector are asked more often, or answered more loosely, than how many qubits a useful quantum computer will require. There is no single qubit count at which a quantum computer becomes useful, because the right number depends entirely on what the machine is meant to do, on which type of machine is being used, and on how cleverly the underlying problem has been formulated. Press releases announce systems with hundreds of qubits, then thousands, with the implicit suggestion that the field is approaching a finish line, when the reality is more nuanced and considerably more interesting. A more productive approach is to abandon the abstract question and ask instead how many qubits each specific application needs on each specific type of machine, then to organise the answers into bands of ambition and difficulty.

Why Qubit Numbers Can Mislead

Not all qubits are created equal. A qubit is the basic building block of a quantum computer, the equivalent of a classical bit, but qubits are fragile and error-prone, so serious computation requires bundling many physical qubits into a single reliable logical qubit through quantum error correction. That ratio is decisive, since depending on the technology and the target error rate it can range from roughly two to one at the optimistic end to more than two thousand to one at the demanding end. This is precisely why headline figures mislead, because two systems can both claim thousands of qubits while being nowhere near equivalent if one counts raw hardware and the other counts error-corrected logical qubits. Requirements then vary for four further reasons: the type of problem, since simulating a molecule differs sharply from routing or factoring; how cleverly the problem is represented, as better reformulations have repeatedly shrunk resource estimates; the hardware modality, since gate-based, photonic and annealing machines do not measure resources alike; and how complete the estimate is, since counting only the core algorithm rather than the full error-correction and routing overhead can understate the true figure by an order of magnitude.

Band One: Early Scientific Applications Within Reach

The first band covers carefully chosen scientific problems, mainly in chemistry and materials science, that may become genuinely interesting with somewhere between tens and a few hundred logical qubits. Chemistry and materials are natural early candidates because quantum computers are well suited to modelling the electrons, bonds and strongly correlated materials that are themselves quantum in nature. Recent work suggests useful early simulations of small, well-chosen chemistry problems may be possible with roughly 25 to 100 logical qubits, while condensed-matter models such as the Fermi-Hubbard system have produced estimates in the low hundreds. The experimental hints are already surfacing, with Quantinuum reporting in 2026 that it had used 64 logical qubits to simulate quantum magnetism at a scale it called exceedingly difficult for classical methods, one of the few public examples of a vendor tying error-corrected logical qubits to a recognisable application. 

Commercially relevant work is appearing too, with battery-cathode studies led by researchers at Xanadu producing resource estimates of roughly 100 to 414 logical qubits. The general consensus from band one is that tens to hundreds of logical qubits will be enough to do scientifically recognisable work in carefully chosen areas where classical methods struggle. This is likely to be the first credible frontier of quantum value.

Band Two: Commercially Ambitious Applications

Commercial returns become more plausible in the second band, which covers more realistic chemistry, certain finance problems and industrial materials simulations that tend to require thousands of logical qubits, and it is here that the most striking algorithmic progress is happening. The clearest example is the FeMoco benchmark, a model of nitrogen fixation that a major 2021 study placed at around 2,100 logical qubits before a 2025 algorithmic improvement cut the estimate to roughly 1,500. Cytochrome P450 enzymes, which mediate around three-quarters of human drug metabolism, sit in similar territory. A Google, Boehringer Ingelheim and collaborator study found that, under one set of assumptions, the largest model could be assessed with a few hundred logical qubits at the algorithmic layer, translating to several million physical qubits. Finance belongs here as well, with rigorous studies of derivative pricing putting credible thresholds between roughly 4,700 and 8,000 logical qubits, while elliptic-curve cryptography, which underpins many blockchain and digital-signature systems, looks easier still at around 1,200 to 1,450 logical qubits, plausibly within reach of the 500,000 to one-million-qubit machines on the public 2030 to 2033 roadmaps of IBM, IonQ and Diraq.

All in all, band-two applications often sit somewhere between 1,000 and 10,000 logical qubits, depending on how realistic the molecule is and how aggressively the problem has been compressed.

Band Three: The Headline Tasks That Still Demand Very Large Machines

The applications that dominate public discussion, breaking RSA encryption and proving a generic advantage on hard optimisation, make up the third band, and they still demand very large fault-tolerant systems often running into the millions of physical qubits. Cryptography illustrates how fast the picture is moving. For RSA-2048, a 2021 estimate suggested around 20 million noisy physical qubits could factor a key in roughly eight hours; a 2025 update using better arithmetic cut that to fewer than one million, and a February 2026 architecture based on quantum LDPC codes pushed it down by another order of magnitude to fewer than 100,000 physical qubits, a collapse achieved through software alone. Optimisation is one of the most heavily marketed but algorithmically least certain classes, with the strongest rigorous study on hard problems suggesting universal machines might need tens of millions of physical qubits to cross over classical methods. Quantum annealers tell a different story again, since D-Wave already fields systems with more than 4,400 physical qubits, though those are not comparable to the logical qubits of a fault-tolerant gate-based machine.

What This Means for The Quantum Ecosystem Observers

The single most useful takeaway is that the headline qubit count is almost never the right number to track. The variable that actually governs when a quantum computer becomes commercially relevant is the physical-to-logical ratio, because a company quoting thousands of physical qubits may command only a handful of reliable logical ones, and the gap between the two is where most overstated roadmaps hide. 

The second variable is algorithmic compression, and it is the one most consistently under-priced, since across all three bands the resource estimates have been collapsing through software improvement faster than hardware itself has advanced, with the RSA figure falling roughly two-hundred-fold in five years. The implication is that the layer translating raw qubits into useful computation, namely error correction and algorithm design, may capture value as durably as the hardware does. The trend that rewards fewer and better-orchestrated qubits favours the error-correction and control layer, where the UK's Riverlane has built a specialism, and it sustains the case for scalable silicon approaches such as Quantum Motion, a UK silicon-spin developer, a Firgun Ventures portfolio company. 

The Staircase, Not the Finish Line

The future of quantum computing is better understood as a staircase than a single milestone. Some scientifically and commercially interesting steps may be reached with machines on the order of a hundred logical qubits, others will require thousands, and the hardest tasks will still demand millions of physical qubits, with the spread between those steps wide enough that conflating any two leads to faulty technical and commercial conclusions. Different applications will therefore mature at very different rates, and diligence has to interrogate not just the qubit headline but the logical-qubit translation, the modality assumptions and the algorithmic compression behind any claim. The right question is rarely how many qubits a company has. It is how many reliable qubits its target application actually needs, and how credibly its roadmap closes the gap.